Volume 40, March 2015 ISSN: 0167-6296

JOURNAL OF
HEALTH
ECONOMICS
 JOURNAL OF HEALTH ECONOMICS Publication information: Journal of Health Economics (ISSN 0167-6296). For 2015, volumes 3944 are scheduled for publication.
Subscription prices are available upon request from the Publisher or from the Regional Sales Office nearest you or from this journals website
(http://www.elsevier.com/locate/jhe). Further information is available on this journal and other Elsevier products through Elseviers website:
(http://www.elsevier.com). Subscriptions are accepted on a prepaid basis only and are entered on a calendar year basis. Issues are sent by
standard mail (surface within Europe, air delivery outside Europe). Priority rates are available upon request. Claims for missing issues should
Aims and Scope be made within six months of the date of dispatch.
This Journal seeks articles related to the economics of health and medical care. Its scope will include the following
topics: production of health and health services; demand and utilization of health services; financing of health services; Advertising information: If you are interested in advertising or other commercial opportunities please e-mail Commercialsales@elsevier.com
and your enquiry will be passed to the correct person who will respond to you within 48 hours.
measurement of health; behavioral models of demanders, suppliers and other health care agencies; health behaviors
and policy interventions; efficiency and distributional aspects of health policy; and such other topics as the Editors Funding body agreements and policies
may deem appropriate. Applications to problems in both developed and less-developed countries are welcomed. Elsevier has established agreements and developed policies to allow authors whose articles appear in journals published by Elsevier, to
comply with potential manuscript archiving requirements as specified as conditions of their grant awards. To learn more about existing
agreements and policies please visit http://www.elsevier.com/fundingbodies
Editors
M. CHALKLEY, Centre for Health Economics, University of York, Heslington, York, UK. Orders, claims, and journal enquiries: Please contact the Elsevier Customer Service Department nearest you:
E-mail: martin.chalkley@york.ac.uk St. Louis: Elsevier Customer Service Department, 3251 Riverport Lane, Maryland Heights, MO 63043, USA; phone: (877) 8397126 [toll free
M.E. CHERNEW, Department of Health Care Policy, Harvard Medical School, Boston, MA, USA. within the USA]; (+1) (314) 4478878 [outside the USA]; fax: (+1) (314) 4478077; e-mail: JournalCustomerService-usa@elsevier.com
Tokyo: Elsevier Customer Service Department, 4F Higashi-Azabu, 1-Chome Bldg, 1-9-15 Higashi-Azabu, Minato-ku, Tokyo 106-0044, Japan;
E-mail: chernew@hcp.med.harvard.edu phone: (+81) (3) 5561 5037; fax: (+81) (3) 5561 5047; e-mail: JournalsCustomerServiceJapan@elsevier.com
D. CUTLER, Department of Economics, Harvard University, Cambridge, MA, USA. E-mail: dcutler@harvard.edu Singapore: Elsevier Customer Service Department, 3 Killiney Road, #08-01 Winsland House I, Singapore 239519; phone: (+65) 63490222;
A. LLERAS-MUNEY, Economics Department, University of California at Los Angeles, Los Angeles, CA, USA. fax: (+65) 67331510; e-mail: JournalsCustomerServiceAPAC@elsevier.com
Oxford: Elsevier Customer Service Department, The Boulevard, Langford Lane, Kidlington, Oxford OX5 1GB, UK; phone: (+44) (1865) 843434;
E-mail: alleras@econ.ucla.edu fax: (+44) (1865) 843970; e-mail: JournalsCustomerServiceEMEA@elsevier.com
E. MEARA, The Dartmouth Institute of Health Policy & Clinical Practice, Dartmouth College, Lebanon, NH, USA.
E-mail: ellen.r.meara@dartmouth.edu Author enquiries
For enquiries relating to the submission of articles (including electronic submission) please visit this journals homepage at
N. RICE, Centre for Health Economics, University of York, Heslington, York, UK. http://www.elsevier.com/locate/jhe. For detailed instructions on the preparation of electronic artwork, please visit http://www.elsevier.com/
E-mail: nigel.rice@york.ac.uk artworkinstructions. Contact details for questions arising after acceptance of an article, especially those relating to proofs, will be provided
L. SICILIANI, Department of Economics and Related Studies, University of York, Heslington, York, UK. by the publisher. You can track accepted articles at http://www.elsevier.com/trackarticle. You can also check our Author FAQs at http://www.
elsevier.com/authorFAQ and/or contact Customer Support via http://support.elsevier.com.
E-mail: luigi.siciliani@york.ac.uk
A.D. STREET, Centre for Health Economics, University of York, Heslington, York, UK. Illustration services
E-mail: andrew.street@york.ac.uk Elseviers WebShop (http://webshop.elsevier.com/illustrationservices) offers Illustration Services to authors preparing to submit a manuscript
but concerned about the quality of the images accompanying their article. Elseviers expert illustrators can produce scientific, technical and
medical-style images, as well as a full range of charts, tables and graphs. Image polishing is also available, where our illustrators take your
Associate Editors image(s) and improve them to a professional standard. Please visit the website to find out more.
J.E. ASKILDSEN, University of Bergen, Bergen, Norway.
K. BAICKER, Harvard School of Public Health, Boston, MA, USA. USA mailing notice: Journal of Health Economics (ISSN 0167-6296) is published bimonthly (January, March, May, July, September and
P.P. BARROS, Universidade Nova de Lisboa, Lisbon, Portugal. November) by Elsevier B.V. (P.O. Box 211, 1000 AE Amsterdam, The Netherlands). Periodicals postage paid at Jamaica, NY 11431 and
additional mailing offices (not valid for journal supplements).
A. BASU, University of Washington, Seattle, WA, USA. USA POSTMASTER: Send change of address to Journal of Health Economics, Elsevier Customer Service Department, 3251 Riverport Lane,
H. BLEICHRODT, Erasmus University, Rotterdam, The Netherlands. Maryland Heights, MO 63043, USA.
J. CAWLEY, Cornell University, Ithaca, New York, USA
AIRFREIGHT AND MAILING in USA by Air Business Ltd., c/o Worldnet Shipping Inc., 156-15, 146th Avenue, 2nd Floor, Jamaica, NY 11434,
R.P. ELLIS, Boston University, Boston, MA, USA. USA.
J. GLAZER, Tel Aviv University, Tel Aviv, Israel.
S. GLIED, Columbia University, New York, NY, USA. The paper used in this publication meets the requirements of ANSI/NISO Z39.48-1992
M. GROSSMAN, National Bureau of Economic Research, New York, NY, USA. (Permanence of Paper).
J. GRUBER, Massachusetts Institute of Technology, Cambridge, MA, USA. Printed by Henry Ling Ltd, Dorchester, UK.
R. KAESTNER, University of Illinois at Chicago, Chicago, IL, USA.
M. KIFMANN, University of Hamburg, Hamburg, Germany.
J. MULLAHY, University of Wisconsin-Madison, Madison, WI, USA.
O.A. ODONNELL, Erasmus University, Rotterdam, The Netherlands. For a full and complete Guide for Authors, please go to: http://www.elsevier.com/locate/Jhe
P. OLIVELLA, Universitat Autnoma de Barcelona, Barcelona, Spain.
C. PROPPER, University of Bristol, Bristol, UK.
A. SCOTT, University of Melbourne, Victoria, Australia.

The Journal of Health Economics has no page charges

doi:10.1016/S0167-6296(15)00015-6
 Journal of Health Economics 40 (2015) 19

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

Using provider performance incentives to increase HIV testing and

counseling services in Rwanda
Damien de Walque a,,1 , Paul J. Gertler b,1 , Sergio Bautista-Arredondo c , Ada Kwan c ,
Christel Vermeersch d , Jean de Dieu Bizimana e , Agns Binagwaho f , Jeanine Condo g
a
Development Research Group, The World Bank, United States
b
Haas School of Business, University of California, Berkeley, United States
c
National Institute of Public Health, Cuernavaca, Mexico
d
The World Bank, United States
e
Camris International, United States
f
Ministry of Health, Government of Rwanda, Kigali, Rwanda; Senior Lecturer, Harvard Medical School, Department of Global Health and Social Medicine;
Clinical Professor of Pediatrics, Geisel School of Medicine at Dartmouth College, United States
g
School of Public Health, College of Medicine and Health Sciences, University of Rwanda, Kigali, Rwanda

a r t i c l e i n f o a b s t r a c t

Article history: Paying for performance provides nancial rewards to medical care providers for improvements in per-
Received 1 March 2013 formance measured by utilization and quality of care indicators. In 2006, Rwanda began a pay for
Received in revised form performance scheme to improve health services delivery, including HIV/AIDS services. Using a prospec-
30 November 2014
tive quasi-experimental design, this study examines the schemes impact on individual and couples HIV
Accepted 4 December 2014
testing. We nd a positive impact of pay for performance on HIV testing among married individuals
Available online 12 December 2014
(10.2 percentage points increase). Paying for performance also increased testing by both partners by 14.7
percentage point among discordant couples in which only one of the partners is an AIDS patient.
JEL classication:
I12 2014 Published by Elsevier B.V.
O15

Keywords:
Performance-based nancing
HIV testing and counseling
Africa
Health human resources
Couple testing

1. Introduction evidence demonstrates that antiretroviral treatment (ART) of HIV+

HIV testing and counseling (HTC) is a gateway to improving
prevention and care efforts, and has become a core strategy for
decreasing HIV transmission and incidence (Glick, 2005). There
have been calls to devote more resources to couple HTC since HIV
transmission is high in discordant couples, i.e. couples in which
only one of the partners is infected by HIV/AIDS, especially if the
infected partner either does not know his or her status or has not
revealed it to the uninfected partner (Padian et al., 1993). Recent
Corresponding author at: Development Research Group, The World Bank, 1818
H Street, NW, Washington, DC 20433, United States. Tel.: +1 202 473 2517;
fax: +1 202 614 0234.
E-mail address: ddewalque@worldbank.org (D. de Walque).
1
These authors shared rst authorship.
individuals is very effective in preventing transmission of the HIV
virus within couples (Cohen et al., 2011; Dodd et al., 2010; El-Sadr
et al., 2010; Wagner et al., 2010).
As a result, HTC couple testing, especially among discordant
couples, has become a key component of prevention programs
in generalized epidemic countries (Allen et al., 2003). Despite
the promise of HTC and the large amount of development assis-
tance for HIV/AIDS, HTC uptake has only recently seen modest
improvements (United Nations, 2011). Moreover, there are few
documented successful experiences of HTC programs reaching cou-
ples (Padian et al., 1993; Painter, 2001).2
2
An important exception is from Thornton (2008) who demonstrated that cash
value vouchers doubled the percentage of individuals who obtained their HIV test
results, given that they had been tested.

http://dx.doi.org/10.1016/j.jhealeco.2014.12.001
0167-6296/ 2014 Published by Elsevier B.V.
2 D. de Walque et al. / Journal of Health Economics 40 (2015) 19
A promising, yet largely untested, intervention to increase test-
ing is to pay health providers for increasing participation in HTC
through provider-initiated testing (PIT). This is part of the more
general pay-for-performance (P4P) movement that gives nan-
cial rewards at the facility and/or provider levels to improve
performance measured by specic utilization and quality of care
indicators. P4P is now being piloted or scaled up in about 40 low-
and middle-income countries3 (Eichler and Levine, 2009; Meessen
et al., 2011).
This paper evaluates the impact of Rwandas national P4P
scheme on individual and couple HTC. Building on the lessons
from pilot experiences in a few provinces, Rwanda initiated in
2006 a national P4P scheme at the health facility level to improve
health services delivery, including HIV/AIDS services. We use data
from a prospective impact evaluation we nested into the national
scale-up of P4P in Rwanda, producing evidence from an impact
evaluation at scale with more external validity than closely moni-
tored pilot experiments. The Rwanda P4P scheme provided larger
payment for couple HTC than for individual HTC, allowing us
to explicitly test whether supply-side incentives are an effective
intervention to increase couple HTC and in particular for dis-
cordant couples among whom the risk of HIV transmission is
higher.
An important aspect of our study is the identication of the
effects of incentives in a budget neutral environment. In other
words, we test whether the government is able to purchase more
services for the same amount of money through incentive con-
tracts than through xed budgets. This is important because if
P4P achieves its results just from increased government spend-
ing, then the same results could be achieved from a simple
increase in budget without incurring the administrative costs
associated with implementing the incentive scheme. In order to
identify the incentive effects in a budget neutral setting we hold
constant total government P4P payments by increasing the tra-
ditional input-based budgets of the comparison group by the
average amount of incentive payments to treatment facilities.
As a result, the average amount paid to facilities in the treat-
ment group is equal to the average amount paid to comparison
facilities.4
Our results show a positive impact of P4P on the probability of
individuals having ever been tested. Indeed, when disaggregated
by couple status we nd that individuals living in a couple drive all
of the results. There is no effect on single individuals even when
we condition on being sexually active. However, there is a positive
and statistically signicant impact of 10.2 percentage points for
individuals in couples, which amounts to a 14.5% increase over the
control group testing rate. The impact of P4P on couple testing is
particularly strong among discordant couples (i.e. one partner is
conrmed HIV+ and the other is not), encouraging the partners of
identied HIV patients to come for HTC. These results are consistent
with the fact that the Rwanda P4P strongly encouraged couple and
partner testing, paying US$0.92 per new individual tested for HIV
and US$4.59 per couple/partner jointly tested. For couple/partner
3
See www.rbfhealth.org for an updated list of countries together with a descrip-
tion of programs.
4
While total government spending is held constant in that average facility budg-
ets are equal between the treatment and control group, the distribution of facility
budgets is not necessarily held constant. One would expect a higher variance in
the treatment group as the more capable facilities obtain higher P4P payments in
the treatment group than in the control group and the less capable facilities would
obtain smaller increases in the treatment group than in the control group. Hence,
an individual facilitys budget is not being held constant and therefore we cannot
interpret the estimates as a pure compensated incentive effect for an individual
facility.
testing, it was not necessary for both partners to come together for
testing: either the partners come together for HIV testing, or one
comes after the other during the same reporting period.
These results show that incentive payments are an effective
means of increasing participation in HTC. They are especially
important for Sub-Saharan Africa, where nearly 80% of HIV-infected
adults are unaware of their HIV status and over 90% do not know
whether their partners are infected (World Health Organization,
2009). With only 12% of the global population, Sub-Saharan Africa
is home to 68% of all people living with HIV.5
Our ndings contribute to the limited but growing evidence
base that paying health facilities for performance is a feasible and
effective method for improving health system performance in low-
and middle-income countries. Our work contributes to the general
literature on P4P in medical care, as it is the rst to examine the
impact of P4P incentives on HIV related services.6 More impor-
tantly, the role of incentives in P4P is key. Because the comparison
facilities regular budgets were increased by an amount equal to
the P4P payment to the treatment group, we were able to isolate
the P4P incentive effect from the resource effect.
Our work also contributes to the relatively small literature on
the effects of paying medical care providers for performance in
developing countries.7 There are four well-identied and related
evaluations in other low- and middle-income countries. Hospital-
based physicians in the Philippines who received extra bonus pay
based in part on knowledge of appropriate clinical procedures
reported increases in clinical knowledge (Peabody et al., 2011).
In Indonesia, performance incentives to villages for improvements
in health outcomes led to an increase in labor supply from health
providers (Olken et al., 2012). Miller et al. (2012) found that bonus
payments to schools signicantly reduced anemia among students
in China. Finally, using the same identication strategy as this study
but a different sample of health facilities and households with
recent births, Basinga et al. (2011) found in Rwanda that P4P had
signicant positive impact on institutional deliveries and preven-
tive care visits by young children, and improved quality of prenatal
care, but found no effect on the number of prenatal care visits or on
immunization rates. A follow-on study also reported large impacts
on child health outcomes and provider productivity (Gertler and
Vermeersch, 2012).
The remainder of the paper is organized as follows. Section 2
describes the context of the health sector in Rwanda and the P4P
intervention evaluated. In Section 4, we present our data and we
describe our identication strategy. Section 6 presents our results
while Section 7 concludes.
2. The health sector in Rwanda and the P4P intervention
In 2005, HIV prevalence for adults in Rwanda was estimated
at 3% (Institut National de la Statistique du Rwanda (INSR) and
ORC Macro, 2006). The Government of Rwanda (GoR) decided to
address the HIV epidemic by not only aggressively scaling up HIV
services nationwide, but also utilizing the planned national P4P
model to target HIV preventive services, i.e. HTC, PIT, prevention of
5
In 2011 an estimated 34 million people were living with HIV worldwide, the
number of AIDS-related deaths was 1.7 million and there were 2.5 million new HIV
infections (UNAIDS, 2012).
6
See Witter et al. (2011) for a recent systematic review of health care performance
incentives in low- and middle-income countries. Most of the literature that they cite
do not have control groups and estimate the impact of P4P as jumps in time trends
of the amount of services providers by treatment facilities.
7
There is, however, a growing literature on P4P for medical care in the U.S. and
other high-income countries with mixed results. See Alshamsan et al. (2010), Scott
et al. (2011), Van Herck et al. (2010).
 D. de Walque et al. / Journal of Health Economics 40 (2015) 19 3

Table 1
Output indicators and unit payments by P4P for HIV services.

Service Quantity indicators for HIV Amount paid by P4P per case (US$)

1 HTC Number of clients tested for HIV at the HTC center 0.92
2 HTC/PMTCT Number of couples/partners tested during the reporting month 4.59
3 PMTCT Number of HIV+ pregnant women on ARV treatment during labor 4.59
4 PMTCT Number of infants born to HIV+ mothers tested 9.17
5 Care Number of HIV+ patients who received CD4 test 4.59
6 Care Number of HIV+ patients treated with co-trimoxazole each month 0.46
7 ARV Number of new HIV+ adults on ARV treatment 4.59
8 ARV Number of new HIV+ infants on ARV treatment 6.88
9 HIV prevention Number of HIV+ women on contraception 2.75
10 HIV prevention Total number of HIV+ patients tested for tuberculosis 2.75

Note: P4P: pay-for-performance; HIV: Human Immunodeciency Virus; HTC: HIV testing and counseling; PMTCT: prevention of mother-to-child transmission (of HIV); ARV:
antiretroviral drug.

mother-to-child transmission (PMTCT), and ART for AIDS patients,
and other HIV-related prevention and care services. The GoR ini-
tiated the P4P scheme in 2006 to supplement the input-based
budgets of health centers and hospitals with bonus payments
conditioned to the quantity and quality of key health services
(Ministre de la Sant Rpublique du Rwanda, 2006).
The scheme pays for different dimensions of services, includ-
ing maternal and child health, tuberculosis, and HIV/AIDS. For
HIV/AIDS, the P4P scheme pays for 10 output indicators, such as
the number of clients and the number of couples tested for HIV
(US$4.59 per couple), the number of newly diagnosed HIV-positive
patients on ART (US$0.92 per individual), and the number of HIV-
positive women on contraception (Table 1). The Ministry of Health
(MoH) dened the output indicators and each corresponding unit
payment based on health priorities, available budget and the pre-
vious NGO pilot experiences (Ministre de la Sant Rpublique du
Rwanda, 2008). This analysis focuses on the rst two indicators
dealing with HTC: (i) the number of clients tested for HIV at the
HTC center and (ii) the number of couples/partners tested at the
health facility.
Facilities submit monthly reports and quarterly requests for
payment to the district P4P steering committee, which is respon-
sible for verifying the quality of the data and authorizing payment.
Each committee veries reports by sending auditors to facilities
on unannounced random days each quarter. The auditors verify
the data reported are the same as the data recorded in facil-
ity records. In addition, during the 20062008 period the MoH
nanced one patient tracking survey to conduct face-to-face inter-
views with approximately 1000 patients to verify the accuracy of
the records. This survey found false reporting was below 5% (Health
Development and Performance, 2008).
Quarterly payments go directly to facilities and are used at each
facilitys discretion. In the sample of 10 treatment facilities in our
study, the P4P payments amounted to 14% of overall expenditures
in 2007. On average, facilities allocated 6080% of the P4P funds to
increase personnel compensation.
It is worth noting that the Rwanda P4P scheme was imple-
mented in the context of a larger health sector reform and
during a period in which HIV/AIDS services, including delivery of
antiretroviral treatment, were extensively scaled-up. As of 2005,
83 health facilities were delivering ART to 19,058 persons living
with HIV/AIDS (PLWHA), and 229 facilities were providing HTC ser-
vices with 449,259 individuals ever tested (Center for Treatment
and Research on AIDS Malaria Tuberculosis and Other Epidemics,
2007). By 2008, coverage of ART had increased more than three-
fold and more than doubled for HTC (Center for Treatment and
Research on AIDS Malaria Tuberculosis and Other Epidemics, 2008).
Our methods described below allow separating the P4P impacts
from the effects of the overall scale-up of HIV/AIDS services in
Rwanda.
3. Experimental design
The evaluation exploits data generated from a stratied cluster
randomized design at the district level. In 2006, the Government of
Rwanda began to scale the implementation of P4P for HIV services
nationally. Rwanda manages its health care system at the district
level and P4P is no exception. Districts were randomly assigned into
either a treatment group that began receiving P4P bonus payments
starting in January 2007 or a control group that began receiving
PBF payments in July of 2008, about 18 months after the treatment
group. Since a primary objective of the study was to examine the
effect of P4P incentives in a budget neutral setting from the point of
view of the purchaser (e.g. the government), we held constant the
average level of resources constant across treatment and compar-
ison facilities by increasing the traditional input-based budgets of
the comparison facilities by the average amount of P4P payments
to treatment facilities on a quarterly basis (Fig. 1).
The Government identied 12 of the 30 districts that con-
tained facilities that provided HTC and ART services and had not
previously received PBF payments. The 12 districts had broad
geographic representation from all parts of the country. They
contained 72 health care facilities that provided HTC and ART
services, of which 48 were already receiving PBF before 2006
and hence excluded from the study. The remaining clinics in the
districts were randomly assigned into either the treatment or con-
trol group. The clinics in each district were all assigned to either
the treatment or control group. Of the 24 clinics in the sample,
10 were assigned to the treatment group and 14 to the control
group.8
By and large there was good but not perfect compliance with the
randomization protocol. In 2 of the districts assigned to the control
group, Gakenke and Rwamagana, most of the non-study clinics had
already started receiving PBF payments prior to 2006. For adminis-
trative purposes the Government decided to reassign the remaining
3 clinics from these 2 districts to the treatment group. Hence, 21
out of the 24 clinics in the study complied with the randomization
protocol.
4. Data
We conducted a baseline survey of the facilities from August
until November 2006 and a follow-up survey from April until
July 2008.9 We also conducted a household level survey that
8
The distribution of the number of facilities providing HIV by treatment, con-
trol and excluded from the study due to having PBF prior to 2006 is presented in
Appendix Table A.1 and Map 1.
9
The study team submitted the research protocol to the Rwanda National Ethics
Committee, which approved the research design, methodology and methods for
4 D. de Walque et al. / Journal of Health Economics 40 (2015) 19
Fig. 1. Experimental design. The originally planned evaluation consisted of 18 Phase
I health facilities and 18 Phase II health facilities from 7 and 7 districts, respectively.
Prior to implementation of the baseline survey, the administrative district bound-
aries were redrawn in the context of a decentralization effort. As a result, some of the
experimental areas were combined with areas that already had NGO P4P schemes.
Because P4P could not be removed from health facilities that were already imple-
menting the system, and because P4P was managed at the district level, the GoR
required that all facilities within those new districts be in the rst phase (treatment)
of the rollout. This led the evaluation team to switch the assignment of treatment
and comparison for eight districts from four blocks, as well as add one block to the
sample.
was administered to a sample of 1200 households with an HIV+
positive member, and 400 randomly sampled neighbor house-
holds in the catchment area of the facility. We identied HIV/AIDS
patients either by contacting the health facility where they received
care or via association of PLWHA. We selected them randomly
proportional to the number of HIV/AIDS patients attending each
facility. We obtained informed consent from the patients before
informed consent. Because of the sensitive nature of the sample and survey, the
data collection was managed by the National University of Rwanda School of Pub-
lic Health with guidance from the National AIDS Control Commission (CNLS). The
follow-up surveys were also reviewed by the National Institute of Public Health in
Mexico Institutional Review Board (IPF Code 3627801).
interviewing their household, but, to maintain condentiality, the
other household members were not informed of this selection pro-
cedure.
While only one individual with a non-positive serostatus per
household was interviewed at baseline, all adults with a non-
positive serostatus were interviewed in the follow-up survey.
From the 1017 individuals interviewed at baseline, 395 were re-
interviewed at follow-up. At follow-up, 1248 new individuals were
also interviewed. Our analysis sample is thus best seen as repeated
cross-sections. However, the rate of re-interview at follow-up of
baseline survey respondents was not statistically different across
treatment and control groups both in univariate (42.9% in control
group, 35.2% in treatment group, p-value for difference between
groups: 0.13) and multivariate (coefcient on treatment indicator:
0.06, p-value: 0.12) analyses.10
The outcome measures are constructed using data from the
household surveys. The outcome for individual HTC is an indica-
tor variable for whether the individual has ever been tested for
HIV. For the purpose of the HTC analysis, we exclude individuals
identied in the survey as HIV/AIDS patients who by denition are
HIV positive and are aware of their HIV status. The sample is further
restricted to individuals aged 15 or older. At baseline, the sample
is comprised of 438 individuals in the treatment group and 445 in
the comparison group. Individuals present at follow-up but not at
baseline were not different based on standard socio-demographic
characteristics.
For the analysis of couple testing, we create an indicator using
the question of whether or not the most recent sexual partners
the respondents had in the 12 months prior to the survey had
ever been tested for HIV.11 We further combine the responses
about each respondents individual testing and the testing of their
sexual partners to create an indicator variable for whether both
partners in the couple/sexual partnership have ever been tested.
We then restrict the sample to individuals living with their sexual
partners and who self-reported having had sex in the 12 months
prior to the survey. For this analysis, the unit of observation is
the couple and we include only one report by couple to avoid
double-counting.
5. Summary statistics and balance at baseline
Table 2 reports the baseline means of facility characteristics in
2006. Conrming that the evaluation design achieved balance of
observed characteristics at baseline between the facilities in the
treatment and comparison groups, there are no signicant dif-
ferences between the treatment and comparison groups in terms
of rural location, proportion of district hospitals in the sample
of facilities, proportion of facilities that are government-assisted
or public, size of catchment population, supply of staff, log 2006
expenditures, and allocation of the budget across medical person-
nel, medical supplies and non-medical purposes.
Table 2 further reports the mean 2008 log expenditures
for treatment and comparison facilities with no statistically
signicant difference in the means after the introduction of
P4P in the treatment facilities. This conrms that the program
10
The multivariate analysis also reveals that females, older and wealthier indi-
viduals were more likely to be re-interviewed while individuals with secondary
education were less likely to be interviewed.
11
We chose ever tested as our main dependent variable given that with a 24
months exposure period using tested in the last 12 months would likely be under-
powered. We are conducting our analysis with individuals who are not identied as
HIV patients. If they found out that they were HIV negative during the rst year of
implementation of the P4P program, they might not necessarily need to be tested
again in the last 12 months before the survey.
 D. de Walque et al. / Journal of Health Economics 40 (2015) 19 5

Map 1. Distribution of HIV clinics by study status (treatment and control).

Table 2
Health facility baseline (2006) characteristics.

Treatment Control p-Value

Mean SE Mean SE
(1) (2) (3) (4) (5)

Located in rural area 0.900 0.107 0.714 0.175 0.384

Is a district hospital 0.600 0.142 0.500 0.129 0.612
Facility is public (vs. assisted by the government) 0.400 0.252 0.429 0.174 0.927
Catchment population 135,928 20,229 111,014 20,151 0.402
Number of staff per facility 60.800 16.435 55.000 7.620 0.755
Number of staff per 10 000 population 6.006 0.862 7.208 1.005 0.383
Log total expenditures (2006) 17.432 0.321 17.832 0.303 0.384
Log total expenditures (2008) 18.338 0.744 18.676 0.315 0.684
Medical personnel budget share 0.459 0.047 0.482 0.029 0.685
Medical supply budget share 0.282 0.043 0.264 0.039 0.759
Non-medical budget share 0.259 0.037 0.257 0.020 0.966
Total number of Health Facilities 10 14

Note: Standard errors (SE) were cluster-adjusted using districts as clusters. p-Value is for the difference between the treatment and control groups.

compensated the comparison facilities with an increase in their
traditional input-based budget equal to the increase in treatment
facilities resources and validates the interpretation of any esti-
mated impacts being caused by the introduction of P4P incentives,
as opposed to an increase in nancial resources.
For the analysis of HTC at the individual level, the sample con-
sists of all adults aged 15 and above who were not identied in
the survey as being HIV/AIDS patients: 438 in the treatment group
and 445 in the comparison group. Table 3 reports the baseline
characteristics of all respondents grouped and by marital status.12
There are no statistical differences in baseline means of the out-
come variable ever been tested. For the control variables used
12
For marital status, we dened as married or living in couple both those legally
married and those cohabiting together even without formal marriage.
6 D. de Walque et al. / Journal of Health Economics 40 (2015) 19

Table 3
Baseline (2006) characteristics of adults (>15 years) not identied as HIV patients.

All Not living in couple Living in couple

Treatment Control Treatment Control Treatment Control

(N = 438) (N = 445) (N = 217) (N = 215) (N = 221) (N = 230)

Variable Mean SE Mean SE p Mean SE Mean SE p Mean SE Mean SE p

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15)

Female 0.628 0.039 0.582 0.034 0.391 0.682 0.046 0.661 0.042 0.736 0.575 0.048 0.509 0.043 0.330
Age 34.332 1.191 35.984 1.025 0.318 30.899 1.751 30.498 1.541 0.867 37.703 0.772 41.112 0.765 0.011
Education
No education 0.292 0.027 0.283 0.025 0.809 0.272 0.029 0.200 0.029 0.110 0.312 0.050 0.361 0.044 0.481
Primary 0.616 0.020 0.605 0.021 0.693 0.645 0.055 0.651 0.048 0.936 0.588 0.050 0.561 0.044 0.690
Secondary or 0.091 0.028 0.112 0.024 0.579 0.083 0.044 0.149 0.038 0.284 0.100 0.030 0.078 0.026 0.606
higher
Marital status
Married 0.505 0.039 0.517 0.034 0.819
0.201 0.039 0.180 0.033 0.689 0.406 0.077 0.372 0.066 0.749
Divorced/widow
Never 0.294 0.043 0.303 0.036 0.878 0.595 0.077 0.628 0.066 0.749
married
Log household 11.915 0.310 11.919 0.253 0.992 11.742 0.379 11.621 0.321 0.813 12.084 0.263 12.197 0.223 0.750
asset value
Sexual activity
Never had 0.192 0.042 0.160 0.035 0.572 0.378 0.086 0.312 0.072 0.566
sex
Ever had sex 0.358 0.047 0.391 0.040 0.608 0.530 0.068 0.558 0.059 0.761 0.190 0.037 0.235 0.033 0.388
but not in
past 12
month
Had sex past 0.450 0.032 0.449 0.029 0.994 0.092 0.023 0.130 0.022 0.258 0.801 0.035 0.748 0.032 0.288
12 months
Partner of an 0.155 0.029 0.173 0.025 0.652 0.308 0.073 0.335 0.060 0.779
HIV patient
Ever been 0.580 0.018 0.539 0.020 0.164 0.447 0.050 0.391 0.045 0.425 0.710 0.023 0.678 0.026 0.376
tested

Note: Standard errors (SE) were cluster-adjusted using districts as clusters. p-Value is for the difference between treatment and control groups.

in the regression models, the samples are generally well balanced.
All samples and sub-samples are well balanced in terms of sexual
activity, marital status and assets values.
Strictly speaking, a sero-discordant couple is a couple where one
of the partners is HIV positive and the other is HIV negative. Since
we did not test the study participants for HIV, the only participants
whose HIV status is known to us are those identied as HIV/AIDS
patients who are HIV positive. Hence our denition of discordant
couple is different: we dene as discordant couple a couple where
one partner is identied as an HIV patient and one is not. The pro-
portion of individuals whose partner is identied as HIV patient
is well balanced so that the proportion of discordant couples as
dened above, i.e. a couple where one partner is identied as an
HIV patient and one is not, is well balanced across treatment and
control.
For the analysis of HTC at the couple level, the sample consists of
all adults aged 15 and above who were identied as HIV negative,
who self-reported having had sex in the 12 months preceding the
survey, and living with their sexual partners: 179 in the treatment
group and 180 in the comparison group. Table 4 reports the baseline
characteristics of respondents. There are no statistical differences
in baseline means of the 3 outcome variables: has the respondent

Table 4
Baseline (2006) characteristics of adults (>15 years) not identied as HIV patients, living with their sexual partners and who self-reported having had sex in the 12 months
preceding the survey.

All respondents

Treatment Control
(N = 179) (N = 180)

Variable Mean SE Mean SE p

(1) (2) (3) (4) (5)

Female 0.559 0.060 0.467 0.052 0.272

Age 35.978 0.915 39.150 0.865 0.030
Education
No education 0.263 0.070 0.311 0.059 0.605
Primary 0.643 0.065 0.611 0.056 0.723
Secondary or higher 0.095 0.037 0.078 0.032 0.731
Log household asset value 11.948 0.221 12.099 0.196 0.621
Partner of an HIV patient 0.374 0.079 0.356 0.066 0.860
Ever been tested for HIV 0.788 0.033 0.711 0.033 0.130
Partner has been tested for HIV 0.821 0.047 0.739 0.042 0.218
Couple has been tested for HIV 0.721 0.039 0.650 0.038 0.223

Note: Standard errors (SE) were cluster-adjusted using districts as clusters. p-Value is for the difference between treatment and control groups.
 D. de Walque et al. / Journal of Health Economics 40 (2015) 19
Table 5
Estimated impact of P4P on HIV testing and counseling at the individual level.
All In couple
(1) (2)
** 0.061 0.102
SE (0.040) (0.041)
p-Value 0.126 0.012
%*** 10.09% 14.12%
N 2215 920
Note: ** is the estimated effect of P4P controlling for year, and respondents characteristics including age, gender, age, years of schooling, and log household wealth. Standard
errors (SE) were multi-way cluster-adjusted using districts, survey year and their intersection following the method developed by Cameron et al. (2011) and all models used
a health facility xed effect. p is the p-value for the difference between treatment and control groups; and % ** * = (/Control mean) 100, where the control mean equals
the mean of the dependent variable for the control group at endline (2008).
ever been tested, has the sexual partner of the respondent ever
been tested and have the respondent and his/her sexual partner
ever been tested. The only difference is that, overall, respondents
in the control group are about 3 years older than those in treatment
group. All other variables including education and asset value are
well balanced. However, even if not statistically signicant,13 some
of the differences observed at baseline in Tables 3 and 4 are substan-
tial and we therefore use a difference-in-difference specication.
5.1. Estimation
Given the reassignment of 3 of clinics from the comparison
group to the treatment group before the start of the study,
we view our study as quasi-experimental. While the sample
is balanced at baseline on outcomes and characteristics, it is
possible that the reassignment of districts was correlated with
something unobservable to us and related to health outcomes.
However, redrawing of administrative units took place within
the context of a decentralization agenda that was led by the
Ministry of Local Government, and we nd no evidence that it
was driven by or related to health outcomes (MINALOC (Ministry
of Local Government), 2006).14 Given this reassignment, we will
use difference-in-differences methods that control for unob-
served time invariant characteristics and any potential baseline
imbalance.15 Difference-in-difference specically controls for
any unobserved targeting criteria of decentralization that caused
the reassignment of the clinics from control to treatment status.
This method compares the change in outcomes in the treatment
group to the change in outcomes in the comparison group. By
comparing changes, we control for observed and unobserved time
invariant characteristics as well as for time-varying factors that are
common to the treatment and comparison groups. As we discussed
13
As a further test of balance, we also replaced the facility xed effects in Eq. (1)
and by an indicator for the treatment districts: the indicator for treatment was not
signicantly different than zero in both Tables 5 and 6 (p-values of 0.270 and 0.370,
respectively).
14
According to MINALOC (Ministry of Local Government) (2006), the objective
of the decentralization was to enhance institutional development and capacity
building for responsive local governance, to develop efcient, transparent and
accountable scal and nancial management systems at local government and
grassroots levels.
15
An alternative, sometimes used in the literature, is the intent to treat estimator
that compares the originally assigned treatments to controls. In this case, how-
ever, we would have misassigned 40% of the observations and would be grossly
underpowered. Also, all of the examples we could nd use the ITT in cases where
the study entered the eld intending to implement the original design and where
behavioral choices by the study participants compromised the study design. In our
case, the design was changed before we entered the eld and was not compro-
mised by the study participants. Hence, while our difference in difference estimator
requires stronger assumptions, we believe that it is appropriate in terms of identi-
cation and is valid based on the balance tests and knowledge of the institutional
context that drove the change in design. In our view, the difference in difference
choice maximizes potential power without sacricing internal validity.
7
Not in couple Not in couple and ever had sex
(3) (4)
0.003 0.039
(0.062) (0.074)
0.959 0.599
0.06% 6.55%
1295 683
above, the nal assignment to the treatment and comparison
groups is orthogonal to pre-intervention observable variables,
leading us to believe that there is likely no correlation between
this assignment and unobservables that would drive program
effects.
We treat the 2006 and 2008 household surveys as repeated
cross-sections and estimate the following regression specication
of the difference-in-difference model for individual outcomes:

Yijt = j + 2008 + P4Pj I2008 + k Xkijt + ijt (1)
k
where Yijt is the HTC outcome of individual or couple i living in
facility js catchment area in year t; P4Pj is a dummy variable that
takes value 1 if facility j belongs to Phase I (i.e. started receiving P4P
in 2007) and 0 otherwise; j is a facility xed effect;  2008 is a xed
effect for 2008; I2008 is a dummy variable that takes value 1 if the
year of observation is 2008 and 0 otherwise; the Xkijt are individ-
ual characteristics; and ijt is a zero mean error term. We compute
robust standard errors using multi-way cluster-adjustment by dis-
tricts, survey year and their intersection following the method
developed by Cameron et al. (2011) to account for potential corre-
lation of the error terms at both the cross-section and the temporal
level.
6. Results
Table 5 reports the estimated P4P program impacts on HTC
outcomes using the individual as the unit of analysis. We present
analyses for the entire sample and then conduct sub-group analyses
by marital status. In all of the estimated models, we control for age,
education, and household assets and for gender and marital status
when relevant. Assets are measured as the value of land, durables
in the house, farm animals, farm equipment, and microenterprise
equipment.
In column (1) of Table 5, we nd a positive but not statistically
signicant impact (p-value 0.126) of 6.1 percentage points in the
probability to have ever been tested with respect to the comparison
group. When we restrict the sample to individuals living in a couple
(column 2), we nd a positive impact of 10.2 percentage points
that is statistically signicant at the 5% level, representing a 14.5%
increase from baseline. However, there is no impact on individuals
not currently in a couple regardless of whether they are sexually
active or not (columns 3 and 4). Those larger impacts of P4P on HTC
among married individuals are consistent with the fact that the
P4P scheme strongly encouraged couple and partner testing since
it paid US$0.92 per new individual tested for HIV and US$4.59 per
couple/partner jointly tested.
In Table 6 focusing on the analysis where the couple is the unit
of observation, there are positive but not signicant (p-value 0.190)
impacts of P4P on the likelihood that the respondent reports that
both partners have ever been tested (column 1). That increase is
8 D. de Walque et al. / Journal of Health Economics 40 (2015) 19
Table 6
Estimated impact of PBF on HIV Testing and Counseling at the couple level.
All Discordant couples Non discordant couples
(1) (2) (3)
** 0.086 0.147 0.072
SE (0.066) (0.068) (0.070)
p-Value 0.190 0.030 0.304
%*** 12.68% 17.57% 12.04%
N 572 229 343
Note: ** is the estimated effect of P4P controlling for year, and respondents char-
acteristics including age, gender, age, years of schooling, and log household wealth.
Standard errors (SE) were multi-way cluster-adjusted using districts, survey year
and their intersection following the method developed by Cameron et al. (2011)
and all models used a health facility xed effect. p is the p-value for the difference
between treatment and control groups; and % ** * = (/Control mean) 100, where
the control mean equals the mean of the dependent variable for the control group
at endline (2008).
however signicant and especially strong among couples in which
one of the partners has been identied as living with HIV/AIDS (dis-
cordant couple): the results in column 2 indicate an increase of 14.7
percentage points, signicant at the 5% level and representing an
18.14 percent increase from baseline. The increase is lower and
not statistically signicant for couples, which are not discordant
(column 3).
This analysis with the couples as the units of observation con-
rms that the larger P4P incentives for joint testing especially
encourage both partners in the couple to be tested. The impact of
P4P on couple testing is particularly strong among discordant cou-
ples where one of the partners has been identied as living with
HIV/AIDS, encouraging the partners of identied HIV patients to
come for HTC.
7. Conclusions
Our study examines the impact of the national P4P scheme
in Rwanda on individual HTC and couple HTC, using data from a
prospective experimental design. The results indicate a positive
impact of P4P, concentrated among individuals in couples, on the
probability of individuals having ever been tested. The results also
indicate larger impacts of P4P on HIV testing by both partners, espe-
cially among discordant couples in which only one of the partners is
identied as HIV positive. Our data set oversampled households in
which at least one member knew his/her HIV status and had either
been seeking treatment or contacted an association of PLWHA. As
such, this limits the external validity of our ndings to the subpo-
pulation that is well-connected to the health system.
Our results show signicant increase of HTC coverage in the con-
text of a massive scaling-up of HIV services. P4P was implemented
in the context of a larger health sector reform and during a period
in which HIV/AIDS services, including delivery of ART, were exten-
sively scaled-up. We are not able to identify how this context of
increase of HIV service coverage interacted with the P4P program.
Arguably a P4P intervention could have even greater impacts in a
more static context of HIV service delivery.
Strong encouragement of couple and partner testing is a key
component of the P4P program for HTC in Rwanda. While indi-
vidual HTC is recognized as the necessary gateway for HIV/AIDS
treatment, the prevention benets of individual HTC remain under
discussion (Denison et al., 2008). Joint couple or partner testing on
the other hand appears to have stronger prevention benets, espe-
cially in the case of discordant couples (Allen et al., 2003; Cohen
et al., 2011). However, despite the apparent importance of couple
testing for treatment and prevention purposes, there have been few
successful experiences of HTC programs reaching couples (Padian
et al., 1993; Painter, 2001). Furthermore, recent evidence on the
effectiveness of ART for prevention of HIV transmission among
couples makes this a key intervention of prevention programs in
generalized epidemic countries (Dodd et al., 2010; El-Sadr et al.,
2010; Wagner et al., 2010). Recent evidence on the prevention
effectiveness of ART points to a 95% protection rate among dis-
cordant couples (Cohen et al., 2011). Our results show that P4P is
an effective intervention to target discordant couples for HTC.
The stronger results among individuals in couples and for joint
testing are consistent with the fact that the Rwanda P4P strongly
encouraged couple and partner testing, paying US$0.92 per new
individual tested for HIV and US$4.59 per couple/partner jointly
tested. In general, services for which prices were higher and
providers found easier to implement had the larger responses. For
maternal care in the same Rwandan P4P, large incentives for in
facility delivery, antenatal care quality and child growth moni-
toring resulted in improved services, while lower incentives for
antenatal care utilization had no impact (Basinga et al., 2011). The
resulting improvements were associated with substantially better
child health outcomes (Gertler and Vermeersch, 2012). Our nd-
ings contribute to the growing evidence base that paying health
facilities for performance may be a feasible and effective method
for improving health system performance under certain circum-
stances. Because the comparison facilities regular budgets were
increased by an amount equal to the P4P payment to the treatment
group, we were able to isolate the P4P incentive effect from the
resource effect. The equivalent monetary transfer to the control
group did not achieve the same results. This suggests that the role
of incentives in P4P is key. Similarly, in a very different context,
Argentinas Plan Nacer uses a relatively small amount of resources
(24% of total health expenditure) to provide incentives to health
providers to use resources more efciently and for higher-quality
care to program beneciaries, reducing low birth-weight and in-
hospital neo-natal mortality (Gertler et al., 2014).
Acknowledgements
We thank Anita Asiimwe, Paulin Basinga, Stefano Bertozzi, Gyuri
Fritsche, Alex Kamurase, Kathy Kantengwa, Gayle Martin, Rigobert
Mpendwazi, Cyprien Munyanshongore, Vedaste Ndahirdwa, Isaac
Ntahobakulira, Miriam Schneidman, Jennifer Sturdy, Claude Sek-
abaraga, Louis Rusa, Adam Wagstaff, and many others for their
contributions to the project; the Rwandan Ministry of Health, the
National AIDS Control Commission (CNLS) in Rwanda, Management
Sciences for Health, the Belgian Technical Cooperation, Cordaid,
GTZ, Healthnet, USAID, and the World Bank and their staff for their
collaboration in the implementation of the P4P rollout plan and
supporting the evaluation; and the University of Rwanda School
of Public Health for data collection. We thank the World Banks
Bank-Netherlands Partnership Program and Spanish Impact Eval-
uation Fund, the British Economic and Social Research Council, the
University of California Berkeleys East Africa Social Science Trans-
lation (EASST) Initiative and the Government of Rwanda for their
nancial support. The ndings, interpretations, and conclusions
expressed in this paper are entirely those of the authors. They
do not necessarily represent the views of the International Bank
for Reconstruction and Development/World Bank and its afliated
organizations, or those of the Executive Directors of the World Bank
or the governments they represent.
Appendix A.
See Table A.1.
 D. de Walque et al. / Journal of Health Economics 40 (2015) 19 9

Table A.1
Distribution of HIV clinics by study status (treatment, control and excluded).

Province District # Facilities # Facilities # Facilities with Total # of facilities

assigned to assigned to prior PBF excluded
treatment group comparison group from study

North Gakenke 2 0 9 11
Musanze (Ruhengeri) 0 3 1 4
South Kamonyi 0 2 1 3
Gikongoro (Nyamagabe) 0 1 1 2
Nyaruguru 1 0 4 5
East Kirehe 0 2 3 5
Nyagatare (Umutara) 0 1 5 6
Rwamagana 1 0 7 8
West Karongi (Kibuye) 0 4 3 7
Ngororero 3 0 5 8
Nyabihu 0 1 4 5
Rutsiro 3 0 5 8

Total 10 14 48 72

References
Allen, S., Meinzen-Derr, J., Kautzman, M., Zulu, I., Trask, S., Fideli, U., Musonda, R.,
Kasolo, F., Gao, F., Haworth, A., 2003. Sexual behavior of HIV discordant couples
after HIV counseling and testing. AIDS 17 (5), 733740.
Alshamsan, R., Majeed, A., Car, J., Millett, C., 2010. Impact of pay for performance
on inequalities in health care: systematic review. Journal of Health Services
Research & Policy 15 (3), 178184.
Basinga, P., Gertler, P.J., Binagwaho, A., Soucat, A.L., Sturdy, J., Vermeersch, C.M., 2011.
Effect on maternal and child health services in Rwanda of payment to primary
health-care providers for performance: an impact evaluation. The Lancet 377
(9775), 14211428.
Cameron, A.C., Gelbach, J.B., Miller, D.L., 2011. Robust inference with multiway clus-
tering. Journal of Business & Economic Statistics 29 (2), 238249.
Center for Treatment and Research on AIDS Malaria Tuberculosis and Other Epi-
demics, 2007. Rapport annuel du TRAC 2006. Center for Treatment and Research
on AIDS Malaria Tuberculosis and Other Epidemics, Kigali.
Center for Treatment and Research on AIDS Malaria Tuberculosis and Other Epi-
demics, 2008. 2008 Annual Report. Center for Treatment and Research on AIDS
Malaria Tuberculosis and Other Epidemics, Kigali.
Cohen, M.S., Chen, Y.Q., McCauley, M., Gamble, T., Hosseinipour, M.C., Kumarasamy,
N., Hakim, J.G., Kumwenda, J., Grinsztejn, B., Pilotto, J.H.S., Godbole, S.V., Mehen-
dale, S., Chariyalertsak, S., Santos, B.R., Mayer, K.H., Hoffmann, I.F., Eshleman,
S.H., Piwowar-Manning, E., Wang, L., Makhema, J., Mills, L.A., de Bruyn, G., Sanne,
I., Eron, J., Gallant, J., Havlir, D., Swindells, S., Ribaudo, H., Elharrar, V., Burns, D.,
Taha, T.E., Nielsen-Saines, K., Celentano, D., Essex, M., Fleming, T.R., for the HPTN
052 Study Team, 2011. Prevention of HIV-1 infection with early antiretroviral
therapy. The New England Journal of Medicine 365 (6), 493505.
Denison, J.A., OReilly, K.R., Schmid, G.P., Kennedy, C.E., Sweat, M.D., 2008. HIV
voluntary counseling and testing and behavioral risk reduction in developing
countries: a meta-analysis, 19902005. AIDS and Behavior 12 (3), 363373.
Dodd, P.J., Garnett, G.P., Hallett, T.B., 2010. Examining the promise of HIV elimination
by Test and Treat in hyper-endemic settings. AIDS 24 (5), 729735.
Eichler, R., Levine, R., 2009. Performance Incentives for Global Health: Potential and
Pitfalls. Center for Global Development, Washington, DC.
El-Sadr, W.M., Affrunti, M., Gamble, T., Zerbe, A., 2010. Antiretroviral therapy: a
promising HIV prevention strategy. Journal of Acquired Immune Deciency
Syndromes 55 (Suppl. 2), S116S121.
Gertler, P.J., Vermeersch, C., 2012. Using Performance Incentives to Improve Health
Outcomes. World Bank Policy Research Working Paper No. 6100. The World
Bank.
Gertler, P.J., Giovagnoli, P., Martinez, S., 2014. Rewarding Provider Performance to
Enable a Healthy Start to Life. Evidence from Argentinas Plan Nacer. World Bank
Policy Research Working Paper No. 6884. The World Bank.
Glick, P., 2005. Scaling up HIV voluntary counseling and testing in Africa: what
can evaluation studies tell us about potential prevention impacts? Evaluation
Review 29 (4), 331357.
Health Development and Performance, 2008. Rapport dEnqute de Contre vrica-
tion par la Communaut dans les districts de Nyamasheke, Nyanza, Nyaruguru
et Rulindo.
Institut National de la Statistique du Rwanda (INSR), ORC Macro, 2006. Rwanda
Demographic and Health Survey. INSR and ORC Macro, Calverton, MD, USA.
Meessen, B., Soucat, A., Sekabaraga, C., 2011. Performance-based nancing: just a
donor fad or a catalyst towards comprehensive health-care reform? Bulletin of
the World Health Organization 89, 153156.
Miller, G., Luo, R., Zhang, L., Sylvia, S., Shi, Y., Foo, P., Zhao, Q., Martorell, R., Medina,
A., Rozelle, S., 2012. Effectiveness of provider incentives for anaemia redu-
cation in rural child: a cluster randomized trial. British Medical Journal 345,
e4809.
MINALOC (Ministry of Local Government), 2006. Making Decentralized Service
Delivery Work in Rwanda: Putting the People at the Center of Service Provision,
Policy Note. Government of Rwanda, Kigali.
Ministre de la Sant Rpublique du Rwanda, 2006. Guide de lApproche Con-
tractuelle. Modules de Formation des Acteurs Concerns pour le Dmarrage PBF
dans Rwanda.
Ministre de la Sant Rpublique du Rwanda, 2008. Module de rfrence de forma-
tion continue en planication familial. A lusage des formateurs, superviseurs et
prestataires au niveau des formations sanitaires.
Olken, B.A., Onishi, J., Wong, S., 2012. Should Aid Reward Performance? Evidence
from a Field Experiment on Health and Education in Indonesia. NBER Working
Paper No. 17892.
Padian, N.S., OBrien, T.R., Chang, Y.C., Glass, S., Francis, D.P., 1993. Prevention
of heterosexual transmission of Human Immunodeciency Virus through
couple counseling. Journal of Acquired Immune Deciency Syndromes 6,
10431048.
Painter, T.M., 2001. Voluntary counseling and testing for couples: a high-leverage
intervention for HIV/AIDS prevention in sub-Saharan Africa. Social Science and
Medicine 53 (2001), 13971411.
Peabody, J., Shimkhada, R.R., Quimbo, S., Florentino, J., Bacate, M.F., MuCulloch, C.,
Solon, O., 2011. Financial incentives and measurement improved physicians
quality of care in the Philippines. Health Affairs 30, 773.
Scott, A., Sivey, P., Ait Ouakrim, D., Willenberg, L., Naccarella, L., Furler, J., Young,
D., 2011. The effect of nancial incentives on the quality of health care pro-
vided by primary care physicians. Cochrane Database of Systematic Reviews 9,
CD008451.
Thornton, R.L., 2008. The demand for, and impact of, learning HIV status. American
Economic Review 98 (5), 18291863.
UNAIDS, 2012. AIDS Epidemic Update: November 2011. UNAIDS, Geneva,
Switzerland.
United Nations, 2011. The Millennium Development Goals Report 2011. United
Nations, New York.
Van Herck, P., De Smedt, D., Annemans, L., Remmen, R., Rosenthal, M.B., Ser-
meus, W., 2010. Systematic review: effects. Design choices, and context of
pay-for-performance in health care. BioMed Central Health Services Research
10, 247260.
Wagner, B.G., Kahn, J.S., Blower, S., 2010. Should we try to eliminate HIV epidemics
by using a Test and Treat strategy? AIDS 24 (5), 775776.
Witter, S., Fretheim, A., Kessy, F.L., Lindahl, A.K., 2011. Paying for performance
to improve the delivery of health interventions in low- and middle-
income countries. Cochrane Database of Systematic Reviews 2 (Art. No:
CD007899).
World Health Organization, 2009. Towards Universal Access: Scaling Up Priority
HIV/AIDS Interventions in the Health Sector. Progress Report 2009. World Health
Organization, Geneva, Switzerland.
 Journal of Health Economics 40 (2015) 1025

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

Public health insurance and disparate eligibility of spouses:

The Medicare eligibility gap
Allison Witman
3040 E. Cornwallis Road, P.O. Box 12194, Research Triangle Park, NC 27709-2194, United States

a r t i c l e i n f o a b s t r a c t

Article history: I exploit the age-based eligibility structure of Medicare and the age gap between spouses to examine
Received 14 February 2013 the impact of Medicare eligibility of an older spouse on the insurance coverage of younger, Medicare-
Received in revised form 4 June 2014 ineligible spouses. Using a regression discontinuity framework, I nd that Medicare eligibility of an older
Accepted 19 October 2014
spouse can crowd-out the health insurance coverage of a younger spouse. Medicare eligibility of older
Available online 24 December 2014
wives increases the likelihood that younger husbands are uninsured. Younger wives are less likely to
be covered through an employer-based plan and more likely to have non-group coverage after an older
JEL classication:
husband turns 65.
I13
I18 2014 Elsevier B.V. All rights reserved.
G22

Keywords:
Public health insurance
Private health insurance
Medicare
Crowd-out
Regression discontinuity

1. Introduction McKnight, 2008; Lichtenberg, 2002; McWilliams et al., 2003, 2007).

An estimated 10,000 people per day turn 65 and become eligi-
ble for Medicare, a program that provides insurance coverage for 50
million people and comprises 21% of national health care spending
(Kaiser Family Foundation, 2012).1 Given its size and role as the pri-
mary health insurance provider to elderly Americans, Medicares
effects have been widely studied by economists interested in the
programs impact on beneciaries. Medicare beneciaries experi-
ence sudden changes in health insurance and health care use at
age 65, increasing the likelihood of insurance coverage, the rate
of multiple coverage, the use of preventative services, hospitaliza-
tion, and prescription medications (Card et al., 2008, 2009; Decker
and Rapaport, 2002; Duggan and Morton, 2010; Finkelstein and
I am grateful to Kelly Bedard, Olivier Deschenes, Shelly Lundberg, Peter Kuhn,
Heather Royer, Jenna Stearns, and seminar participants at the UCSB Economics Labor
Lunch for helpful comments and suggestions.
Tel.: +1 9195417211.
E-mail address: awitman@rti.org
1
In addition to being at least 65, age-based qualication for Medicare requires at
least 10 years of Medicare-covered employment by ones self or spouse. Individuals
under age 65 with certain disabilities also qualify for the program.
Despite the breadth of research investigating the effects of Medi-
care, the programs spillovers to Medicare-ineligible individuals
have been largely overlooked.
Medicare is unlike the private health insurance plans that cover
most Americans under age 65 in that it has no coverage provision
for family members or other dependentsMedicare only covers the
eligible individual. The age-based eligibility structure of Medicare
means that households with an age difference between spouses
will inevitably encounter a period of time in which the older spouse
is eligible for the program and the younger spouse is not, creating a
Medicare-eligibility gap between spouses.2 Medicare eligibility of
an older spouse creates the potential for spillovers to a younger
spouse, caused by an interaction between the individual-based
nature of Medicare and the household-based nature of private
insurance. When an older spouse turns 65 the household will have
to decide whether the older spouse will take up Medicare cover-
age, in which case the younger spouse will need to have their own
coverage or go uninsured.
2
Assuming the younger spouse does not qualify based on disability status.

http://dx.doi.org/10.1016/j.jhealeco.2014.10.007
0167-6296/ 2014 Elsevier B.V. All rights reserved.
 A. Witman / Journal of Health Economics 40 (2015) 1025
In this paper, I use data from the National Health Interview
Survey (NHIS) and a regression discontinuity design to examine
the impact of an older spouses Medicare eligibility on the insur-
ance coverage of a younger, Medicare-ineligible spouse. I nd that
Medicare eligibility of an older spouse can crowd-out the health
insurance coverage of a younger spouse, reducing coverage on the
extensive margin and the generosity of coverage. In households
with an older wife and a younger husband, Medicare-ineligible hus-
bands are 3.49 percentage points more likely to be uninsured just
after a wife turns 65. This is a 44% increase in the fraction of married
men in this age range without insurance coverage. I also provide
evidence that the men who are going uninsured are less educated,
working, and relatively healthy. Second, Medicare-ineligible wives
experience large changes in the source of private coverage when a
husband becomes eligible for Medicare. Ineligible wives move out
of private plans in their husbands name (10.43 percentage point
fall) and into private plans in their own name (9.10 percentage
point increase), with a shift toward coverage that can be charac-
terized as less generous. Additionally, I nd that healthier women
are more likely to be uninsured just after a husband turns 65.
These ndings add to existing knowledge on the effects of the
Medicare program as well as identify Medicares age-based eligibil-
ity structure as a new source of crowd-out. Crowd-out is a reduction
in private health insurance coverage caused by eligibility for public
health insurance and has exclusively been studied in the context
of Medicaid and Veterans Insurance. Newly-eligible individuals
may drop private coverage in favor of public insurance (Blumberg
et al., 2000; Boyle and Lahey, 2010; Busch and Duchovny, 2005;
Card and Shore-Sheppard, 2004; Dubay and Kenney, 1996, 1997;
Sasso and Buchmueller, 2004; Yazici and Kaestner, 2000; Shore-
Sheppard et al., 2000); however, declines in private coverage may
not be limited to those who become eligible for public insurance.
Crowd-out can spill over to the family members of public insurance
eligibles as documented by Cutler and Gruber (1996) and Gruber
and Simon (2008), who report that men whose family members
become eligible for Medicaid reduce employer insurance cover-
age. Similar to this paper, Boyle and Lahey (2012) exploit disparate
spousal eligibility for public health insurance. The authors show
that wives increased labor supply after Veterans Insurance expan-
sions granted coverage to their husbands. One explanation for this
result is that wives increased labor supply in order to provide
health insurance for themselves. In contrast to previous studies, the
crowd-out generated by Medicare results from a perfectly antici-
pated increase in access to public coverage. Although the crowd-out
experienced by younger spouses is temporary because they will
eventually age into Medicare themselves, continuous and adequate
insurance coverage is important for the near-elderly because they
are high users of health care and at a higher risk for health shocks
due to their age.
This paper has the advantage of using a regression discontinu-
ity design previously used once in the crowd-out literature to study
Medicaid-generated crowd-out (Card and Shore-Sheppard, 2004).
Medicares age-65 threshold lends itself to regression discontinu-
ity and the research design has been used to estimate the causal
impact of eligibility on the insurance coverage, health outcomes,
and labor market participation of Medicare-eligible individuals
(Card et al., 2008, 2009; Decker, 2005; Fairlie et al., 2012). Regres-
sion discontinuity has the advantage of relying on relatively weak
identifying assumptions that approach a randomized experiment;
therefore providing credible evidence that Medicare eligibility of
older spouses reduces the health insurance coverage of younger
spouses.
The younger spouses of Medicare-eligible individuals are a
group to watch as the Patient Protection and Affordable Care Act
(ACA) is implemented. The insurance coverage mandate combined
11
with new options for younger spouses (e.g. health insurance
exchanges and Medicaid expansions) will likely reduce the number
of younger spouses going uninsured. Given that relatively healthy
younger spouses are forgoing coverage, the addition of these peo-
ple to the risk pool will potentially reduce average costs in this
age range. Additionally, new coverage options for younger spouses
that are not tied to employment will reduce the incentive for older
spouses to remain in the workforce in order to provide health
insurance to the household. Post-ACA, older spouses may take up
Medicare at a higher rate and transitions in coverage for younger
spouses could increase when an older spouse turns 65.
As the baby boom generation ages into Medicare eligibility,
policymakers will undoubtedly propose changes to the benet
structure of the program in an effort to reduce costs while provid-
ing coverage to aging Americans. These results demonstrate that
changes to the Medicare program will not only affect eligible indi-
viduals, but may also impact their family members. Analyses of
proposed modications to the program should include assessments
of the impact on beneciaries as well as potential spillovers to the
spouses of Medicare beneciaries.
The next section provides an introduction to the Medicare pro-
gram with a discussion of how eligibility of one spouse may impact
an ineligible spouse. Section 3 presents the data, econometric spec-
ication and identication assumptions. Section 4 presents results
and Section 5 concludes.
2. Background
Medicare is government subsidized insurance available to
American citizens ages 65 years and older.3 Eligible individuals
have access to the four components of Medicare, each covering
a specic type of service. Part A covers hospital-related expenses
and is free for most enrollees. Part B covers doctor visits, outpa-
tient care, and other general services at a cost between $96 and
$115 per month.4 Medicare Parts A and B have relatively high cost
sharing requirements and do not cover certain benets including
dental and long term care. As a result, most beneciaries have
some form of supplemental insurance (e.g. employer retiree cov-
erage, a Medigap plan, or Medicaid) to ll in holes in the benet
package and reduce Medicares cost sharing burden (Kaiser Family
Foundation, 2010). An alternative to purchasing supplemental cov-
erage is Medicare Part C, also known as Medicare Advantage.
Medicare Part C is comprised of government subsidized, privately-
administered plans that offer more comprehensive benets than
Parts A and B alone. In 2006, a prescription drug insurance plan
was introduced as Medicare Part D.
An older spouse deciding to take up Medicare coverage rep-
resents a tradeoff for families between individual-based and
household-based coverage. Using the example of a Medicare-
eligible husband with a Medicare-ineligible wife, let us consider
several cases. First, a wife who provided her own health insur-
ance prior to her husbands Medicare eligibility will experience no
required change in coverage when her husband turns 65. Second,
if the Medicare-eligible husband is the primary provider of health
insurance to the family, then he can continue to provide health
insurance for the ineligible wife, the wife can purchase insurance
3
Some individuals under age 65 who are disabled or have certain diseases also
qualify for Medicare.
4
Enrollment in Medicare Parts A and B at age 65 is automatic for individuals
receiving Social Security benets. Individuals not receiving Social Security benets
at age 65 must contact the Social Security Administration in order to enroll in the
program. Those automatically enrolled due to receipt of Social Security benets have
the right to refuse Part B coverage.
12 A. Witman / Journal of Health Economics 40 (2015) 1025

on her own, or she could go uninsured. Options for obtaining health Table 1
Selected summary statistics, younger spouses.
insurance for the ineligible wife include enrolling in a sponsored
plan from her own employer, purchasing directly from an insurance Women Men
company in the non-group market, and qualifying for means-tested Panel A: demographic characteristics
government insurance programs.5 % of sample 80.15 19.85
Medicare eligibility is both an income and a price shock to Age 58.81 59.78
the household. While Medicare is relatively low cost for the older Age gap between spouses 5.00 3.40

spouse, the cost of insurance coverage for the younger spouse could Education
increase if the household moves from group to individual coverage. % less than high school 14.91 18.17
% high school 37.88 30.06
An individual-based, private plan for the younger spouse may cost
% some college 24.69 23.55
more on a per-person basis than previously held group coverage for % more than college 22.52 28.22
both spouses. This increase in cost for a younger spouse can result
Race
in a reduction or elimination of health insurance coverage.
% White 83.41 82.37
The welfare implications of an older spouses Medicare eligi- % Black 6.32 6.93
bility vary depending on the choice of household model.6 Under % Hispanic 3.86 3.03
a unitary framework in which couples maximize a single house- % Other 6.41 7.67
hold utility function, revealed preference dictates that a household N 67,554 16,527
choosing to put the older spouse on Medicare will be better off even
in the case of the younger spouse losing coverage. However, house- Panel B: insurance characteristics prior to spousal Medicare eligibility
hold bargaining models generate ambiguous predictions for the % insured 91.16 91.08
% government coverage 8.66 12.32
utility of the younger spouse. While the income effect experienced
% private coverage 85.32 82.72
by the older spouse has the potential to make both spouses better % 2+ sources of coverage 8.75 9.06
off, the increase in the older spouses bargaining power could result
Source of private coverage
in an equilibrium with lower utility for the younger partner. The % self (work) 33.52 57.71
possibility that Medicares age-based eligibility structure is leaving % self (non-work) 3.34 5.44
younger spouses worse off merits a discussion of whether or not % spouse (work) 48.33 21.19
policy interventions are warranted, beginning with an analysis of % spouse (non-work) 4.05 1.04
the ACAs impact on younger spouses in the Medicare eligibility Note: Summary statistics are from pooled 1993 to 2011 NHIS data. Panel A presents
gap. demographic characteristics of younger husbands and wives whose older spouse
is within ten years of Medicare eligibility (between 55 and 75 years old). Panel B
Households can perfectly anticipate Medicare eligibility, mean-
restricts the sample to households with both a husband and a wife that are below
ing that the younger spouse could adjust their insurance coverage age 65.
prior to the older spouse turning 65. For younger spouses who
can obtain employer coverage, anticipatory adjustments would
most likely be made during the employers annual open enroll- survey of insurance coverage and health. The rst year in the analy-
ment period. Medicare eligibility of a spouse is a qualifying event sis is 1993 because several key insurance variables enter the survey
that allows a younger spouse to reduce coverage or cancel a group in the that year.
health insurance plan any time of year; however, increases in cov- The sample contains married households with an older spouse
erage due to spousal Medicare eligibility are not permitted outside within 10 years of Medicare eligibility (i.e. between 55 and 75 years
of the open enrollment period.7 If a younger wife provides health old). I construct the sample by rst selecting married households.
insurance to the couple from her employer plan, she can remove her Next, I choose couples with an older spouse that is between the
husband from the plan immediately when he turns 65 and main- ages of 55 and 75. Finally, I exclude couples that were born in the
tain her own coverage. Alternatively, if the older husband provides same quarter.8 Same-aged couples are excluded from the sample
insurance coverage through his employer, the younger wife would because they effectively age into Medicare at the same time and the
need to sign up for her own employers coverage during the open identication strategy rests on one spouse aging into Medicare rst.
enrollment period preceding eligibility of her husband to avoid a There are two types of households in the sample: older husband,
lapse in coverage. I investigate this possibility later in the paper younger wife couples and older wife, younger husband couples. I
and nd some evidence that households are making anticipatory focus on the outcomes of the younger spouse in each of these house-
adjustments in coverage. hold types, but will control for characteristics of both spouses in
the regressions. The nal sample includes 84,081 younger spouses,
comprised of 67,554 younger wives and 16,527 younger husbands.
3. Data and methodology
Table 1 presents selected summary statistics by gender for
younger spouses. The rst column includes demographic and
3.1. Data
insurance characteristics for women who are younger than their
husband and the second column includes the same information
I use data from the National Health Interview Survey (NHIS)
for men who are younger than their wife. Row 1 shows that 80% of
for years 19932011 (National Center for Health Statistics, 2012).
households in the sample are comprised of older husband, younger
The NHIS is a cross-sectional, nationally representative household
wife couples and 20% of households are comprised of older wife,
younger husband households. The average age for younger wives
and husband is 58.81 years and 59.78 years, respectively.9 The age
5
Beginning on October 1, 2013, individuals could apply for coverage using the gap between spouses is larger for older husband, younger wife
health insurance marketplaces established by the Affordable Care Act.
6
For a survey of theories of the family see Bergstrom (1995). Bargaining models
are generally more accepted than unitary models of household behavior. Seminal
8
papers in the household bargaining literature include Manser and Brown (1980), Age is measured in quarters in the NHIS.
9
McElroy and Horney (1981), Woolley (1988), and Lundberg and Pollak (1991). No restrictions are made on the age of the younger spouse. The age of younger
7
Internal Revenue Service Regulation 1.125-4(e). spouses ranges from 19 to 74.75 years old. Fig. A1 plots the age cumulative
 A. Witman / Journal of Health Economics 40 (2015) 1025
couples than for older wife, younger husband couples. On aver-
age, a younger wife will have 5 years after her husband turns 65
until she is Medicare-eligible while the average Medicare eligibility
gap for a husband that is younger than his wife is over 3 years. The
mean age difference is driven up by a small fraction of households
with a large age gap between spouses; the median age difference
for each group is smaller at 3.75 and 2 years, respectively.
Panel B of Table 1 limits the sample to households with both
spouses under age 65 to show insurance coverage of younger
spouses prior to Medicare eligibility of an older spouse.10 Insurance
variables are point-in-time measures of coverage at the time of the
NHIS interview. Among younger spouses, 91% report being insured.
Insurance coverage can be obtained through a government pro-
gram such as Medicare, Medicaid, or Veterans Insurance or through
a private health insurance plan. Private coverage can be purchased
through work or non-work sources and can be in an individuals
own name or under a spouses name and is categorized into four
sources: (1) Self (Work) is an insurance plan from an individ-
uals own employer, (2) Spouse (Work) is a plan from a spouses
employer, (3) Spouse (Non-Work) is a plan obtained from a non-
work source under a spouses name, and (4) Self (Non-Work) is
a non-work plan an individuals own name. Non-work insurance
is primarily comprised non-group coverage.11,12 The husbands
employer is the primary source of coverage for the majority of
households regardless of whether the male is the older or younger
spouse. Younger wives are most frequently covered via a spouses
work plan (48%), followed by a plan in the womans own name from
work (34%). Conversely, men are most often covered via their own
work plan (58%) followed by a wifes work plan (21%). The fraction
of younger spouses with non-work coverage is low.
3.2. Methodology
In order to test whether Medicare eligibility of an older spouse
impacts the health insurance coverage of a Medicare-ineligible
spouse, I compare younger spouses with an older spouse on either
side of the age 65 Medicare eligibility threshold. If individuals are
similar on either side of the threshold, the difference in outcomes
can be interpreted as the causal impact of spousal Medicare eli-
gibility. The empirical analysis is completed using a reduced-form
regression discontinuity design:
Yij = 0 + 1 Spouse65i + 2 Self 65i + 3 Xij + h(Agei , Agej ) + ij ground characteristics change at the threshold. Second, I plot the
(1)
where Yij is a point-in-time measure of the insurance coverage of
younger spouse i with older spouse j. Yij is a dummy variable indi-
cating whether or not i has a particular type of insurance coverage.
Spouse65i is a dummy variable for Medicare eligibility of is spouse,
set to 1 if is spouse is 65 or older and 0 otherwise. Self65i is a dummy
for Medicare eligibility of the younger spouse i, set to 1 if i is 65
distribution function, demonstrating that the majority of younger spouses are above
50 years old.
10
Categories of insurance coverage in Table 1 are not mutually exclusive because
some individuals report multiple forms of coverage.
11
Private coverage from non-employment sources includes plans purchased in
the non-group market, school plans, and plans obtained through local government
or community programs. Approximately 90% of private, non-work coverage for
Medicare-ineligible individuals in the sample are non-group plans.
12
The NHIS questionnaire asks whether a specic plan is in a persons own name
or a family members name. It is not possible to determine if the family member is
a spouse. It is reasonable to assume that for the majority of married households in
this age range, coverage by a family member is spousal coverage.
13
or older and 0 otherwise.13 Xij includes controls for the younger
spouses education, race, and an early Social Security eligibility
dummy for both the older and younger spouse.14 The specication
includes a quadratic in both spouses ages (h(Agei , Agej )) that is fully
interacted with post-65 age dummies (one for the older spouse and
one for the younger spouse) to allow the functional form to vary
on either side of the age 65 cutoff. Survey year and region xed
effects are included. Lastly, ij is an error term that accounts for the
impact of unobserved characteristics on the outcome of interest
and standard errors are clustered by age of the older spouse. Eq. (1)
will be estimated separately for men and women because spousal
Medicare eligibility has heterogeneous effects by gender.
The coefcient 1 is the effect of interest, measuring the impact
of an older spouse turning 65 on the younger spouses insurance
coverage.15 1 compares the insurance of younger spouses with a
partner that has recently turned 65 to the insurance coverage of
younger spouses with a partner that is just below age 65. Thus,
1 measures the effect of turning 65 relative to a pre-65 control
group that knows with certainty when they will enter eligibility
for Medicare. In contrast to regression discontinuity designs that
compare treated and untreated groups, age-based treatment makes
a comparison between those who are treated and those who will
soon be treated.16
3.3. Identication
The regression discontinuity research design is compelling
because it overcomes the omitted variable bias problem given a
modest set of assumptions. Moreover, some of the assumptions
are partially testable. Interpretation of the effects of Medicare eli-
gibility on outcomes as causal requires two assumptions (Lee and
Lemieux, 2009). First, the conditional expectation functions of the
potential outcomes must be continuous with respect to age across
the Medicare eligibility threshold, which is analogous to saying
that in the absence of treatment at age 65 outcomes would trend
smoothly. There are two standard test for violation of the continu-
ity assumption. First, there should be no changes in other variables
across the threshold if this assumption holds. I test for discontinu-
ities in education, race, labor market participation, health status,
and Social Security receipt and present the results in Fig. A3 and
Table A2. These estimates are all statistically insignicant except
for a decline in the fraction of the sample that is white, which is
likely due to sampling error given that none of the other back-
density function of the running variable (spousal age) and test for
a discontinuity at the age-65 threshold of Medicare eligibility per
McCrary (2008) in Fig. A4. I nd no evidence of violation of the
continuity assumption with this second test.
13
The coefcient 2 on Self65i gives the impact of a younger spouses own Medi-
care eligibility on their own insurance coverage. This is the effect measured in Card
et al. (2008, 2009) and is not the focus of this paper. These coefcients are presented
in Table A3.
14
Individuals have the option of taking Social Security benets as early as age 62.
The Social Security eligibility dummy equals 1 if the individual is greater than 62
and 0 otherwise.
15
1 captures the reduced-form effect of spousal Medicare eligibility rather than
take up. Fig. A2 shows the rst-stage graph and estimate of 0.64 (i.e., the effect of an
older spouse turning 65 on take up of Medicare). Instrumental variables estimates
of spousal Medicare take up are 1 /0.64.
16
Age-based treatment is common in regression discontinuity settings. Other
examples include the minimum legal drinking age, age notches in the calculation of
public benets, the age 19 limit on parental insurance coverage, and other papers
using the age 65 Medicare eligibility threshold (Decker and Rapaport, 2002; Card
and Shore-Sheppard, 2004; Card et al., 2008, 2009; Lemieux and Milligan, 2008;
Carpenter and Dobkin, 2009; Yoruk and Yoruk, 2011; Shigeoka, 2014).
14 A. Witman / Journal of Health Economics 40 (2015) 1025

Secondly, the monotonicity assumption requires that crossing
the age 65 threshold cannot cause some people to take up Medi-
care coverage and others to reject it. The monotonicity assumption
would be violated if the older spouse j had Medicare coverage
before turning 65 but dropped it upon reaching the 65 threshold.
Although I cannot test whether this assumption holds given the
cross-sectional nature of the data, it is unlikely that a person who
qualied for Medicare prior to age 65 due to disability or illness
would drop coverage specically because they turned 65.
4. Results
4.1. Main ndings
I begin with a graphical depiction of the insurance coverage of
the younger spouse as a function of the older spouses Medicare eli-
gibility in Fig. 1. Panels (a) and (b) plot the average rate of insurance
coverage and private coverage for younger wives. Panels (c) and (d)
plot the average rate of insurance coverage and private coverage
for younger husbands. The horizontal axis is the older spouses age
in quarters and each graph contains a vertical line at the spousal
age of 65 to indicate Medicare eligibility. Solid curves show the
quadratic t of the raw means on either side of the Medicare eligi-
bility cutoff. A dashed line indicates the fraction of younger spouses
that are age-eligible for Medicare, corresponding to the right axis
of each graph.
There is no evidence of crowd-out among Medicare-ineligible
wives in panels (a) and (b) of Fig. 1, with no clear discontinuity in
insurance coverage or private coverage at the spousal age of 65.
Table 2 presents estimates of discontinuities in insurance cover-
age upon spousal Medicare eligibility corresponding to Fig. 1 using
Eq. (1). The top row of the table lists the dependent variable for
each regression. The column labeled Mean presents the fraction
of younger spouses with the given coverage in the two years prior to
the older spouses Medicare eligibility. The estimated discontinu-
ities and standard errors are presented under the column labeled
RD. The results for women are presented in Panel A. The frac-
tion of women with insurance coverage falls by 1.24 percentage
points just after a husband turns 65, but the point estimate does not
reach statistical signicance. Husband Medicare eligibility causes

Fig. 1. Health insurance coverage of younger spouses. Note: Graphs show the reduced form effect of spousal Medicare eligibility on the health insurance coverage of the
younger spouse. Scatterplots are means of the raw data and solid lines are quadratic polynomial ts. The dashed line shows the fraction younger spouses eligible for Medicare
using the right axis.
 A. Witman / Journal of Health Economics 40 (2015) 1025 15

Table 2
The impact of spousal Medicare eligibility on insurance coverage of the younger spouse.

Insured Government Private 2+ plans

Mean RD Mean RD Mean RD Mean RD

Panel A: women
Spouse65 91.54 1.24 11.32 0.69 83.77 0.41 8.75 0.08
(0.80) (0.67) (1.14) (0.66)

Panel B: men
Spouse65 92.10 3.49* 17.63 0.52 80.50 4.73* 9.69 1.98
(1.47) (2.01) (2.12) (1.74)

Note: Table presents reduced form estimates of spousal Medicare eligibility on the insurance coverage of a younger spouse using NHIS data from 1993 to 2011. The dependent
variable in each regression is listed above the columns labeled Mean and RD. Entries in the Mean column are the fraction of younger spouses with the given type of
coverage in the two years prior to an older spouses Medicare eligibility. Regressions include controls for race and education of the younger spouse, quadratics in the age
of both spouses separately interacted with dummy variables for Medicare eligibility of the younger and older spouse at age 65, dummy variables for both spouses Social
Security eligibility at age 62, region, and year xed effects. Standard errors are clustered by spouse age and are presented in parentheses.
*
p < 0.05.
**
p < 0.01.
***
p < 0.001.

Fig. 2. Sources of private coverage of younger spouses. Note: Graphs show the reduced form effect of spousal Medicare eligibility on the source of private health insurance
for the younger spouse. Scatterplots are means of the raw data and solid lines are quadratic polynomial ts. The dashed line shows the fraction younger spouses eligible for
Medicare using the right axis.
16 A. Witman / Journal of Health Economics 40 (2015) 1025
no changes in the fraction of women with government coverage,
private coverage, or multiple sources of coverage. On the surface,
husband Medicare eligibility does not appear to cause a reduction
in the fraction of women with health insurance, which is reassuring
given that women are more likely to be insured via their husband
and are less likely to have an alternate source of coverage through
their own employer.
Medicare-ineligible men, on the other hand, experience reduc-
tions in coverage when their wife turns 65. In panels (c) and (d) of
Fig. 1, the fraction of men reporting any insurance coverage and pri-
vate coverage falls discontinuously to the right of the spousal age
65 threshold. Panel B of Table 2 presents the estimated disconti-
nuities for younger husbands. The fraction of men who are insured
falls by 3.49 percentage points and the fraction of men with private
coverage falls 4.73 percentage points when an older wife turns 65.
This is a 44% increase in the number of men who are uninsured and
a 6% decrease in private coverage. The approximately 2 percent-
age point difference between the fall in private coverage and any
coverage is comprised of a fall in multiple forms of coverage for
younger husbands.
The asymmetry of ndings for men and women is striking and
possible explanations will be explored at the end of this section.
First, I investigate private coverage more closely by focusing on the
source of private health insurance.
Fig. 2 plots the sources of private insurance coverage for younger
spouses with corresponding estimates presented in Table 3. For
women, spousal Medicare eligibility is associated with a movement
out of coverage under a husbands name and into coverage in the
wifes own name. Panels (a) and (b) of Fig. 2 show insurance cover-
age of younger wives in a husbands name and in a wifes own name,
respectively. Both work and non-work coverage from an older hus-
band fall discontinuously when a husband turns 65. As reported in
Panel A of Table 3, the largest decline in coverage is from a hus-
bands employer with an estimated reduction of 8.04 percentage
points compared to a 2.39 percentage point reduction in spousal
non-work coverage.
Wives offset the decline in private coverage from a husband
by increasing coverage in their own name, as shown in panel
(b) of Fig. 2. The estimated increase in coverage from a wifes
own employer is 4.62 percentage points and the increase non-
work coverage in their own name is 4.48 percentage points as
shown in Table 3. Thus, women take up insurance from both their
employer and non-work sources approximately equally when an
older husband reaches Medicare eligibility. As a result of this move-
ment, the composition of private coverage for Medicare-ineligible
women is different on either side of the husband Medicare eligibil-
ity threshold. Husband Medicare eligibility causes net movement
out of employer-provided plans and net movement into non-work
(usually non-group) plans for coverage. Historically, non-group
coverage has been considered less generous than employer cov-
erage, covering fewer benets and requiring higher out-of-pocket
costs.
Panels (c) and (d) of Fig. 2 plot the fraction of men with private
coverage in their wifes name and own name with corresponding
estimates presented in Panel B of Table 3. The fraction of husbands
covered via a wifes employer falls by 6.98 percentage points when
an older wife turns 65 and can be seen in Panel (a) of Fig. 2. Unlike
younger wives, younger husbands do not offset this fall in spousal
coverage by taking up coverage in their own name. There are no
changes in coverage in a husbands own name form either work or
non-work sources.
In summary, when an older spouse turns 65 younger spouses
experience a decline in coverage from the older partner. Younger
wives offset this fall in spousal coverage by taking up plans in their
own name from work and non-work sources. Younger husbands do
not offset the fall in spousal coverage and as a result, the fraction
of men without insurance increases.
There are several possible explanations for the higher propen-
sity of males to go uninsured. First, varying attitudes toward risk
could explain some component of mens willingness to forgo cov-
erage. Evidence from experimental economics demonstrates that
men are less risk averse than women (Eckel and Grossman, 2008;
Croson and Gneezy, 2009; Charness and Gneezy, 2012). Younger
husbands increased tolerance for risk is consistent with the results
in Cutler et al. (2008), who nd that individuals who have a higher
risk tolerance are more willing to go uninsured. Second, the avail-
able coverage options for younger husbands may be different than
for younger wives. The decline in coverage for men is driven by a
reduction in an older wifes work plan. The men who lose coverage
from their wife may not have access to an employer plan because
if they did, they likely would have already been the provider of
health insurance to the family. Additionally, gender asymmetries
in access and cost in the non-group market may explain the het-
erogeneous results. Men in this age range are on average in worse
health than women, suggesting that men may have been declined
coverage at a higher rate than women.17 Additionally, coverage in
the non-group market for men over 60 was also more expensive
than for women prior to the Affordable Care Act. In addition to dif-
ferential preferences for coverage across men and women, forces in
the non-group market may contribute to mens higher propensity
to lose coverage in the Medicare eligibility gap.
4.2. Heterogeneity
I now examine heterogeneity of effects across three margins:
completed education, labor force participation, and health status.18
While each of these categories is smooth across the spousal age 65
threshold, both labor market participation and health are declin-
ing functions of spousal (and own) age.19 As a consequence, the
bandwidth in a regression discontinuity design would ideally be
smaller than the bandwidth used in this paper. However, given the
sample size required for adequate power I estimate Eq. (1) with the
full bandwidth. The results provide interesting evidence that cer-
tain groups are more affected by the Medicare eligibility gap, but
should be interpreted with caution.
Table 4 presents coefcients from a separate regression for the
group listed in the rst column of the table.20 Panel A shows results
by education level. For younger wives, there is no differential effect
on insurance coverage or private coverage by education level when
a husband turns 65. However, education level does play a role for
men. Younger husbands with a high school or less education are
5.31 percentage points less likely to have insurance coverage after
a wife turns 65. This result is statistically different from the insignif-
icant point estimate for husbands with at least some college. Panel
B presents results by labor market participation status, which is
not a source of notable heterogeneity for younger wives. The fall
in insurance coverage for men who are working is 4.05 percentage
points, which is a larger estimated effect than the 2.46 percentage
point fall in coverage for non-working men, although not statisti-
cally different. Lastly, Panel C includes coefcients from estimates
17
The Affordable Care Act now prohibits denial of coverage based on preexisting
conditions.
18
Estimates by race are available upon request. The relatively small sample size
of non-white respondents makes identifying effects by race infeasible.
19
Table A2 presents estimated discontinuities in education, labor force participa-
tion, and health status at spousal Medicare eligibility. Accompanying graphs are in
Fig. A3.
20
Regressions are run separately for each group. Results are similar when the
groups are pooled and the Spouse65 variable is interacted with group categories.
 A. Witman / Journal of Health Economics 40 (2015) 1025 17

Fig. 3. Fraction of younger spouses with insurance coverage by subpopulation. Note: Graphs show the reduced form effect of spousal Medicare eligibility on insurance of
the younger spouse for different subpopulations. Scatterplots are means of the raw data and solid lines are quadratic polynomial ts.
18 A. Witman / Journal of Health Economics 40 (2015) 1025

Table 3
The impact of spousal Medicare eligibility on the source of private insurance coverage of the younger spouse.

Spouse (work) Spouse (non-work) Self (work) Self (non-work)

Mean RD Mean RD Mean RD Mean RD

Panel A: women
Spouse65 45.53 8.04*** 5.18 2.39*** 32.18 4.62** 4.46 4.48***
(1.66) (0.63) (1.54) (0.43)

Panel B: men
Spouse65 18.54 6.98** 1.21 0.49 55.13 0.45 7.51 0.69
(2.26) (0.73) (2.73) (1.36)

Note: Table presents reduced form estimates of spousal Medicare eligibility on the insurance coverage of a younger spouse using NHIS data from 1993 to 2011. The dependent
variable in each regression is listed above the columns labeled Mean and RD. Entries in the Mean column are the fraction of younger spouses with the given type of
coverage in the two years prior to an older spouses Medicare eligibility. Regressions include controls for race and education of the younger spouse, quadratics in the age
of both spouses separately interacted with dummy variables for Medicare eligibility of the younger and older spouse at age 65, dummy variables for both spouses Social
Security eligibility at age 62, region, and year xed effects. Standard errors are clustered by spouse age and are presented in parentheses.
*
p < 0.05.
**
p < 0.01.
***
p < 0.001.

by health status. Healthy wives are 1.10 percentage points less and have less than a high school education. This group of men may
likely to have insurance coverage and healthy husbands are 5.42 have jobs that do not offer health insurance coverage, may nd the
percentage points less likely to have insurance coverage after an cost of non-group coverage prohibitively high, and may not be able
older spouse turns 65. Results for insurance coverage of younger to afford to pay for treatment out of pocket after a health shock. The
spouses from Table 4 are presented graphically in Fig. 3. Affordable Care Acts insurance mandate may bring these relatively
A new result emerges from this analysis showing that a higher healthy individuals into the health insurance market, improving
fraction of healthy men and women are uninsured after a spouse the health of individuals in the risk pool and potentially driving
turns 65. The propensity for healthier individuals to go uninsured down average costs for this age group.
is evidence of adverse selection in the health insurance market,
which is a standard empirical nding summarized in Cutler and 4.3. Robustness checks
Zeckhauser (2000) and recently revisited by Einav et al. (2010).
Consequently, healthy individuals may be making an optimal In Table 5, I present several robustness checks of the main
choice to go uninsured if they do not forgo necessary care or expe- ndings. The main identifying assumption in a regression dis-
rience a health shock during the Medicare eligibility gap. However, continuity design is that the treatment and control group are
the men who are becoming uninsured also appear to be working comparable; therefore, removing demographic controls from the
regression should not affect the estimates. The rst row of Table 5
Table 4
presents the main estimates from Tables 2 and 3 using a second
The impact of spousal Medicare eligibility on insurance coverage of the younger order polynomial and controlling for background characteristics.
spouse, by selected characteristics of the younger spouse. The next row, labeled Second Order Polynomial, No Controls
presents results from regressions without background characteris-
Women Men
tics, including only a quadratic polynomial in the older spouses age
Insured Private Insured Private that differs on either side of the eligibility threshold, a dummy for
Panel A: high school or less and some college or more the older spouses eligibility, a dummy for the younger spouses
High school or less 1.25 0.17 5.31* 4.00 eligibility, year xed effects, and region xed effects. The point
(1.26) (1.88) (2.56) (3.05)
estimates are very similar to the main specications that include
Some college or more 1.15 0.81 1.39 5.12
(0.81) (1.28) (1.45) (2.62) demographic controls. One exception is that the point estimate on
government coverage for younger wives becomes statistically sig-
Panel B: working and not working
nicant. There is no reason to believe that government coverage
Working 0.28 0.41 4.05* 4.03
(1.01) (1.27) (1.87) (2.34) should fall discontinuously and this result is likely due to sampling
Not working 2.07 0.27 2.46 3.23 error.
(1.11) (1.61) (2.17) (3.73) Additionally, estimates can vary with the choice of polyno-
Panel C: health status poor/fair and good/very good/excellent mial. The main specication in this paper uses a second order
Poor/fair 1.30 4.88 0.90 2.94 polynomial, which is standard in papers using the NHIS and
(2.30) (3.34) (2.44) (6.43) a regression discontinuity design. Nevertheless, the third row
Good/verygood/excellent 1.10* 1.10 5.42** 7.98**
of Table 5 presents coefcients from regressions in which the
(0.54) (0.80) (1.61) (2.49)
age of the older and younger spouse is with a third degree
Note: Table presents reduced form estimates of spousal Medicare eligibility on the
polynomial. On the whole, varying the degree of polynomial
insurance coverage of a younger spouse using NHIS data from 1993 to 2011. Each
coefcient is from a separate regression for the group listed in the leftmost column
produces similar results as the main specication with few
of the table. Regressions include quadratics in the age of both spouses separately exceptions.
interacted with dummy variables for Medicare eligibility of the younger and older The regression discontinuity framework identies discrete
spouse at age 65, dummy variables for both spouses Social Security eligibility at age changes in coverage at the age 65 cutoff of spousal Medicare eligi-
62, region, and year xed effects. Standard errors are clustered by spouse age and
bility; however, households anticipating Medicare eligibility of an
are presented in parentheses.
*
p < 0.05. older spouse may adjust the coverage of a younger spouse ahead
**
p < 0.01. of time. Anticipatory adjustments will not be embodied in the esti-
***
p < 0.001. mated discontinuities in coverage, potentially biasing coefcients
 A. Witman / Journal of Health Economics 40 (2015) 1025 19

Table 5
Robustness checks.

Insured Gov Private 2+ plans Spouse (work) Spouse (non-work) Self (work) Self (non-work)

Panel A: women
2nd order polynomial (original estimates)
Spouse65 1.24 0.69 0.41 0.08 8.04*** 2.39*** 4.62** 4.48***
(0.80) (0.67) (1.14) (0.66) (1.66) (0.63) (1.54) (0.43)
2nd order polynomial, no controls
Spouse65 1.34 1.33* 0.42 0.90 7.21*** 2.89*** 3.99** 4.51***
(0.72) (0.58) (1.02) (0.61) (1.50) (0.64) (1.50) (0.42)
3rd order polynomial
Spouse65 0.40 0.54 1.01 0.09 7.35*** 2.59*** 5.12** 4.83***
(0.64) (0.73) (1.18) (0.67) (2.00) (0.75) (1.81) (0.49)
Donut RD
Spouse65 2.88* 0.20 2.18 0.86 10.80*** 2.19* 6.51** 3.78***
(1.09) (1.11) (1.35) (1.04) (1.58) (1.03) (1.95) (0.72)
Age gap > 1 year
Spouse65 0.84 0.65 0.37 0.24 7.60*** 2.38** 4.28* 4.54***
(0.87) (0.73) (1.25) (0.66) (1.76) (0.74) (1.81) (0.55)

Panel B: men
2nd order polynomial (original estimates)
Spouse65 3.49* 0.52 4.73* 1.98 6.98** 0.49 0.45 0.69
(1.47) (2.01) (2.12) (1.74) (2.26) (0.73) (2.73) (1.36)
2nd order polynomial, no controls
Spouse65 4.61 **
1.36 5.91** 3.88* 6.74** 0.28 0.19 0.81
(1.39) (2.10) (1.94) (1.66) (2.19) (0.68) (2.53) (1.31)
3rd order polynomial
Spouse65 3.15 0.68 2.20 0.26 6.56* 0.15 4.58 0.58
(1.65) (2.43) (2.32) (1.89) (2.79) (0.72) (2.49) (1.41)
Donut RD
Spouse65 4.52* 2.88 7.21* 1.66 9.38*** 1.17 2.25 1.38
(2.19) (2.65) (3.15) (2.69) (2.61) (1.29) (4.22) (2.42)
Age gap > 1 year
Spouse65 2.40 0.22 4.12 3.96* 7.58** 1.08 0.22 0.18
(2.09) (2.48) (2.55) (1.72) (2.79) (0.83) (3.34) (1.28)

Note: Table presents reduced form estimates of spousal Medicare eligibility on the insurance coverage of a younger spouse using NHIS data from 1993 to 2011. Each
coefcient is from a separate regression using the specication described in the leftmost column of the table. Standard errors are clustered by spouse age and are presented
in parentheses.
*
p < 0.05.
**
p < 0.01.
***
p < 0.001.

toward zero and underestimating the impact of spousal Medicare
eligibility. The donut regression discontinuity design removes
observations near the cutoff, generating estimates that embody
changes at and around the discontinuity (Barreca et al., 2011).21
In the fourth row of Panels A and B in Table 5, I re-estimate Eq. (1)
excluding couples with an older spouse within 6 months of Medi-
care eligibility, leaving a 1 year donut in the data. These estimates
include changes in insurance coverage that not only occur right at
age 65, but also 6 months before and after an older spouse is 65.
This exercise yields a statistically signicant 2.88 and 4.52 percent-
age point reduction in the fraction of younger wives and husbands
with insurance coverage. Overall, the point estimates from this
specication are slightly larger in absolute value than the main
estimates for both younger husbands and wives, indicating that
some adjustments in coverage are made before and after the older
spouse 65.
Lastly, I address the concern that the results could be con-
founded by couples close in age because the younger spouse ages
into Medicare just after the older spouse. In the fth row of Table 5,
I limit the sample to couples with at least a one year age gap
between the older and younger spouse. Although the standard
errors increase slightly, the point estimates are very similar to those
21
The donut design is most frequently cited as a method to account for heaping
in the data, but can also be used to address changes in behavior due to anticipation
of treatment as in Barcellos and Jacobson (2014).
presented in Tables 2 and 3. On the whole, results do not appear to
be driven by couples that are very close in age.
5. Discussion
As the baby boom generation begins to reach the age of Medi-
care eligibility, an increasing number of married households will
face a tradeoff between household-based coverage and individual-
based coverage, or may even consider forgoing coverage for
a Medicare-ineligible spouse. In this paper, I nd that Medicare
eligibility of an older spouse can crowd-out the insurance coverage
of a younger spouse. Younger husbands who are relatively healthy,
working, and have less than a high school education are less likely
to have insurance just after a wife turns 65. Relatively healthy
women are more likely to be uninsured after a husband turns
65; but the biggest change in coverage for women occurs on the
intensive margin. Private coverage in a husbands name falls and
coverage in a wifes own name increases with an increase in the
fraction of women insured via plans that can be considered less
generous.
Younger spouses of Medicare-eligible individuals will be one
group impacted by the changes in the insurance market after
the Patient Protection and Affordable Care Act (ACA). The health
insurance exchanges are a new source of coverage for younger
spouses and may be an attractive option given the projected decline
in premiums in the over-60 non-group market (OConnor, 2013)
combined with the premium subsidies available for households
between 138 and 400% of the poverty line. The evidence of adverse
20 A. Witman / Journal of Health Economics 40 (2015) 1025

selection in the under-65 market generated by Medicare eligibil-

ity of older spouses presented in this paper suggests that some
new entrants to the insurance market in this age range will be
relatively healthy. Importantly, the availability of an alternative
source of coverage for younger spouses means that older spouses
will be more likely to take up Medicare and will have less incen-
tive to remain in the workforce to provide health insurance for a
Medicare-ineligible spouse.
Medicare will undoubtedly remain a focal point of policy-
makers trying to balance providing adequate insurance coverage
for the growing number of elderly Americans with the need to curb
rapidly increasing health care costs. Because most Americans have
household-level coverage prior to Medicare eligibility, proposals
such as raising the age of eligibility to 67 or changing the benet
structure of the program will not only affect newly eligible individ-
uals, but also their family members. Analyses of proposed changes Fig. A2. First stage estimate. Note: The graph shows the fraction of older spouses
to the program should include assessments of the impact on ben- that report having Medicare coverage as a function of the age of the older spouse. The
eciaries as well as potential spillovers to the spouses of Medicare scatterplot is the raw data and the curves are quadratic polynomial ts estimated
beneciaries. separately on each side of the age-65 threshold. The rst-stage estimate of the take
up of Medicare at age 65 among older spouses is 63.67%.

Appendix A.

See Figs. A1A4 and Tables A1A3.

Fig. A1. Age distribution of younger spouse sample. Note: Graph shows the cumu-
lative distribution function of younger spouses age by gender.
 A. Witman / Journal of Health Economics 40 (2015) 1025 21

Fig. A3. Education, race, labor force participation, and health status of younger spouses. Note: Graphs show the fraction of younger spouses with each education level and
race plotted against the age of the older spouse. Scatterplots are means of the raw data and solid lines are quadratic polynomial ts.
22 A. Witman / Journal of Health Economics 40 (2015) 1025

Fig. A3. (Continued ).

 A. Witman / Journal of Health Economics 40 (2015) 1025 23

Fig. A4. Density of spousal age. Note: The graph plots the number of older spouses in the sample at each age to test for violation of the continuity assumption. The curves
are quadratic polynomial ts estimated separately on each side of the age-65 threshold. The estimated discontinuity at age 65 is not statistically signicant.

Table A1
Comparison of bandwidth and polynomial choices in regressions of insurance coverage on spousal and own Medicare eligibility.

Polynomial Panel A: women Panel B: men

24 quarters 32 quarters 40 quarters 24 quarters 32 quarters 40 quarters

First order
Spouse65 1.14 1.73** 1.65** 3.84** 3.23** 3.24**
(0.68) (0.59) (0.53) (1.31) (1.14) (1.06)
Self65 6.79*** 7.37*** 7.64*** 6.98*** 7.07*** 7.31***
(0.53) (0.49) (0.46) (1.41) (1.18) (1.07)

Second order
Spouse65 0.15 0.47 1.24 3.25 3.79* 3.49*
(0.71) (0.71) (0.80) (1.63) (1.54) (1.47)
Self65 6.10*** 6.88*** 7.33*** 6.14** 6.63*** 6.53***
(0.72) (0.62) (0.56) (1.91) (1.60) (1.44)

Third order
Spouse65 1.78** 0.65 0.40 2.57 2.92 3.15
(0.65) (0.65) (0.64) (1.76) (1.68) (1.65)
Self65 5.52*** 6.47*** 6.83*** 6.28* 5.93** 6.24**
(0.80) (0.74) (0.67) (2.57) (2.14) (1.91)

Fourth order
Spouse65 1.59* 1.64* 0.92 2.29 1.68 2.92
(0.72) (0.67) (0.68) (2.03) (1.89) (1.77)
Self65 5.54*** 6.18*** 6.54*** 5.28 4.85 5.18*
(0.98) (0.90) (0.85) (3.14) (2.68) (2.44)

Note: Table presents reduced form regression discontinuity estimates of spousal and own-Medicare eligibility on a dummy variable for own insurance coverage. Columns
report estimates using a bandwidth of 24, 32, and 40 quarters around the age of 65. Estimates are presented for rst, second, third, and fourth order polynomials in spousal
and own age. Regressions include and age polynomial interacted with dummy variables for own and spousal Medicare eligibility at age 65, own race and education controls,
dummy variables for own and spousal Social Security eligibility at age 62, region, and year xed effects. Standard errors are clustered by spouse age and are presented in
parentheses.
*
p < 0.05.
**
p < 0.01.
***
p < 0.001.
24 A. Witman / Journal of Health Economics 40 (2015) 1025

Table A2
Estimates of discontinuities in education, race, health status, labor force participation, and social security benets of younger spouses.

Less than high school High school Some college College or more

Spouse65 0.27 0.74 1.39 0.91

(1.11) (1.41) (0.95) (0.97)

White Black Hispanic Good/very good/excellent health

Spouse65 1.72* 0.89 0.59 0.01

(0.82) (0.65) (0.34) (1.20)

Working Hours worked Health insurance offered at work , Receive social security income

Spouse65 0.38 0.19 1.62 0.28

(1.05) (0.63) (1.24) (1.12)

Note: Table presents estimates of discontinuities upon an older spouse reaching Medicare eligibility at age 65. All dependent variables are point-in-time except for receipt
of Social Security income, which is during the previous calendar year. Regressions include quadratics in age interacted with dummy variables for own and spousal Medicare
eligibility at age 65, dummy variables for own and spousal Social Security eligibility at age 62, region, and year xed effects. Standard errors are clustered by spouse age and
are presented in parentheses.
*
p < 0.05.
**
p < 0.01.
***
p < 0.001.

Conditional on working.

Regression includes years 19972011.

Table A3
The impact of Medicare eligibility of the younger spouse on insurance coverage of the younger spouse.

Panel A: women

Insured Government Private 2+ plans

Mean RD Mean RD Mean RD Mean RD

Self65 91.36 7.33*** 15.60 67.34*** 82.78 8.10*** 10.74 49.18***

(0.56) (0.88) (0.91) (1.00)

Spouse (work) Spouse (non-work) Self (work) Self (non-work)

Mean RD Mean RD Mean RD Mean RD

Self65 36.98 3.93** 4.65 5.44*** 30.31 11.62*** 12.87 1.31

(1.29) (0.73) (0.96) (0.81)

Panel B: men

Insured Government Private 2+ plans

Mean RD Mean RD Mean RD Mean RD

Self65 91.81 6.53*** 22.60 58.23*** 78.19 10.00*** 12.48 38.93***

(1.44) (2.14) (2.28) (2.35)

Spouse (work) Spouse (non-work) Self (work) Self (non-work)

Mean RD Mean RD Mean RD Mean RD

Self65 14.43 0.04 1.11 1.12 53.06 17.27*** 10.84 5.96**

(1.59) (0.76) (2.43) (2.01)

Note: Table presents reduced form estimates of spousal Medicare eligibility and own Medicare eligibility on own insurance coverage using NHIS data from 1993 to 2011.
Entries in the Mean column are means of own insurance coverage for the two years prior to spousal Medicare eligibility (rst row of each panel) or own Medicare eligibility
(second row of each panel). Regressions include quadratics in age interacted with dummy variables for own and spousal Medicare eligibility at age 65, own race and education
controls, dummy variables for own and spousal Social Security eligibility at age 62, region, and year xed effects. Standard errors are clustered by spouse age and are presented
in parentheses.
*
p < 0.05.
**
p < 0.01.
***
p < 0.001.
 A. Witman / Journal of Health Economics 40 (2015) 1025
References
Barcellos, S.H., Jacobson, M., March 2014. The Effects of Medicare on Medical Expen-
diture Risk and Financial Strain. Working Paper 19954. National Bureau of
Economic Research.
Barreca, A.I., Lindo, J.M., Waddell, G.R., September 2011. Heaping-Induced Bias in
Regression-Discontinuity Designs. NBER Working Paper 17408.
Bergstrom, T., 1995. A survey of theories of the family. In: Rosenzweig, M., Stark,
O. (Eds.), Handbook of Population and Family Economics. Elsevier, Amsterdam,
pp. 2159.
Blumberg, L.J., Dubay, L., Norton, S.A., 2000. Did the Medicaid expansions for chil-
dren displace private insurance? An analysis using the SIPP. Journal of Health
Economics 19 (1), 3360.
Boyle, M.A., Lahey, J.N., May 2012. Spousal Labor Market Effects from Government
Health Insurance: Evidence from a Veterans Affairs Expansion. Working Paper,
Center for Retirement Research at Boston College wp2012-16. Center for Retire-
ment Research.
Boyle, M.A., Lahey, J.N., 2010. Health insurance and the labor supply decisions of
older workers: evidence from U.S. Department of Veterans Affairs Expansion.
Journal of Public Economics 94, 467478.
Busch, S.H., Duchovny, N., 2005. Family coverage expansions: impact on insurance
coverage and health care utilization of parents. Journal of Health Economics 24
(5), 876890.
Card, D., Shore-Sheppard, L.D., 2004. Using discontinuous eligibility rules to iden-
tify the effects of the federal Medicaid expansions on low-income children. The
Review of Economics and Statistics 86 (3), 752766.
Card, D., Dobkin, C., Maestas, N., 2008. The impact of nearly universal insurance
coverage on health care utilization: evidence from Medicare. The American
Economic Review 98 (5), 22422258.
Card, D., Dobkin, C., Maestas, N., 2009. Does Medicare save lives? The Quarterly
Journal of Economics 124 (2), 597636.
Carpenter, C., Dobkin, C., 2009. The effect of alcohol consumption on mortality:
regression discontinuity evidence from the minimum drinking age. American
Economic Journal: Applied Economics 1 (1), 164182.
Charness, G., Gneezy, U., 2012. Strong evidence for gender differences in risk taking.
Journal of Economic Behavior & Organization 83 (1), 5058.
Croson, R., Gneezy, U., 2009 June. Gender differences in preferences. Journal of
Economic Literature 47 (2), 448474.
Cubanski, J., Huang, J., Damico, A., Jacobson, G., Neuman, T., 2010. Medicare Chart-
book, 4th ed. Kaiser Family Foundation, Menlo Park.
Cutler, D.M., Finkelstein, A., McGarry, K., 2008. Preference heterogeneity and insur-
ance markets: explaining a puzzle of insurance. American Economic Review 98
(2), 157162.
Cutler, D.M., Gruber, J., 1996. Does public insurance crowd out private insurance?
The Quarterly Journal of Economics 111 (2), 391430.
Cutler, D.M., Zeckhauser, R.J., 2000. Chapter 11: The anatomy of health insurance.
In: Culyer, A.J., Newhouse, J.P. (Eds.), In: Handbook of Health Economics, vol. 1,
Part A. Elsevier, Amsterdam, pp. 563643.
Decker, S., Rapaport, C., January 2002. Medicare and Disparities in Womens Health.
Working Paper 8761. National Bureau of Economic Research.
Decker, S.L., 2005. Medicare and the health of women with breast cancer. Journal of
Human Resources XL (4), 948968.
Dubay, L.C., Kenney, G.M., 1996. The effects of medicaid expansions on insurance
coverage of children. The Future of Children 6 (1), 152161.
Dubay, L.C., Kenney, G.M., 1997. Did medicaid expansions for pregnant women
crowd-out private insurance? Health Affairs 16 (1), 185193.
Duggan, M., Morton, F.S., 2010. The effect of Medicare Part D on pharmaceutical
prices and utilization. American Economic Review 100 (March (1)), 590607.
Eckel, C.C., Grossman, P.J., 2008. Men, women and risk aversion: experimental evi-
dence. In: Plott, C.R., Smith, V.L. (Eds.), In: Handbook of Experimental Economic
Results, vol. 1, Part 7, 1st ed. Elsevier, Amsterdam, pp. 10611073 (Chapter 113).
25
Einav, L., Finkelstein, A., Cullen, M.R., 2010. Estimating welfare in insurance mar-
kets using variation in prices. The Quarterly Journal of Economics 125 (3),
877921.
Fairlie, R.W., Kapur, K., Gates, S., April 2012. Job Lock: Evidence from a Regression
Discontinuity Design. Working Papers 201215. School of Economics, University
College Dublin.
Finkelstein, A., McKnight, R., 2008. What did Medicare do? The initial impact of
Medicare on mortality and out of pocket medical spending. Journal of Public
Economics 92 (7), 16441668.
Gruber, J., Simon, K., 2008. Crowd-out 10 years later: have recent public insurance
expansions crowded out private health insurance? Journal of Health Economics
27 (2), 201217.
Kaiser Family Foundation, November 2012. Medicare at a Glance. Number 1066-15.
Lee, D.S., Lemieux, T., February 2009. Regression Discontinuity Designs in Eco-
nomics. Working Paper 14723. National Bureau of Economic Research.
Lemieux, T., Milligan, K., 2008. Incentive effects of social assistance: a regres-
sion discontinuity approach. Journal of Econometrics 142 (2), 807828,
https://ideas.repec.org/a/eee/econom/v142y2008i2p807-828.html
Lichtenberg, F.R., 2002. The effects of Medicare on health care utilization and out-
comes. Forum for Health Economics & Policy 5 (Frontiers in Health Policy
Research), Article 3.
Lundberg, S., Pollak, R.A., 1991. Separate Spheres Bargaining and the Marriage Mar-
ket. Discussion Papers in Economics at the University of Washington 91-08.
Department of Economics at the University of Washington.
Manser, M., Brown, M., 1980. Marriage and household decision-making: a bargain-
ing analysis. International Economic Review 21, 3144.
McCrary, J., 2008. Manipulation of the running variable in the regression dis-
continuity design: a density test. Journal of Econometrics 142 (2), 698
714.
McElroy, M.B., Horney, M.J., 1981. Nash-bargained household decisions: toward
a generalization of the theory of demand. International Economic Review 22,
333349.
McWilliams, J.M., Meara, E., Zaslavsky, A.M., Ayanian, J.Z., 2007. Use of health ser-
vices by previously uninsured Medicare beneciaries. New England Journal of
Medicine 357 (2), 143153.
McWilliams, M.J., Zaslavsky, A., Meara, E., Ayanian, J., 2003. Impact of Medicare cov-
erage on basic clinical services for previously uninsured adults. Journal of the
American Medical Association 290 (6), 757764.
National Center for Health Statistics, 2012. Data File Documentation, National
Health Interview Survey, [Machine-Readable Data File and Documentation].
National Center for Health Statistics, Centers for Disease Control and Prevention,
Hyattsville, MD.
OConnor, J.T., 2013. Comprehensive Assessment of ACA Factors That Will Affect
Individual Market Premiums in 2014. Technical Report. Milliman.
Sasso, A.T.L., Buchmueller, T.C., 2004. The effect of the state childrens health insur-
ance program on health insurance coverage. Journal of Health Economics 23 (5),
10591082.
Shigeoka, H., 2014. The effect of patient cost sharing on utilization, health, and risk
protection. American Economic Review 104 (7), 21522184.
Shore-Sheppard, L., Buchmueller, T.C., Jensen, G.A., 2000. Medicaid and crowding out
of private insurance: a re-examination using rm level data. Journal of Health
Economics 19 (1), 6191.
Woolley, F., 1988. A Non-Cooperative Model of Family Decision Making. Working
Paper. TIDI/125. London School of Economics.
Yazici, E., Kaestner, R., 2000. Medicaid expansions and the crowding out of private
health insurance among children. Inquiry 37 (Spring (1)), 2332.
Yoruk, B.K., Yoruk, C.E., 2011. The impact of minimum legal drinking age laws on
alcohol consumption, smoking, and marijuana use: evidence from a regression
discontinuity design using exact date of birth. Journal of Health Economics 30
(4), 740752.
 Journal of Health Economics 40 (2015) 2639

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

The heterogeneity of concentrated prescribing behavior: Theory and

evidence from antipsychotics
Ernst R. Berndt a,b , Robert S. Gibbons c,b , Anton Kolotilin d , Anna Levine Taub e,
a
Louis E. Seley Professor in Applied Economics, MIT Sloan School of Management, United States
b
NBER, United States
c
Sloan Distinguished Professor of Management, MIT Sloan School of Management, United States
d
University of New South Wales, School of Economics, Australia
e
Cornerstone Research, United States

a r t i c l e i n f o a b s t r a c t

Article history: We present two new ndings based on annual antipsychotic US prescribing data from IMS Health on 2867
Received 24 January 2013 psychiatrists who wrote 50 or more prescriptions in 2007. First, many of these psychiatrists have prescrip-
Received in revised form tion patterns that are statistically signicantly different than random draws from national market shares
18 September 2014
for prescriptions by psychiatrists. For example, many have prescription patterns that are signicantly
Accepted 14 November 2014
Available online 12 December 2014
more concentrated than such draws. Second, among psychiatrists who are the most concentrated, differ-
ent prescribers often concentrate on distinct drugs. Motivated by these two ndings, we then construct
a model of physician learning-by-doing that ts these facts and generates two further predictions: both
JEL classication:
I10
concentration (on one or a few drugs) and deviation (from the prescription patterns of others) should be
I11 smaller for high-volume physicians. We nd empirical support for these predictions. Furthermore, our
D80 model outperforms an alternative theory concerning detailing by pharmaceutical representatives.
D83 2014 Elsevier B.V. All rights reserved.

Keywords:
Antipsychotic prescriptions
Pharmaceutical concentration
Physician learning

1. Introduction
1.1. Motivation and overview
Consider a physician seeing a patient with a conrmed diagnosis
for which several alternative pharmaceutical treatments are avail-
able. Suppose that, given the clinical evidence, patient response to
a given treatment is idiosyncratic and unpredictable in terms of
both efcacy and side effects. What treatment algorithms might
the physician employ to learn about the efcacy and tolerability of
the alternative drug therapies for this and future similar patients?
One possibility is for the physician to concentrate her prescrib-
ing behaviorin the extreme, on just one drug. By observing this
and future patients responses to that drug, the physician can learn
by doing, thereafter exploiting her accumulated knowledge about
Corresponding author at: 1000 El Camino Real, Suite 250, Menlo Park, CA, 94025,
United States. Tel.: +1 650 470 7014.
E-mail address: ataub@cornerstone.com (A.L. Taub).
this drug. For example, the physician will learn how to counsel
patients on the efcacy and side-effect responses they might expe-
rience, possible interactions with other drugs, and the best time of
day to take the drug; in addition, she will learn how to adjust the
dosage depending on patients factors such as smoking behavior,
thereby improving patient outcomes and engaging the patient in
adherence and symptom remission.
Alternatively, the physician might diversify her prescriptions
across several drugs, hoping to deliver the best match between
different drugs and current and future similar patients. Speci-
cally, based on information from a patients history, familiarity with
the existing scientic and clinical literature, conversations with
fellow medical professionals in the local and larger geographical
community, and perhaps interactions with pharmaceutical sales
representatives, the physician might select the therapy that a priori
appears to be the best match with the particular patients charac-
teristics (even if the physician is less able to counsel the patient on
the side effects, interactions, and other aspects of the drug).
In short, the physician can learn from exploiting or explor-
ing, concentrating or diversifying. In addition, physicians with

http://dx.doi.org/10.1016/j.jhealeco.2014.11.003
0167-6296/ 2014 Elsevier B.V. All rights reserved.
 E.R. Berndt et al. / Journal of Health Economics 40 (2015) 2639
concentrated prescriptions may converge (exhibiting near una-
nimity on the choice of a favorite drug) or diverge (with different
physicians concentrating on distinct drugs). We explore these
issues using data on a particular therapeutic class of drugs known as
antipsychotics. Later in this introduction, we provide a brief back-
ground on the history of antipsychotic drugs and the illnesses they
treat.
In the body of the paper, we begin by describing our data
on the antipsychotic prescriptions written by psychiatrists who
wrote 50 or more prescriptions in 2007. We then present two
new ndings. First, many of these psychiatrists have prescrip-
tion patterns that are statistically signicantly different than
random draws from national market shares for prescriptions
by psychiatrists. For example, many have prescription patterns
that are signicantly more concentrated than such draws. (As
a stark illustration, in the data we describe and analyze below,
the average share of antipsychotic prescriptions written for the
psychiatrists most-prescribed drug is 41%.) Second, but less
formally, among psychiatrists who are the most concentrated,
different prescribers concentrate on distinct drugsthe het-
erogeneous concentration in our title. In short, even in our
sample of psychiatrists with high prescription volumes, national
market shares for this population do not reect homogeneous
physicians each prescribing drugs in proportions approximating
national shares, but rather the aggregate of heterogeneous physi-
cians many of whom are highly concentrated, albeit on different
drugs.
Motivated by these two ndings, we construct a model of
physician learning-by-doing. Our model predicts how different
physicians locate along the concentrationdiversication contin-
uum. In particular, path-dependence in learning by doing is a
strong force towards the heterogeneous concentration we observe.
On the other hand, the model also predicts that the effects
of this path-dependence should be smaller for physicians with
larger prescription volumes. More specically, both concentration
(on one or a few drugs) and deviation (from the prescriptions
patterns of others) should be smaller for high-volume physi-
cians.
We nd support for both of these predictions. In addition, we
can distinguish our learning model from a competing hypoth-
esis emphasizing detailing by pharmaceutical representatives:
our model predicts that high-volume young physicians should
write a higher share of their prescriptions for old drugs than
low-volume young physicians should. We nd support for this
additional prediction, even when we dene old drugs to be
those that had ceased to be detailed before the young physi-
cians careers begana nding the detailing alternative cannot
explain.
The issues in this paper are important: advances in the practice
of evidence-based medicine are likely to be constrained consider-
ably if physicians limit the evolution of their evidentiary platform to
their own learning-by-doing experiences, down-weighting accu-
mulating evidence reported by other prescribers. Concentrated,
one size ts all prescribing behavior could fail to exploit oppor-
tunities to successfully tailor or personalize medical treatments
to patients idiosyncratic genetic, environmental and behavioral
characteristics. In contemplating responses to these problems,
however, it is of course important to understand the forces that
drive concentrated prescribing. In the learning-by-doing model
that we propose, physicians concentrate on a single drug because
learning how to properly manage the dosing and side effects asso-
ciated with a variety of different drugs would worsen expected
health outcomes for their patients while such learning took
place. In this setting, policies that either diverted patients to
more specialized physicians or lowered the costs associated with
27
learning about how to manage the side effects and dosing of
drugs would lead to more personalized medical treatments for
patients.
Given the importance of this topic, it is not surprising
that the issues we explore have been discussed by others.
For example, Coscelli (2000), Coscelli and Shum (2004) and
Frank and Zeckhauser (2007) also study concentrated prescribing
behaviorempirically and, in some cases, theoretically. In con-
trast to existing work on concentrated prescribing behavior, our
learning-by-doing model offers a new theory of how physician
beliefs form, and (if necessary) how they can be shaped issues
that are a key challenge for future research (Cutler et al., 2013, p.
28).1
In complementary work, focusing on patients instead of physi-
cians, Crawford and Shum (2005) and Dickstein (2011) analyze
how a given patients treatment regime evolves over time. More
specically, whereas we study what a physician can learn from one
patients experience with a given drug to forecast a future patients
experience with that drug, these studies of how a given patients
treatment evolves assume that there is no useful information from
one patients experience to improve the physicians forecast of a
future patients experience. We envision interesting and testable
implications from combining our focus on learning across patients
with this focus on learning within patients, and we hope that future
work will pursue such possibilities.
Many papers have analyzed whether unmeasured patient
heterogeneity is responsible for physician-level ndings in empir-
ical analyses like ours. The overwhelming nding from this
literaturewith contributions both by health economists (e.g.,
Hellerstein, 1997; Zhang et al., 2010; Cutler et al., 2013) and
by academic clinicians (e.g., Solomon et al., 2003; Schneeweiss
et al., 2005)is that the estimated role of physicians in inu-
encing treatment regimes is largely unaffected by incorporating
patient-specic data. We nd three examples particularly strik-
ing. First, as Coscelli (2000, p. 354) summarized his early work
with patient-level data: These patterns demonstrate clearly that
the probability of receiving a new treatment is signicantly inu-
enced by the doctors identity, and that doctors differ in their choice
among . . . drugs for the same patient. Second, the results obtained
by Frank and Zeckhauser (2007) suggest that, other than through
demographics, variations in patient condition severity and clini-
cal manifestations are remarkably unrelated to physician practice
behavior: the empirical results they obtained are largely quanti-
tatively unaffected with alternative specications incorporating
patient-specic data. Third, Cutler et al. (2013) nd that patient
demand is relatively unimportant in explaining variations in
Medicare expenditures across Hospital referral regions, whereas
the single most important factor is physician beliefs about treat-
ment. In short, similar to our hope that future theory will combine
learning across patients and learning within patients, our hope
is that future empirical work will combine longitudinal data on
both physicians and patients, but the existing empirical literature
1
Coscelli does not use a formal model. Coscelli and Shum use a two-armed bandit
model. Frank and Zeckhauser informally discuss a Sensible Use of Norms hypoth-
esis also based on a bandit model. See Section 3.2 for a comparison of bandit models
to our approach. Frank and Zeckhauser also discuss a My Way hypothesis, where
physicians regularly prescribe a therapy that is quite different from the choice that
would be made by other physicians (p. 1008). They interpret their My Way hypoth-
esis as physicians engaging in some highly suboptimal therapeutic practices (p.
1125), whereas in our model such heterogeneous concentration by physicians may
be optimal. Finally, none of these alternative models generates predictions about
either (a) the effect of volume on concentration and deviation or (b) the prescriptions
for old drugs by young physicians.
28 E.R. Berndt et al. / Journal of Health Economics 40 (2015) 2639

suggests that our results from physician-level data will very likely
persist.

1.2. Antipsychotics for the treatment of schizophrenia and related

conditions

Schizophrenia is an incurable mental illness characterized

by gross distortions of reality, disturbances of language and
communications, withdrawal from social interaction, and disorga-
nization and fragmentation of thought, perception and emotional
reaction (Anderson, 1998, 1456). Symptoms are both positive
(hallucinations, delusions, voices) and negative (depression, lack Fig. 1. Number of typical and atypical prescriptions, annually 19962007.
of emotion). The prevalence of schizophrenia is 12%, with
genetic factors at play but otherwise unknown etiology. The ill-
the FDA and launched in the US, including Risperdal (risperi-
ness tends to strike males in late teens and early twenties, and
done, 1993), Zyprexa (olanzapine, 1996), Seroquel (quetiapine,
females ve or so years later. As the illness continues, per-
1997), Geodon (ziprasidone, 2001) and Abilify (aripiprazole, 2002).
sons with schizophrenia frequently experience unemployment,
Guidelines from the American Psychiatric Association state that
lose contact with their family, and become homeless; a substan-
although each of these ve second-generation atypicals is approved
tial proportion undergo periods of incarceration (Domino et al.,
for the treatment of schizophrenia (some later also received FDA
2004).
approval for treatment of bipolar disease and major depressive
Because schizophrenia is a chronic illness affecting virtually
disorder, as well as various pediatric/adolescent patient subpo-
all aspects of life of affected persons, the goals of treatment are
pulation approvals), they also note that In addition to having
to reduce or eliminate symptoms, maximize quality of life and
therapeutic effects, both rst- and second-generation antipsy-
adaptive functioning, and promote and maintain recovery from
chotic agents can cause a broad spectrum of side effects. Side
the adverse effects of illness to the maximum extent possible
effects are a crucial aspect of treatment because they often determine
(American Psychiatric Association, 2004, p. 9). In the US, Medicaid
medication choice and are a primary reason for medication discon-
is the largest payer of medical and drug benets to people with
tinuation. (American Psychiatric Association, 2004, p. 66, italics
schizophrenia (Duggan, 2005).
added). Learning about such side effects is central to our theoretical
From 1955 through the early 1990s, the mainstays of pharma-
model.
cological treatment of schizophrenia were conventional or typical
Initially these SGAs were perceived as having similar efcacy for
antipsychotic (also called neuroleptic) drugs that were more effec-
positive symptoms and superior efcacy for negative symptoms
tive in treating the positive than the negative symptoms, but
relative to typicals, but without the older drugs extrapyrami-
frequently resulted in extrapyramidal side effects (such as tar-
dal and agranulocytosis side effects. However, beginning in about
dive dyskinesiaan involuntary movement disorder characterized
20012002 and continuing to the present, a literature has devel-
by puckering of the lips and tongue, or writhing of the arms or
oped associating SGAs with weight gain and the onset of diabetes,
legs) that may persist even after the drug is discontinued, and for
along with related metabolic syndrome side effects, particularly
which currently there is no effective treatment.2 In 1989, Clozaril
associated with the use of Zyprexa and clozapine and less so for
(generic name clozapine) was approved by the U.S. Food and Drug
Risperdal. Various professional treatment guidelines have coun-
Administration (FDA) as the rst in a new class of drugs called
seled close scrutiny of individuals prescribed Zyprexa, clozapine
atypical antipsychotics; this drug has also been dubbed a rst-
and Risperdal. The FDA has ordered manufacturers to add bolded
generation atypical (FGA). Although judged by many still to be
and boxed warnings to the product labels, initially for all atypicals,
the most effective among all antipsychotic drugs, for 12% of
and later, to both all typical and all atypical antipsychotic labels.
individuals taking clozapine a potentially fatal condition called
The labels have been augmented further with warnings regarding
agranulocytosis occurs (decrease in white blood cell count, leav-
antipsychotic treatment of elderly patients with dementia, since
ing the immune system potentially fatally compromised). Patients
evidence suggests this subpopulation is at greater risk for stroke
taking clozapine must therefore have their white blood cell count
and death.4
measured by a laboratory test on a regular basis, and satisfactory
Despite this controversy, as seen in Fig. 1, based on a 10% ran-
laboratory test results must be communicated to the pharmacist
dom sample of all antipsychotic prescribers in the U.S. (additional
before a prescription can be dispensed. For these and other rea-
data details below), the number of atypical antipsychotic prescrip-
sons, currently clozapine is generally used only for individuals who
tions dispensed between 1996 and 2007 increased about sevenfold
do not respond to other antipsychotic treatments (Frank et al.,
from about 400,000 in 1996 to 2,800,000 in 2007.5 In comparison,
2004).3
Between 1993 and 2002, ve so-called second-generation atyp-
ical (hereafter, SGA) antipsychotic molecules were approved by
4
Additional controversy emerged when major studies, published in 2005 and
2006, raised issues regarding whether there were any signicant efcacy and
tolerability differences between the costly SGAs and the older off-patent conven-
2
In an earlier version of this manuscript (Taub et al., 2011), we included in our tional antipsychotics, as well as differences among the ve SGAs. Important issues
analyses among the typical antipsychotics an old drug named prochlorperazine regarding the statistical power of these studies to detect differences, were they
(Compazine) that was FDA approved both for treatment of schizophrenia and nau- present, have also been raised, and currently whether there are any signicant
sea. Since its primary use has been for nausea, and since the branded version has now differences among and between the conventional and SGA antipsychotics remains
been withdrawn from the US market, we exclude that drug from our set of antipsy- controversial and unresolved. For further details and references, see the Appendix
chotics. The drugs we therefore count as typical antipsychotics are uphenazine, available from the corresponding author, Timelines U.S. Food and Drug Adminis-
haloperidol, loxapine, molindone, pimozide, perphenazine, thioridazine, thiothix- tration Approvals and Indications, and Signicant Events Concerning Antipsychotic
ene, chlorpromazine, and triuoperazine. Drugs.
3 5
For a history of clozapine and discussion of antitrust issues raised by the labo- We will use prescribed, written and dispensed interchangeably, but the
ratory test results requirement, see Crilly (2007). IMS Health XponentTM data are based on dispensed prescriptions; for a variety of
 E.R. Berndt et al. / Journal of Health Economics 40 (2015) 2639
the number of conventional or typical antipsychotic prescriptions
fell 55% from 1,100,000 in 1996 to about 500,000 in 2003 and has
stabilized at that level since then. As a proportion of all antipsy-
chotic prescriptions, the atypical percentage more than tripled
from about 27% in 1996 to 85% in 2007. It is also noteworthy that,
despite all the concerns about the safety and efcacy of antipsy-
chotics, the total number of antipsychotic prescriptions dispensed
in this 10% random sample typical plus atypical more than
doubled between 1996 and 2007, from about 1,500,000 to about
3,300,000.
2. Data and initial ndings
2.1. Prescription data
Our data on prescribers behaviors are taken from the IMS
XponentTM data source that tracks prescribing behavior by linking
individual retail and mail-order dispensed pharmacy prescriptions
to the prescriber identication number. A 10% random sample
of all prescribers who wrote at least one antipsychotic prescrip-
tion in 1996 was drawn. These prescribers are followed on a
monthly basis from January 1996 through September 2008. Each
year after 1996 the sample is refreshed by adding a 10% sam-
ple of new antipsychotic prescribers. These prescribers are new
in the sense that they are new to the sample; they may have
been prescribing antipsychotics for many years. For each physi-
cian prescriber, we have matched geographical, medical training
and ofce-practice data from the registry at the American Medical
Association.
Our data are a cross-section of prescribers in 2007. Although
manufacturers received approval to market reformulated versions
of several SGAs during the ve years leading up to our 2007 sam-
ple, no new major antipsychotic products were launched in the US
during these years, and 2007 is ten or more years after four of the
six atypicals were introduced.
We link the prescriber identiers in the IMS XponentTM data
base to the American Medical Association (AMA) Masterle
Directory, which provides education, training, specialty certi-
cation and demographic data on most physicians and type of
practice as of 2008. In addition, each prescriber in our sample
is assigned a geographical location based on their 2007 location.
The resulting dataset includes physicians in various special-
ties, such as psychiatry (general, childadolescent, and geriatric),
neurology (general, child, and geriatric), primary care (internal
medicine, family medicine, pediatrics, and general practice), and
others.
Our theoretical approach below applies to a group of physi-
cians each treating patients with symptoms drawn randomly from
a xed distribution, so we hereafter restrict attention to psychia-
trists (although our main empirical ndings are similar when we
analyze the full dataset and interact the main regressors with the
aggregate specialty groups above). In addition, to mitigate the pos-
sible impact of very low-volume prescribers, for the remainder of
the paper we limit the sample to the 2867 psychiatrists who in
2007 wrote at least 50 prescriptions for an antipsychotic. Again,
our main results are qualitatively similar for samples with lower
prescription thresholds (e.g., 12 per year) and higher (e.g., 100 per
year). In fact, our sample shrinks by only 16% when we shift from
reasons, a physician may prescribe a Product X but it may not be dispensed at all, or
in fact after consulting with the prescriber the pharmacist may dispense product Y.
29
the lower prescription threshold of 12 per year to our main sample
with threshold 50.6
Several nal features of the physician data set are also worth
noting. First, we have data on only physicians and their prescribing
behavior, not on the patients they see. Second, IMS keeps track of
prescribers who are deceased or retire, using look-back windows
with no prescribing activity for one year forward and one year back-
ward. Third, because the sample starts with prescribers who wrote
at least one antipsychotic prescription in 1996 (who are then fol-
lowed through September 2008, unless they die or retire), the set
of prescribers in the sample is likely older than would be observed
in an entirely new random sample drawn in, say, 2007.7
2.2. Initial ndings on heterogeneously concentrated prescribing
behavior
To describe our initial ndings, we must rst correct for the
mechanical bias present in other estimators due to sample sizes
(i.e., prescriptions per physician) too small to invoke the law of
large numbers. We begin by exploring the deviation of a physicians
prescriptionssay, from national market shares for the population
of physicians in question. Consider physician i prescribing drug d,
and denote the share of prescriptions written by this physician for
drug d as sid . Let the national market share of drug d be md , where
both sid and md are between zero and one. As a measure of the devi-
ation of physician is prescribing behavior from national market
shares, we calculate

Di = (sid md )2 . (1)
d
If every physician had the same prescription behavior, Di would
equal zero. As physician prescribing behavior heterogeneity
increases, Di increases.
Ellison and Glaeser (1997) note that, at small volumes, there will
be a mechanical reduction in the deviation measure Di as volume
increases. To correct for this small-volume issue in the deviation
measure, they revise the raw deviation measure (1) as follows:
   
Vi  1
i =
D Di 1 m2d . (2)
Vi 1 Vi
d
where Vi is the volume of prescriptions written by physician i. Here-
after we refer to this revised measure of deviation as corrected
deviation.
We document below not only that many physicians have pre-
scription patterns that are statistically signicantly different than
would be predicted by random draws from national market shares,
but also (as an example) that many physicians have prescription
patterns that are statistically signicantly more concentrated than
would be predicted by such random draws. One measure of a pre-
scribers concentration Ci is the Herndahl-Hirschman
 2 Index (HHI)
of the physicians prescriptions, Ci = s . As with deviation,
d id
however, at small volumes, there will be a mechanical reduction in
the HHI measure of concentration as volume increases. To correct
for this small-numbers volume issue in the concentration measure,
we amend the raw HHI index as follows:
Vi
 1

C i = Ci . (3)
Vi 1 Vi
6
Results for these alternative samples are available from the corresponding
author.
7
In a Physician Sample appendix, available from the corresponding author, we
discuss this latter point in more detail.
30
Hereafter we refer to this revised measure of concentration as cor-
rected concentration. 8
Using these corrected measures of deviation and concentra-
tion, we now present two striking initial ndings. First, even in
our sample of psychiatrists with high prescription volumes, many
physicians have prescription patterns that are statistically signi-
cantly different than random draws from national market shares.
To assess the statistical signicance of such ndings, we follow
Ellison and Glaeser (1997) by computing the statistic
 patients is w; so patient t arrives at the physicians ofce at the
Di (1 m2d )/Vi
ti =  d
Var(Di )
for each physician. Given our sample of psychiatrists with high pre-
scription volumes, we can apply the Central Limit Theorem to show
that if |ti | > 2 then the null hypothesis is rejected for psychiatrist i
at the 5% level. In our sample, the 5th percentile in the distribution
of ti is 4.85, meaning that for over 95% of the psychiatrists in our
sample, we reject the null at the 5% level.
Complementing this nding about the statistical signicance of
deviation from national market shares, we also show that many
physicians have prescription patterns that are signicantly more
concentrated than would be predicted by random draws from
national market shares. To do so, we compute an analogous statistic

Ci ([(Vi 1)/Vi ] m2d + [1/Vi ])
Ti =  d
Var(Ci )
for each physician. Applying the Central Limit Theorem again gives
that if |Ti | > 2 then the null hypothesis is rejected for psychiatrist i at
the 5% level. In our sample, the 25th percentile in the distribution
of Ti is 2.83, meaning that for over 75% of the psychiatrists in our
sample, we reject the null at the 5% level.
In addition to these statistical tests, we also argue (although
less formally) that, among psychiatrists who deviate from psy-
chiatrist national market shares by a similar amount, there is
substantial heterogeneity in prescription patterns. For example, if
we (temporarily) limit the sample to psychiatrists with the most
concentrated prescribing, i.e. those for whom |Ti | > 2 (n = 2348),
54% chose Seroquel as their favorite drug, 28% Risperdal, 10% Abil-
ify, 5% Zyprexa, 1% Geodon, and 1% clozapine, with 1% for typical
antipsychotics.9
To recap, even in our sample of psychiatrists with high prescrip-
tion volumes, national market shares for this population do not
reect homogeneous physicians each prescribing drugs in propor-
tions approximating national shares, but rather the aggregate of
heterogeneous physicians many of whom are highly concentrated,
albeit on different drugs.
8
To see why we use this corrected measure of concentration, suppose that a
physician i prescribes a drug d with probability pd independently across periods  2
and that the realized share of a drug d is sid . Then the expectation of C i is pd .
Specically,

 We assume  d and dt are independent normally distributed ran-
Vi
 1
 dom variables for all d and t, with mean zero and variances d2 and
E[C i ] = E 2
sid = p2d
Vi 1 Vi
d d
because
  p pd )
2 d (1
E sid = (Var(sid ) + p2d ) =
Vi
d d d d
9 10
In comparison, in our 2007 sample of high-volume psychiatrists, the national
market percentages of the six atypicals were Seroquel 30%, Risperdal 24%, Abilify
15%, Zyprexa 11%, Geodon 8%, and clozapine 3%, with 9% for typical antipsychotics.
E.R. Berndt et al. / Journal of Health Economics 40 (2015) 2639
3. Towards a theory of prescriber learning and treatment
behavior
3.1. A model of prescriber learning-by-doing
We assume that patients arrive sequentially to be seen by a
physician (say, a female) and are indexed by periods in which they
arrive t N = {1, 2, . . .}. That is, there are innitely many patients
and one physician. The time between the arrivals of successive
(4) point in time tw. The continuous time discount rate is r.
The physician observes that patient t has symptom s randomly
drawn from the set of all possible symptoms S = {s1 , . . ., sS } with
the corresponding probabilities p1 , . . ., pS . Symptoms are drawn
independently across patients. The set of available drugs that treat
these symptoms consists of D = {d1 , . . ., dD }. The maximum possi-
ble benet of drug d for symptom s is Bsd . The ideal drug treatment
for a given symptom s is d* (s), meaning that Bsd *(s) > Bsd for all
d =/ d * (s). The physician knows Bsd for all combinations of s in
S and d in D. That is, the learning in our model is not about the
maximum possible benet derived from drug d for a patient with
symptom s; that ideal benet is already known by the physi-
cian.
The therapy for a patient includes not only the drug d that
the physician prescribes, but also any complementary action a
(5) that the physician undertakes, such as adjusting the dosage of
the drug (a process known as titrating), perhaps because the
patient is a heavy smoker, or any actions that affect the patients
adherence and outcomes, such as communicating information on
possible side effects and their duration, possible adverse interac-
tions with other drugs, and/or the best time of the day to take
the drug (e.g., take once-a-day sedating drugs at night).10 To
achieve the maximum potential benet from a drug, the physi-
cian must undertake the ideal complementary action. It is this
ideal complementary action that the physician learns about in our
model.
To formalize the process of learning about complementary
actions, we assume that the realized effectiveness of drug d pre-
scribed for patient t with symptom s is
bsdt = Bsd (a xdt )2 , (6)
where a denotes the complementary action the physician under-
takes, and
xdt = d + dt . (7)
Thus, to achieve the maximum possible benet (bsdt = Bsd ) from
drug d for patient t with symptom s, the physician must choose the
ideal complementary actions for drug d and patient t (a = xdt ), where
these actions depend on both the drug ( d ) and the patient (dt ). As
|a xdt | increases, the realized benet from drug d decreases at an
increasing rate; as a result, even the optimal drug for the patients
d
symptom, d * (s), can yield very poor outcomes if |a xdt | is large.
2 , respectively.
Recall that the physician knows the maximum potential ben-
et from each drug Bsd . The only uncertainty the physician faces
is what complementary actions will work best for a specic drug

Vi 1
 and a particular patient. From (7), the ideal complementary actions
1
+ p2d = p2d + .
Vi Vi
We are indebted to Marcela Horvitz-Lennon, M.D., for discussion of physicians
common complementary actions when prescribing antipsychotic drugs to people
with schizophrenia.
 E.R. Berndt et al. / Journal of Health Economics 40 (2015) 2639
for a given drug have a patient-specic component dt and a gen-
eral component  d . Because the patient-specic components dt are
independent across patients, there is nothing to learn from, say, d1
about dt for t > 1. Rather, the physician uses experience with this
drug from past patients to learn the general component  d , which
is relevant for future patients receiving this drug.
To simplify our analysis, we make a seemingly strong (but ulti-
mately inconsequential) assumption: after prescribing drug d to
patient t and undertaking complementary actions a, the physician
observes the complementary action that would have been optimal
(i.e., xdt ). Note that the physician does not observe the ideal actions
had that patient been given another drug (i.e., xd t for d =/ d) or the
ideal actions for another patient given that drug (i.e., xdt for t =
/ t).
In short, our assumption gives the physician an unrealistically large
amount of information about the patient just treated, but even all
this information still leaves the physician with much to learn about
the average ideal complementary actions for a given drug (and an
average patient),  d .
The intuition underlying our model is simple. The physician
learns about  d by prescribing drug d and subsequently observing
the ideal complementary action xdt for patient t. Because the physi-
cian does not observe  d , she cannot learn everything she needs
to know about a drug from treating one patient with this drug.
We have assumed that the variance of  d may depend on drug d,
but the variance of dt depends neither on drug d nor on patient
t. Therefore, initially the physician may have different uncertain-
ties associated with distinct drugs. However, the speed of learning
the complementary action  d for each drug d depends on only how
often the physician prescribes drug d, not on the drug or patient
identity.
3.2. Discussion of the model
Our model builds on Jovanovic and Nyarko (1996), in which
a decision-maker also knows all parameters of the environ-
ment except the optimal complementary action. Their model also
assumes a quadratic objective function and normally distributed
random variables. The novel aspect of our model is random symp-
toms, which implies that the long-run prescribing behavior of the
physician depends on the initial history of symptoms presented to
her.
Our model has the same reduced form as another class of
models, also called learning modelsnamely, models of learn-
ing curves or learning by doing, where benets for each drug
increase deterministically with the number of times the drug is
prescribed. In particular, Eqs. (8) and (9) imply that in our model
the expected benets from prescribing drug d for symptom s are
equal to
2 d2
Bsd 2 ,
2 + d2 #d
where # d is the number of times the physician prescribed drug d.
In addition, if there is full learning about each drug after one pre-
scription of the drug (i.e., if 2 = 0), then our model is equivalent
to the following conceptually different model. There are benets
Bsd that the physician obtains if she prescribes drug d for symp-
tom s. The physician incurs a xed cost of d2 when she prescribes
drug d for the rst time, and thereafter she incurs no cost when she
prescribes drug d. This xed cost can represent either the phys-
ical cost of reading instructions on how to use a new drug or
the cognitive costs of switching from a customary drug to a new
drug.
Our model also differs from the multi-armed bandit models
(see e.g., Bergemann and Valimaki, 2006). In the multi-armed
31
bandit analog of our model, the effectiveness of each drug Bsd
would be unknown and there would be no complementary actions.
That is, patients experiences would be noisy signals for the true
quality of a drug. Then, similarly to our model, in some cases
physicians prescribing choices may diverge even if initially they
had the same beliefs about the efcacy of each drug. Crawford
and Shum (2005), Ferreyra and Kosenok (2010), and Dickstein
(2011) estimate models in this spirit, but they do not focus
on either concentration or deviation in prescriptions by physi-
cians (not to mention the effects of volume on concentration and
deviation).11
While bandit models have important applications when physi-
cians are trying to learn the true qualities of drugs, these models
are less useful in our setting where physicians need to learn the
complementary actions for a drug. In particular, in a two-armed
bandit model, if players observe each others decisions, then even-
tually all players settle on the same decision with probability one
(Aoyagi, 1998). And in our setting, a physician can observe the
national market shares of the drugs, which provide that physician
information about what other physicians prescribed (and, implic-
itly, some information about what other physicians learned about
the efcacy of various drugs) and still make different prescrib-
ing decisions. Thus, a two-armed bandit model would contradict
one of our main empirical ndingsheterogeneous concentra-
tion. More generally, in a multi-armed bandit model, if physicians
observe national market shares of all drugs, it is not clear that
any of our empirical ndings would ariseconcentration, devia-
tion, variation in both concentration and deviation with volume,
and variation in young physicians use of old drugs with vol-
ume.
In contrast, in our learning-by-doing model, the physicians
prescribing behavior does not depend on whether the physician
observes national market shares, because the underlying efcacy of
each drug is already known by each physician. As a result, observing
other physicians prescription decisions conveys no useful informa-
tion: a physician must learn how to use a drug, and no amount of
being told that other physicians have learned how to use it can
teach the physician. That is, from the prescribers perspective, each
drug is an experience good rather than a search good.12
3.3. Analysis of the model and preliminary comparative statics
We assume that the physician seeks to maximize the expected
present value of the sum of the realized effectiveness (bsdt ) of the
drugs she prescribes to the sequence of patients she treats (t = 1,
2, . . .). The optimal prescribing behavior of the physician can be
characterized in a simple manner because our model is station-
ary and the realized effectiveness has a quadratic structure with
normally distributed uncertainty components. Denote the physi-
cians history through patient t by ht = t1 (s , d , a , xd  ). The
=1 
physicians policy decision is to choose a drug d and complemen-
tary action a, for each patient t with symptom s and at each history
ht .
Because complementary action a does not affect learning about
 d , the optimal complementary action a and physicians expected
11
More specically, Crawford and Shum (2005) and Dickstein (2011) use
patient-level data, so they can analyze a patients learning but not a prescribers
concentration. In contrast, Ferreyra and Kosenok (2010) share our focus on pre-
scriber learning and analyze prescriber data, but they focus on learning to prescribe
a single new drug, rather than on the steady-state concentration or deviation of
prescriptions.
12
For a model of antipsychotic and antidepressant prescribing behavior incorpo-
rating spillovers depending on the close-knittedness of prescribers, see Domino
et al. (2012).
32 E.R. Berndt et al. / Journal of Health Economics 40 (2015) 2639
instantaneous benet from prescribing drug d for patient t are given
by:
a(ht ) = E[d |ht ],
(8)
E[bsdt |ht ] = Bsd Var(d |ht ) 2 ,
where E[d |ht ] and Var( d |ht ) denote the conditional expectation
and variance of  d at history ht . Moreover, the standard formula
for Bayesian updating with normally distributed random variables
yields:
1 1 #d(ht )
= 2 + , (9)
Var(d |ht ) d 2
where # d(ht ) denotes the number of patients to whom the physi-
cian prescribed drug d during history ht . From (8) and (9), we see
that the more times a physician has prescribed drug d, the closer
she will expect to be to achieving the second-best benets of the
drug d for a patient with symptom s, namely Bsd 2 .
The optimized expected benet from prescribing drug d to
patient t with symptom s that is, E[bsdt |ht ] in (8) depends on d
in two ways: the maximum benet Bsd , which is already known,
and the expected loss from imperfect complementary actions,
Var(d |ht ) + 2 , which depends on the history ht only through
posterior variances Var( d |ht ). That is, the physicians prescribing
behavior can be summarized by D state variables identied with
posterior variances Var( d |ht ) for d D. Therefore, to compare pre-
scribing behavior of physicians with different histories, we need to
compare only their posterior variances of  d .
To prepare for the empirical work below, we now discuss
comparative-static results of the learning-by-doing model with
respect to w, the time between the arrivals of successive patients.
Suppose rst that w is large (i.e., the physician is a low-volume
prescriber). In this case, the physician will eventually concentrate
on a subset of drugs, in the sense that all future prescriptions will
be from this subset, and each drug in this subset will be prescribed
for some symptom. Moreover, this subset of drugs will depend on
the initial history of patients symptoms randomly presented to
the physician. The intuition behind this is as follows. If the physi-
cian observes an initial sequence of patients each of whom has a
given symptom s, then she will choose an appropriate drug, say d,
for them. The physician will learn a great deal about this drug d
and will be unwilling to switch to another drug d when she sees
a patient with symptom s (even if d would be more appropriate
for s if the physician had the same knowledge about drugs d and
d ).
More formally, consider a physicians choice for a patient with
symptom s between two drugs d and d. If the physician is myopic
then the expected benets to the patient from using drugs d and d
are given by
Bs d Var(d |ht ) 2 ,
Bs d Var(d |ht ) 2 .
Therefore, the myopic physician compares the difference between
Bs d and Bs d to the difference between Var(d |ht ) and Var( d |ht ). If
the maximum potential benet from drug d , Bs d , is greater than
that from drug d, Bs d , but the physician has prescribed drug d more
often than drug d in the past so that
Var(d |ht ) < Var(d |ht ) (Bs d Bs d ),
then she will choose drug d.
Now consider physicians who are not myopic. For physicians
with extremely high values of w (i.e., very low patient volumes), the
prescription behavior that optimizes the expected present value of
the realized effectiveness (bsdt ) is essentially the myopic behavior
just described. For physicians with lower values of w (i.e., higher
volumes of patients), however, optimal prescription behavior now
accounts for the fact that learning more about a new drug today
will improve effectiveness for future patients given this drug, who
may arrive soon. In this sense, high-volume prescribers have a
larger incentive to invest in learning how to use new or differ-
ent drugs effectively. The set of drugs a physician eventually uses
will still depend on the initial history of symptoms the physi-
cian has seen, but this dependence becomes weaker as patient
volume increases. Therefore we expect to see lower concentra-
tion and lower deviation with increases in patient volume, all else
equal.
Finally, as w approaches zero (i.e., the physician sees patients
almost continuously), the set of drugs that the physician eventually
prescribes ceases to depend on the symptoms of the initial patients
that the physician randomly sees. More formally, if we assume that
there are sufciently many different symptoms such that each drug
d in D is optimal for some symptoms s in S (i.e., for each d there exists
s such that d* (s) = d), then a physician with very high patient volume
will eventually learn a great deal about optimal complementary
actions  d for each drug d in D and prescribe d* (s) for every s.
To exposit all these ideas in a simple setting, in Appendix A we
solve an example of our model. To accelerate physicians progress
towards steady-state prescription behaviors, we assume that 2 =
0, so that a physician learns everything about a drugs comple-
mentary actions after prescribing the drug just once. Proposition
1 describes the solution to this example, and Corollaries 1 and 2
then show, respectively, that expected concentration and expected
deviation are decreasing with volume.
4. Heterogeneous concentration, deviation and
prescription volumes
4.1. Empirical framework and econometric methods
The cross-sectional regression specication we take to the 2007
data is of the following general form:
Yi = ln Vi + Xi + i (10)
where Yi is either the corrected measure of deviation from national
psychiatrist market shares in (2) or the corrected measure of con-
centration in (3), Vi is the number of antipsychotic prescriptions
written by psychiatrist i in 2007, and Xi is a vector of covariates
described below. We take the age of the psychiatrist from the AMA
Masterle Directory. As a exible specication, we use age quartiles
(age < 45, 45 age < 54, 54 age < 62, and age 62) as indicator-
variable regressors instead of the raw age metric.
Although our theory has nothing to say about gender, female
and male psychiatrists may use this technology in different ways.
Therefore we control for the gender of the psychiatrist. In addition,
some psychiatrists ask that their prescribing data not be shared
with pharmaceutical or other for-prot organizations; thus we add
an opt-out indicator variable to the specication.
The theory above emphasizes differential learning costs (2
and d2 in our model), and we might expect the learning costs for
psychiatrists to depend on their training and/or current practice
environment. In particular, we control for whether the psychia-
trist practices in a group or has a solo practice, is hospital-based or
not, the population of the county in which the psychiatrist prac-
tices (in thousands), and whether the psychiatrist has an MD or DO
degree.13
13
DO is doctor of osteopathy. Anderson (1998, p. 1169) denes osteopathy as
. . .a therapeutic approach to the practice of medicine that uses all the usual forms
 E.R. Berndt et al. / Journal of Health Economics 40 (2015) 2639 33

Table 1
Summary statistics for 2007 prescriber sample.

Variable Mean Std. dev. Minimum Maximum

HHI of Individual Physicians Antipsychotic Prescribing 0.29 0.11 0.12 0.95

Corrected HHI of Individual Physicians Antipsychotic Prescribing 0.28 0.11 0.12 0.95
Deviation of Physicians Antipsychotic Prescribing from Nat. Mkt. Shares 0.08 0.08 0.002 0.86
Corrected Deviation of Physicians Antipsychotic Prescribing from Nat. Mkt. Shares 0.08 0.08 0.01 0.86
Total Yearly Antipsychotic Prescriptions 726 849 50 7186
Prescriber Age 53 12 28 92
Solo Practice 0.16 0.37 0 1
Population (county) in thousands 1240 1840 3 9735
Female 0.34 0.47 0 1
Hospital based Physician 0.14 0.34 0 1
DO Flag 0.04 0.18 0 1
Physician Opt Out 0.04 0.19 0 1
Number of observations 2867

All values calculated using IMS Health Incorporated XponentTM general prescriber sample 2007 data for psychiatrists writing at least 50 antipsychotic prescriptions.

Summary statistics for all the dependent and explanatory vari- ways. In particular we estimated the model using 1/Vi as well as
ables in our analyses are presented in Table 1. In part because we just linear Vi . Although neither of those t the data as well as ln Vi ,
limit our sample to psychiatrists writing at least 50 prescriptions, the negative sign of the estimated effect of volume on psychiatrist
the raw and corrected concentration and deviation are very simi- deviation and HHI concentration prescribing was robust.
lar. We also experimented with using different antipsychotic pre-
scribing frequency cutoffs when constructing our dataset we use
4.2. Results: deviation and concentration models in our regressions (at least 12, 75, and 100 in 2007). The sign of the
estimated effect of volume on psychiatrist deviation and concen-
Results from OLS estimation with corrected deviation from tration remained unchanged.
national psychiatrist market shares and corrected concentration as We also explored whether similar results hold if we include
dependent variables are reported in Table 2. The omitted reference physicians from all of the specialties in our full dataset. In this
case is a male psychiatrist under age 45 in a group practice that is regression specication we interacted the volume measure with
not hospital-based, who has not opted out of the AMA Masterle physician-specialty dummies (psychiatrist, neurologist, primary
registry and has an MD degree. We present parameter estimates care, and other) thereby allowing exibility in how high and low
on only the ln Vi , age quartiles and female gender variables; esti- volume physicians in different specialties differ from one another.
mates for the other covariates are available from the corresponding Consistent with Table 2, we found that both psychiatrist deviation
author.14 from national market shares and psychiatrist HHI concentration
As seen in the rst column of Table 2, in the corrected deviation were negatively related to volume. In addition, we found nega-
model the coefcient on ln Vi is negative and statistically signi- tive estimated effects for the other specialties. In fact, in this full
cant, consistent with our learning-by-doing theoretical framework. dataset (and with the lower prescription threshold of 12), the
Psychiatrists in the two oldest age quartiles are more deviant volume effects were even stronger and the R2 was more than
than psychiatrists under age 45, although the relationship is not twice as high. The estimated volume effects were generally largest
monotonic with age. While female psychiatrists are very slightly for other physicians, followed by primary care physicians, then
less deviant than males, this effect is not statistically signi- neurologists, and smallest for psychiatrists. One plausible inter-
cant. pretation of this last nding is that greater residency training
Also consistent with our theoretical framework, corrected con-
centration declines with increases in prescribing volume, and
Table 2
signicantly so. Although none of the age quartile coefcient
OLS estimates of corrected deviation and corrected concentration.
estimates is statistically signicant, female psychiatrists are
slightly more concentrated in their antipsychotic prescribing Corrected deviation Corrected HHI
behavior than are their male counterparts. Log(Total Yearly Antipsychotic 0.025*** 0.034***
Prescriptions)
4.2.1. Robustness [0.001] [0.002]
Age Quartile 4553a 0.006 0.004
We have undertaken a number of robustness checks. For exam- [0.004] [0.005]
ple, we repeated the analysis allowing volume to enter in different Age Quartile 5461a 0.016*** 0.007
[0.004] [0.005]
Age Quartile 62+a 0.011*** 0.006
[0.004] [0.006]
of medical diagnosis and therapy, including drugs, surgery, and radiation, but that Femalea 0.001 0.010**
places greater emphasis on the inuence of the relationship between the organs and [0.003] [0.004]
the musculoskeletal system than traditional medicine does. Osteopathic physicians Number of observations 2867 2867
recognize and correct structural problems using manipulation. Although the vast R2 0.13 0.14
majority of psychiatrists in our sample have an MD degree, a number have DO
training and degree. Dependent variables: Corrected Deviation of Psychiatrist Antipsychotic Prescrib-
14
In the corrected-deviation regression, the only coefcient estimate that is sig- ing from National Psychiatrist Market Shares as in (2), Corrected Concentration of
nicantly different from zero is that for the hospital-based psychiatrist: 0.010, Psychiatrists Antipsychotic Prescribing Shares as in (3). Standard errors in square
standard error of 0.004. In the corrected-concentration regression, the only statis- brackets.
a
tically signicant estimates are those for the hospital-based psychiatrist (0.012, Dummy variable.
**
standard error 0.006) and that for the opt-out psychiatrist (0.038, standard error p < 0.05.
***
0.010). p < 0.01.
34 E.R. Berndt et al. / Journal of Health Economics 40 (2015) 2639
concerning antipsychotics to some extent substitutes for physi-
cian experience gained by prescribing a certain type of drug more
frequently.
5. Exploring a competing hypothesis: detailing
There may be alternative frameworks that help explain the
variations we observe in physician prescribing behavior. One
plausible competing hypothesis involves pharmaceutical sales
representatives (called detailers) who may target their sales
efforts at high-volume prescribers. More specically, suppose that,
rather than high-volume prescribing behavior generating greater
prescription heterogeneity through the logic of our learning-by-
doing model, high-volume prescribers were instead exposed to
detailing by a greater number of different pharmaceutical man-
ufacturers than are low-volume prescribers (because of the large
returns potentially realized by pharmaceutical detailing when
a high-volume prescriber is persuaded to prescribe a particular
branded drug by a detailer). In this competing hypothesis it is
the increase in detailing that leads to less concentrated prescrib-
ing by high-volume physiciansperhaps because some detailers
provide persuasive information, or because writing a few pre-
scriptions for each detailed drug is a reciprocal form of behavior
providing some positive feedback from the prescriber to the various
detailers.15
In evaluating this competing hypothesis, it is important to note
that drugs are detailed only when they are on patent or have market
exclusivity for other reasons. (After a branded drug faces generic
competition, there are few incentives for its manufacturer to detail
physicians: the brand would be unable to appropriate many bene-
ts, which for the most part would instead accrue to the generics.16 )
An implication is that drugs having lost market exclusivity many
years ago are unlikely to have been detailed to young doctors prac-
ticing in 2007.
In order to compare the predictions of the competing hypoth-
esis (that physician detailing drives heterogeneous prescribing
behavior) to the predictions of our learning-by-doing model, we
distinguish old antipsychotic drugs approved and launched in the
US before 1990 (i.e., Clozaril and all the typical antipsychotic drugs)
from new antipsychotic drugs (i.e., all SGA atypical antipsy-
chotics, the earliest of which was Risperdal, approved in 1993).
Note that the ten typical drugs prescribed in our 2007 sample were
approved for marketing by the FDA between 1957 and 1984, and
Clozaril (also an FGA) was approved in 1989; they all experienced
generic entry by 1996, many much earlier in the 1980s. An impli-
cation is that none of these old drugs was detailed after 1996. Thus,
any effect that detailing of old drugs might have on prescriptions
in 2007 would be a very long-run effect. More importantly, for our
purposes, such detailing of old drugs could not have been targeted
at physicians who were not practicing when this detailing occurred,
before 1996.
To highlight prescriber-age issues, in this section we limit
the sample to the oldest (age 62 and up in 2007) and youngest
(age below 45 in 2007) age quartiles of psychiatrists writing at
least 50 antipsychotic prescriptions in 2007 (1410 psychiatrists).
The youngest quartile of psychiatrists were between the ages of
17 and 33 in 1996, when the last old drug experienced generic
entry. These youngest psychiatrists are therefore very unlikely
ever to have been detailed on an old drug. In contrast, the old-
est quartile of psychiatrists were age 51 and older in 1996. These
15
For a model of reciprocal behavior in response to gift giving and experimental
evidence, see Malmendier and Schmidt (2011).
16
For discussion and empirical evidence, see Berndt et al. (2003).
Table 3
Tobit estimates of percentage of prescriptions for old drugs.
% Rxs for new drugs
a
Physician Age 62+ 13.278***
[4.023]
Log(Total Yearly Antipsychotic Prescriptions) 2.252***
[0.476]
(Physician Age 62+)(Log(Total Yearly Antipsychotic 1.331**
Prescriptions))
[0.659]
Femalea 2.157***
[0.812]
Number of observations 1410
Pseudo R2 0.0123
Left censored 221
Right censored 0
Mean of dependent variable 11.01
All values calculated using IMS Health Incorporated XponentTM general prescriber
sample 2007 data, and population estimates from the US Census Bureau. New drugs
are dened as SGA atypicals. Sample is comprised of the oldest (62 +) and youngest
(2844) quartile of psychiatrists.
a
Dummy variable.
**
Signicance at 5% level.
***
Signicance at 1% level.
oldest psychiatrists therefore could have been detailed on an old
drug.
If pharmaceutical detailing were the primary driver of psychia-
trists choice of which antipsychotic class of drugs to prescribe (old
vs. new), then the youngest psychiatrists should prescribe very few
of the older drugs. In addition, we would expect high-volume young
psychiatrists (who are likely visited the most by pharmaceutical
detailers promoting new drugs) to be the least likely to prescribe
older drugs.
In contrast, under our learning-by-doing model, the share of
old drugs prescribed by young psychiatrists should increase with
volume for high enough volumes (see Appendix A): in our frame-
work, high-volume young psychiatrists have an incentive to invest
in learning the complementary actions for old drugs because these
drugs deliver the highest benets for some (albeit a small minor-
ity of) patients. On the other hand, young psychiatrists with low
volumes typically have insufcient incentive to invest in learning
the complementary actions for old drugs (because the rst set of
patients they encountered typically had symptoms best treated by
the new drug, so they prescribed the new drugs and learned about
their complementary actions).
As the dependent variable we employ the psychiatrists share of
total antipsychotic prescriptions written for the old drugs, where
the share ranges from zero to 100. (Since in our data this share
sometimes is zero, we employ Tobit rather than OLS estimation.)
If the detailing hypothesis were the primary driver of prescriber
choice, for young psychiatrists the old share would decrease with
volume. If instead our learning-by-doing model were the primary
driver of prescriber choice, for young psychiatrists the old share
would increase with volume. To allow for differential volume effects
by age, we specify a model with an interaction between volume and
age > 62. The results are shown in Table 3.
Consistent with both the detailing and learning-by-doing
hypotheses, Table 3 shows that the oldest psychiatrists prescribe
larger shares of the old drugs, but that this share decreases with an
older psychiatrists prescription volume. More importantly, since
the estimated effect of the ln Vi is positive and signicant, the high-
est volume psychiatrists in the youngest quartile prescribe a larger
share of old drugs. This is consistent with our learning-by-doing
framework, but is at odds with the detailing hypothesis, because
these youngest high-volume psychiatrists are likely to have been
 E.R. Berndt et al. / Journal of Health Economics 40 (2015) 2639
Table 4
Means, standard deviations and coefcients of variation for antipsychotic HHIs: alternative geographical aggregates, 2007.
Geographic aggregate Mean corrected HHI HHI std. dev.
Individual prescriber 0.28 0.11
County 0.24 0.08
Hospital referral region 0.22 0.05
State (plus District of Columbia) 0.20 0.02
Nation 0.19 na
IMS Health Incorporated XponentTM 2007 data general prescriber sample data. Includes all psychiatrists who wrote at least 50 prescriptions for antipsychotics in 2007.
heavily detailed on new drugs, but are likely never to have been
detailed on the old drugs.
We conclude, therefore, that while the predictions of our
learning-by-doing model are generally observed in the prescribing
data, a crucial prediction of the detailing hypothesis is at odds with
the prescribing behavior we observe among young psychiatrists:
High-volume young psychiatrists prescribe old drugs more often than
do low-volume young psychiatrists.
6. Discussion and conclusion
We conclude by (1) exploring the connection between our
results and the literature on regional-variation, (2) discussing pos-
sibilities for future work, and (3) summarizing the paper.
6.1. Regional variation
Our ndings of heterogeneous concentration raise an intrigu-
ing possibility. The highly publicized regional-variation literature
documents that within-region treatment variations for selected
conditions experienced by Medicare patients are relatively small
compared to much larger and persistent between-region differ-
ences in treatments and costs (e.g., Skinner and Fisher, 1997; Fisher
et al., 2003a,b; Yasaitis et al., 2009). Could it be that our ndings of
heterogeneous concentration are driven by correspondingly large
between-region variability in antipsychotic prescribing behav-
ior? Alternatively, is most variability in antipsychotic prescribing
behavior psychiatrist-specic, with regional patterns similar to
each other?
To analyze regional variation, we return to our full sample of
2867 psychiatrists writing at least 50 antipsychotic prescriptions in
2007. We compute mean HHIs and their variability (both standard
deviations and coefcients of variation) at alternative levels of
regional aggregation. While most of these regional aggregates are
familiar, we note that hospital referral regions (HRRs) represent
306 regional health-care markets that have played a prominent
role in the Dartmouth regional-variation and related literatures.
Results are given in Table 4.
As expected from our results in Section 2.2, prescriber behavior
is more concentrated at the individual level than are national mar-
ket shares (mean corrected HHI falls from 0.28 at the individual
level to 0.19 nationally), with monotonically decreasing concen-
tration as one considers larger aggregate regions (from county to
HRR to state to nation). Also consistent with Section 2.2, there is
more variability in prescribing behavior across psychiatrists at the
individual level than across larger aggregate regions: the coef-
cient of variation of corrected HHI is 0.39 at the individual level
but falls monotonically to 0.10 at the state level, and the mean
corrected deviation (dened in Section 2.2) likewise falls mono-
tonically from 0.077 at the individual level to 0.007 at the state
level. Phelps (1992, pp. 2526) has categorized coefcients of vari-
ation for surgical procedures in the 0.10.2 range as revealing low
variability, while those at 0.4 and greater are termed high
35
Coef. of variation Mean corrected deviation Deviation std. dev.
0.39 0.077 0.080
0.32 0.041 0.051
0.22 0.023 0.036
0.10 0.007 0.010
na na na
variability procedures. Within that classication scheme, the con-
centration of antipsychotic prescribing behavior exhibits close to
high variability at the individual-prescriber level, moderate vari-
ability at the county level, and low variability at the HRR and
especially at state levels.
In short, in our data, practice heterogeneity is much greater at
the individual level than at the HRR level: the coefcient of varia-
tion of corrected HHI is almost twice as high at the individual level
as at the HRR level, almost reaching Phelps high threshold for the
former and barely exceeding his low ceiling for the latter. Cutler
et al. (2013) offer related evidence that the strongest determinant
of regional variation is differences in physician beliefs about the
efcacy of particular therapies.
6.2. Future work
Several interesting future research projects have emerged from
our study. As noted earlier, the relative efcacy, tolerability and
cost-effectiveness of the various typical and atypical antipsychotics
remain controversial issues, even after publication of a substan-
tial number of articles over the last decade, including those based
on randomized controlled clinical trials.17 What is less controver-
sial is that this dispute has had a substantial impact on changing
prescription shares of the various antipsychotics. Our IMS Health
data reveal that between 2002 and 2008, the Seroquel prescription
percentage increased from 21% to 37%, Abilify from 0% to 16%,
Geodon from 4% to 7%, even as the Risperdal share declined from
35% to 26%, and that of Zyprexa declined most dramatically from
34% to 12%. Who were the prescribers who switched most rapidly
low or high volume, what specialties, gender, age group, solo vs.
group practice and who were those who changed relatively lit-
tle? What were the relative responses by different prescribers to
the FDA issuing bold boxed warnings, to professional associations
revising treatment guidelines, to publication of major ndings in
medical journals? More generally, how well does our theoretical
framework, implemented here in a cross-sectional context, predict
dynamic behavior of physicians? Understanding which prescribers
respond most and which the least would provide valuable informa-
tion to guide future regulatory-related information dissemination
strategies.18
In addition to the dynamics of prescribing behavior for individ-
ual physicians, it would be interesting to study the dynamics of
such behaviors for groups of physicians. In this paper, our model
and empirics ignored learning from others, spillovers, and herding
behavior. Chandra and Staiger (2007) have developed and esti-
mated a model that focuses on productivity spillovers related to
local specialization in heart attack care, whereby excellence in
one clinical approach in a local market raises the average skill of
17
Among the more notable publications are those based on the CATIE study; see,
for example, Lieberman et al. (2005), White (2006) and Kraemer et al. (2009).
18
The only research on this topic of which we are aware is that by Hoblyn et al.
(2006).
36 E.R. Berndt et al. / Journal of Health Economics 40 (2015) 2639
other practitioners of that approach operating in the same market.
This in turn leads to greater specialization and reduces both the
absolute and relative productivity of practitioners using alternative
approaches. Homogeneity in clinical approach within a geograph-
ical area, and substantial heterogeneity across areas, can reect
what may also be sensible and useful since they stem from posi-
tive spillover effects from local specialization. In future research,
it would be useful to attempt to incorporate various types of
spillover effects into physician prescribing behavior. This is partic-
ularly important, since learning from sources other than ones own
prescribing behavior is a critical component in national efforts to
enhance the practice of evidence-based medicine.19
Finally, our ndings (and those of others) suggest that a sig-
nicant proportion of the heterogeneity in the treatments patients
receive depends upon physician preferences concerning treatment
regime. Our model and empirics focus on the roles of initial patients
and future volume in determining a physicians prescription pat-
tern. It would also be informative and useful to identify other
physician experiences that generate differences in practice style
across physicians, perhaps related to location of medical residency
training, analogous to recent investigations characterizing man-
agement style.20
6.3. Summary
We have developed and implemented empirically a model in
which a physician treats a sequence of patients with random symp-
toms. For each patient, the physician prescribes a drug and chooses
a complementary action. The physician knows the maximum pos-
sible benet from prescribing any drug for any symptom (i.e., the
benet to the patient if the optimal complementary actions are
taken), but does not know ex ante the complementary actions
that achieve this maximum benet for any given drug. By pre-
scribing a drug, choosing complementary actions, and observing
the patients response, the physician learns about the appropriate
complementary actions for that drug. Thus, in our model, there is
learning-by-doing, causing physicians to be more adept at choosing
complementary actions for drugs they have prescribed previously
than for drugs they have not yet prescribed. On the other hand,
knowing that some drugs are well suited to certain symptoms,
physicians may optimally prescribe an unfamiliar drug in response
to a new symptom, especially if this and other symptoms that may
be well addressed by this drug are likely to recur in future patients.
The main predictions of our model arise from considering dif-
ferences in optimal prescribing behavior for physicians treating
different volumes of patients. In particular, past volume inu-
ences the extent of learning-by-doing and hence a physicians
ability to choose appropriate complementary actions for familiar
drugs, whereas future volume inuences the expected benets to
future patients from prescribing an unfamiliar drug for the current
patient, so as to learn more about its appropriate complementary
actions. High-volume physicians are thus expected to prescribe a
wide range of drugs. Low-volume physicians, in contrast, may opti-
mally treat the patients they see by learning a great deal about
appropriate complementary actions for a small subset of the avail-
able drugs and not prescribing drugs from outside this subset.
Furthermore, the drugs optimally included in this subset depend
on the random symptoms presented by the patients the physician
treats early in her career. In short, both concentration and deviation
decrease with volume.
19
For an attempt to incorporate spillovers from close knitted prescribers in the
context of antipsychotic prescribing, see Domino et al. (2012).
20
See, for example, Bertrand and Schoar (2003) and Kaplan et al. (2008).
We have confronted this model with cross-sectional data on
antipsychotic prescriptions, regressing corrected deviation and
corrected concentration on the volume and other characteristics
of psychiatrists writing at least 50 antipsychotic prescriptions in
2007. As predicted by our model, we observe that higher-volume
psychiatrists have lower deviation and concentration in their pre-
scribing behavior.
To compare our learning model to a model of detailing by sales
representatives to psychiatrists, we regress the share of prescrip-
tions written for new drugs on the psychiatrists age quartile, total
volume of antipsychotic prescriptions written, and the interaction
of the two. Consistent with our learning model but at odds with
the detailing model, we nd that the higher volume psychiatrists
in the youngest age quartile prescribe a larger share of old drugs
than do their lower volume counterparts.
Acknowledgements
This research has beneted enormously from the IMS Health
Services Research Network that has provided data and data assis-
tance. Special thanks are due to Stu Feldman, Randolph Frankel,
Cindy Halas, Robert Hunkler and Linda Matusiak at IMS Health.
We have also beneted from feedback by seminar participants
at Wharton, Northeastern University, Boston University School of
Public Health, the NBER, the University of Chicago, and the Univer-
sity of California Los Angeles, and from the comments of Joseph
Doyle, Marcela Horvitz-Lennon, Ulrike Malmendier, David Moli-
tor, Jonathan Skinner, Douglas Staiger and Richard Zeckhauser.
The statements, ndings, conclusions, views and opinions con-
tained and expressed in this manuscript are based in part on
19962008 data obtained under license from IMS Health Incor-
porated: National Prescription AuditTM , XponentTM and American
Medical Association Physician MasterleTM . Such statements, nd-
ings, conclusions, views and opinions are not necessarily those
of IMS Health Incorporated or any of its afliated or subsidiary
entities. This research has not been sponsored. Prof. Berndt has
received an unrestricted gift from Pzer, Inc., to perform health
policy research administered through the MIT Sloan School of
Management. All authors contributed equally to this research; the
alphabetical ordering is by convention.
Appendix A. A 2 2 example
To obtain more precise comparative-static results (and to illus-
trate the logic of the model more generally), consider a simple
example that satises the following assumption:
Assumption 1. erw = , S = {s1 , s2 }, D = {d1 , d2 }, Pr(s2 ) = p2 > 1/2,
12 = 22 = c > 1, 2 = 0, B12 = B21 = 0, B11 = B22 = 1.
A verbal interpretation of Assumption 1 is the following. We
dene as = erw . Therefore, a higher value of corresponds to a
physician who has a shorter time between the arrivals of successive
patients and hence sees a higher volume of patients. There are two
drugs d1 and d2 , and two symptoms s1 and s2 . Symptoms s2 and
s1 are realized with probabilities p2 and p1 = 1 p2 , respectively.
Symptom s2 occurs more often than symptom s1 (i.e., p2 > 1/2).
Therefore, drug d2 is more likely to be ideal for a randomly drawn
symptom. In all other respects, drugs and symptoms are symmetric
(i.e., B11 = B22 , B12 = B21 , and 12 = 22 ).
Before seeing any patients, the physician has the same uncer-
tainty  d about the ideal complementary action for each drug d
(i.e., 12 = 22 > 0). However, the physician learns the ideal com-
plementary action precisely after one prescription (i.e., 2 = 0). As
discussed in Section 3.2 of the main text, this learning assumption
 E.R. Berndt et al. / Journal of Health Economics 40 (2015) 2639
Fig. A1. Physicans prescribing behavior in a 2 2 example. (For interpretation of the references to color in the text, the reader is referred to the web version of the article.)
implies that the physician incurs a xed cost c =  2 when she pre-
scribes drug d for the rst time, and thereafter she incurs no cost
when she prescribes drug d.
The ideal drugs for given symptoms are normalized in such a
way that d * (s1 ) = d1 and d * (s2 ) = d2 (i.e., B11 > B12 and B22 > B21 ).
Without loss of generality, we can normalize B12 = B21 = 0 because
only the relative benets B22 B21 and B11 B12 matter for the
physicians choice of drug d. Likewise, without loss of generality we
can jointly rescale B11 , B22 , and  2 so that B11 = B22 = 1. Finally, to
make the analysis interesting, we assume that the myopic physi-
cian concentrates on the drug prescribed to the rst patient (i.e.,
c > 1).
In Proposition 1, we fully characterize the physicians optimal
prescribing behavior under Assumption 1. Fig. A1 illustrates differ-
ent cases that arise in the model depending on parameter values.
The explicit formulas for the boundaries of different regions of
Fig. A1 are given in the proof of Proposition 1 in an online appendix.
Proposition 1. Let Assumption 1 hold. There are six different cases
that can arise in the model that correspond to the combination of a
color (green, yellow, red) and a shade (light, dark) shown in Fig. A1.
(The dark red area exists iff c > 2.)
In the rst period, the physician prescribes:
the ideal drug in the light color areas;
the drug d2 in the dark color areas.
Starting from the second period the physician prescribes:
the ideal drug in the green area;
the ideal drug or the drug d2 depending on whether d1 or d2 was
prescribed in the rst period, respectively, in the yellow areas;
the drug prescribed in the rst period in the red areas.
To provide intuition for Proposition 1, we explain color and
shade regions of Fig. A1 in turn. We begin by explaining different
colors in Fig. A1. A low-volume physician (red area) never exper-
iments. She always concentrates on the drugs prescribed in the
past. An intermediate-volume physician (yellow area) is willing to
experiment and prescribe a new drug only if this new drug is more
likely to be the ideal drug than the drug she prescribed in the past.
As the probability that the new drug is ideal increases, a physician
has higher incentives to experiment with the new drug. This cor-
responds to the decreasing boundary between the red and yellow
areas in Fig. A1. A high-volume physician (green area) is always
willing to experiment and prescribe a new drug. As the probability
that the new drug is ideal decreases, a physician has lower incen-
tives to experiment with the new drug. This corresponds to the
increasing boundary between the yellow and green areas in Fig. A1.
37
We now explain shades (light and dark areas) in Fig. A1. Shades
determine what drug a physician prescribes at the beginning of
her career. In light areas, an inexperienced physician prescribes the
ideal drug (drug di for symptom si ), whereas in dark areas she pre-
scribes the more popular drug (drug d2 ) regardless of symptoms.
Note that in dark areas, the inexperienced physician prescribes the
more popular drug even though this drug may be suboptimal for the
patient. This occurs because the inexperienced physician expects
the more popular drug to be optimal for most future patients, so
she invests in learning how to use this drug at the beginning of her
career. Note that in the dark yellow area the physician concentrates
on the most popular drug her entire career. However, she would
diversify and always prescribe the ideal drug in the long run if she
were forced to prescribe the less popular drug at the beginning of
her career.
Finally, we explain why a physician prescribes the more popular
drug at the beginning of her career only if she sees an intermediate
volume of patients and the more popular drug is very likely to be
ideal (i.e., why the dark area occurs at intermediate values of and
high values of p2 ). A low-volume physician prescribes the ideal
drug because she is not willing to invest in learning any drug (e.g.,
as volume goes to zero, the physician becomes myopic and so does
what is best for the current patient). In contrast, a high-volume
physician prescribes the ideal drug because she is willing to invest
in learning complementary actions for both drugs. Therefore, only
an intermediate-volume physician can invest in learning only the
more popular drug. The intermediate-volume physician invests in
learning only about the more popular drug only if this more popular
drug is very likely to be ideal in the future.
Proposition 1 immediately implies that under reasonable
restrictions on model parameters, concentration and deviation
decrease with volume. For the concentration result, we just need
to assume that parameters are such that the left panel of Fig. A1
applies. For the deviation result, we also need to assume that the
market shares are not extreme. In particular, we assume that the
market share of the more popular drug is higher than the fre-
quency of the symptom for which this drug is ideal. This assumption
automatically holds if the economy is populated with physicians
who may differ in volume but otherwise are identical. Further, we
assume that the share of physicians who prescribe only the more
popular drug is less than a half.
Corollary 1. Suppose that Assumption 1 holds and c > 2. Then the
expected concentration of a physician decreases with volume.
Corollary 2. Suppose that Assumption 1 holds, c > 2, and the market
share m2 of drug d2 satises m2 [p2 , (1 + p2 )/2]. Then the expected
deviation of a physician decreases with volume.
38 E.R. Berndt et al. / Journal of Health Economics 40 (2015) 2639
A.1. Comparing cohorts of physicians and eras of drugs
We now use this 2 2 example to build intuition for what our
model predicts about the prescriptions of typical versus atypical
antipsychotics by old versus young physicians. Specically, con-
sider the following sequence of eras denoted T = 1, 2, 3: at T = 1, a
cohort of old physicians is trained and has access to only typical
antipsychotics; at T = 2, a cohort of young physicians is trained
(and the old continue to practice) and all physicians have access
to both typical and atypical drugs; nally, at T = 3, both cohorts are
practicing and have access to both kinds of drugs. We will view
T = 3 as 2007, the year of our data. We now explore what the 2 2
example predicts about prescriptions in T = 3.
In T = 1, there are two possible symptoms (s1 and s2 ), a cohort
of physicians beginning their prescribing careers (hereafter, old
physicians), and only one drug available (which we will interpret
as a typical antipsychotic and label as d1 ). For these old physicians
during T = 1, all they can do is prescribe d1 , so they do so for all
symptoms (s1 and s2 ). As a result, because Assumption 1 implies full
learning after one prescription, these old physicians know perfectly
how to take complementary actions for d1 in the future.
In T = 2, another drug becomes available (which we will inter-
pret as an atypical antipsychotic and label as d2 ) and a new cohort
of physicians begin its prescribing careers (hereafter, new physi-
cians). Both old and new physicians know that drug di is the best
prescription for symptom si , in the sense that this prescription max-
imizes Bsd . The only difference between the new and old physicians
is that the new physicians do not yet know how to take complemen-
tary actions for either drug (d1 or d2 ), whereas the old physicians
do know how to do this for the typical (d1 ) but not for the atypical
(d2 ).
Because the market share of atypicals relative to typicals is
very large (much greater than 0.5) in 2007, we assume that
Pr(s2 ) = p2 > 1/2, again in keeping with Assumption 1. For example,
let us set p2 = 6/7. If we then proceed upwards in Fig. A1 along a
vertical line at p2 = 6/7, we are comparing physicians with different
volumes.
Recall that old and new physicians have different histories
at T = 3. For new physicians, T = 3 is their second period, so their
prescription at T = 3 depends on their history at T = 2. For old
physicians, T = 3 is their third period, so their prescription at T = 3
depends on their history at T = 1 and the fact that the new drug
arrived at T = 2. Designating (x, y) to mean that a physician is
prescribing fraction x of d1 and fraction y of d2 , where x + y = 1, we
then have the following prescription behaviors as a function of the
colored and shaded regions in Fig. A1.
Old physicians New physicians
Light red All are (1, 0) 1 p2 are (1, 0) ; p2 are (0, 1)
Dark red All are (1, 0) All are (0, 1)
Dark yellow All are (1 p2 , p2 ) All are (0, 1)
Light yellow All are (1 p2 , p2 ) 1 p2 are (1 p2 , p2 ); p2 are (0, 1)
Light green All are (1 p2 , p2 ) All are (1 p2 , p2 )
For old physicians, concentration falls with volume, the num-
ber of atypicals increases with volume, and the share of atypicals
increases with volume. For new physicians, concentration falls with
volume, the number of atypicals weakly increases with volume,
and the share of atypicals falls with volume for sufciently high
volumes. The last of these results is the most important: high-
volume young physicians have an incentive to invest in learning
the complementary actions for old drugs (typical antipsychotics)
because these drugs deliver the highest benets for some (albeit a
small minority) of patients. Alternatively, viewing the table from
the opposite perspective, both old and young physicians with
low volumes have insufcient incentive to invest in learning the
complementary actions for a drug, but for old physicians it is the
new drug about which they do not learn (because they learned
about the old drug when it was the only one available), whereas
for new physicians it is most often the old drug about which they
do not learn (because their rst patient had symptom s2 and so
the physician prescribed d2 and learned about its complementary
actions).
Appendix B. Supplementary data
Supplementary data associated with this article can be found,
in the online version, at http://dx.doi.org/10.1016/j.jhealeco.
2014.11.003.
References
American Psychiatric Association, 2004. Practice Guideline for the Treatment of
Patients with Schizophrenia, second ed, Available from: http://www.psych.org
Anderson, K.N. (Ed.), 1998. Mosbys Medical, Nursing & Allied Health Dictionary,
fth ed. Mosby-Year Book, Inc., St. Louis.
Aoyagi, M., 1998. Mutual observability and the convergence of actions in a multi-
person two-armed bandit model. Journal of Economic Theory 82, 405424.
Bergemann, D., Valimaki, J., 2006. Bandit problems. In: Durlauf, S., Blume, L. (Eds.),
The New Palgrave Dictionary of Economics. Macmillan Press, Basingstoke.
Berndt, E.R., Kyle, M.K., Ling, D.C.Y., 2003. The long shadow of patent expiration:
generic entry and Rx to OTC switches. In: Feenstra, R.C., Shapiro, M.D. (Eds.),
Scanner Data and Price Indexes. NBER Series on the Conference on Research in
Income and Wealth, vol. 61. University of Chicago Press for the National Bureau
of Economic Research, Chicago, pp. 229267 (Chapter 8).
Bertrand, M., Schoar, A., 2003. Managing with style: the effect of managers on rm
policies. Quarterly Journal of Economics 118 (4), 11691208.
Chandra, A., Staiger, D.L., 2007. Productivity spillovers in health care: evidence from
the treatment of heart attacks. Journal of Political Economy 115 (1), 103141.
Coscelli, A., 2000. The importance of doctors and patients preferences in the pre-
scription decision. Journal of Industrial Economics 48, 349369.
Coscelli, A., Shum, M., 2004. An empirical model of learning and patient spillovers
in new drug entry. Journal of Econometrics 122, 213246.
Crawford, G., Shum, M., 2005. Uncertainty and learning in pharmaceutical demand.
Econometrica 73, 11371173.
Crilly, J., 2007. The history of clozapine and its emergence in the US market: a review
and analysis. History of Psychiatry 18 (1), 3960.
Cutler, D.M., Skinner, J., Stern, A.D., Wennberg, D., 2013. Physician Beliefs and Patient
Preferences: A New Look at Regional Variation in Health Care Spending, Working
Paper No. 19320. National Bureau of Economic Research, Cambridge, MA.
Dartmouth Atlas Project. The Dartmouth Atlas of Health Care. Available from:
http://www.dartmouthatlas.org
Dickstein, M.J., 2011. Efcient Provision of Experience Goods: Evidence from Antide-
pressant Choice (unpublished). Harvard.
Domino, M.E., Frank, R.G., Berndt, E.R., 2012. The Diffusion of Antipsychotic Medica-
tions: Does the Close-knittedness of Provider Networks Matter? Paper Presented
at 2009 Annual Meetings of the American Economic Association, San Francisco,
CA, January 4, 2009. Revised March 2012.
Domino, M.E., Norton, E.C., Morrissey, J.P., Thakur, N., 2004. Cost shifting to jails
after a change to managed mental health care. Health Services Research 39 (5),
13791401.
Duggan, M., 2005. Do new prescription drugs pay for themselves? The case of second
generation antipsychotics. Journal of Health Economics 24, 131.
Ellison, G.D., Glaeser, E.L., 1997. Geographic concentration in U.S. manufacturing
industries: a dartboard approach. Journal of Political Economy 105 (5), 889927.
Ferreyra, M.M., Kosenok, G., 2011. Learning about new products: an empirical study
of physicians behavior. Economic Inquiry 49 (3), 876898.
Fisher, E.S., Wennberg, J.E., Stukel, T.A., et al., 2003a. The implications of regional
variations in Medicare spending. Part 1: The content, quality, and accessibility
of care. Annals of Internal Medicine 138 (4), 273287.
Fisher, E.S., Wennberg, J.E., Stukel, T.A., et al., 2003b. The implications of regional
variations in Medicare spending. Part 2: Health outcomes and satisfaction with
care. Annals of Internal Medicine 138 (4), 288298.
Frank, R.G., B., E.R., B., A.B., Lehman, A.F., 2004. Quality-constant price indexes for the
ongoing treatment of schizophrenia: an exploratory study. Quarterly Review of
Economics and Finance 44 (July (3)), 390409.
Frank, R.G., Zeckhauser, R.J., 2007. Custom-made versus ready-to-wear treatments:
behavioral propensities in physicians choices. Journal of Health Economics 26
(December (6)), 11011227.
Hellerstein, J., 1997. The importance of the physician in the generic versus trade-
name decision. RAND Journal of Economics 27, 108136.
Hoblyn, J., Noda, A., Yesavage, J.A., et al., 2006. Factors in choosing atypical antipsy-
chotics: toward understanding the bases of physicians prescribing decisions.
Journal of Psychiatric Research 40, 160166.
Jovanovic, B., Nyarko, Y., 1996. Learning by doing and the choice of technology.
Econometrica 64 (6), 12991310.
 E.R. Berndt et al. / Journal of Health Economics 40 (2015) 2639
Kaplan, S.N., Klebanov, M.M., Sorensen, M., 2008. Which CEO Characteristics and
Abilities Matter? (unpublished). University of Chicago Booth School of Manage-
ment, Chicago, Available from: skaplan@uchicago.edu
Kraemer, H.C., Glick, I.D., Klein, D.F., 2009. Clinical trials design lessons from
the CATIE study. American Journal of Psychiatry 166 (November (11)),
12221228.
Lieberman, J.A., Stroup, S.T., McEvoy, J.P., et al., 2005. Effectiveness of antipsychotic
drugs in patients with chronic schizophrenia: primary efcacy and safety out-
comes of the clinical antipsychotic trials of intervention effectiveness (CATIE)
schizophrenia trial. New England Journal of Medicine 353, 12091223.
Malmendier, U., Schmidt, K.M., 2011, June. You Owe Me (unpublished working
paper). University of California at Berkeley, Department of Economics, Berkeley,
CA, Available from: ulrike@econ.berkeley.edu
Phelps, C.E., 1992. Diffusion of information in medical care. Journal of Economic
Perspectives 6 (Summer (3)), 2342.
Schneeweiss, S., Glynn, R.J., Avorn, J., Solomon, D.H., 2005. A Medicare database
review found that physician preferences increasingly outweighed patient
39
characteristics as determinants of rst-time prescriptions for COX-2 inhibitors.
Journal of Clinical Epidemiology, Brief Reports 58, 98102.
Skinner, J., Fisher, E., 1997. Regional disparities in Medicare expenditures: an oppor-
tunity for reform. National Tax Journal 50 (3), September.
Solomon, D.H., Schneeweiss, S., Glynn, R.J., Levin, R., Avorn, J., 2003. Determinants
of selective cyclooxygenase-2 inhibitor prescribing: are patient or physician
characteristics more important? American Journal of Medicine 115, 715720.
Taub, A.A.L., Kolotilin, A., Gibbons, R.S., Berndt, E.R., 2011, February. The Diversity of
Concentrated Prescribing Behavior: An Application to Antipsychotics. Working
Paper No. 16823. National Bureau of Economic Research, Cambridge, MA.
White, R.F., 2006. The signicance of CATIE: an expert interview with Jeffrey A.
Lieberman, MD. Medscape Psychiatry & Mental Health, 2006.
Yasaitis, L., Fisher, E.S., Skinner, J.S., Chandra, A., 2009. Hospital quality and intensity
of spending: is there an association? Health Affairs Web Exclusive 28 (May (4)),
W566W572.
Zhang, Y., Baicker, K., Newhouse, J.P., 2010. Geographic variation in Medicare drug
spending. New England Journal of Medicine 363 (5), 405409.
 Journal of Health Economics 40 (2015) 4053

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

Hungry today, unhappy tomorrow? Childhood hunger and subjective

wellbeing later in life
Marco Bertoni
University of Padova, Italy

a r t i c l e i n f o a b s t r a c t

Article history: I use anchoring vignettes to show that, on data for eleven European countries, exposure to episodes of
Received 8 May 2014 hunger in childhood leads people to adopt lower subjective standards to evaluate satisfaction with life in
Received in revised form adulthood. I also show that, as a consequence, estimates of the association between childhood starvation
18 December 2014
and late-life wellbeing that do not allow for reporting heterogeneity are biased towards nding a positive
Accepted 18 December 2014
correlation. These results highlight the need to consider rescaling when drawing inference on subjective
Available online 29 December 2014
outcomes.
2014 Elsevier B.V. All rights reserved.
JEL classication:
C42
I31
J13

Keywords:
Subjective wellbeing
Hunger
Anchoring vignettes

1. Introduction
There is wide evidence in economics showing that family back-
ground and shocks taking place early in life affect human capital
accumulation, health and socio-economic conditions throughout
This paper draws from a chapter of my PhD thesis. I am extremely grateful for
supervision from Giorgio Brunello, for advice from Danilo Cavapozzi, and for sug-
gestions and comments from the Editor, two anonymous referees, Viola Angelini,
Andrea Brandolini, Massimiliano Bratti, Martina Celidoni, Luca Corazzini, Michele De
Nadai, Marta De Philippis, Enkelejda Havari, Monica Langella, Andrea Moro, Omar
Paccagnella, Mario Padula, Paolo Roberti, Silvana Robone, Lorenzo Rocco, Alfonso
Rosolia, Elisabetta Trevisan, Guglielmo Weber, participants at the 3rd Potsdam
Workshop in Empirical Economics, the 2014 BOMOPAV Workshop, the 2014 ASSET
meeting, and at seminars in Padua, Venice and the Bank of Italy. All results men-
tioned in the paper but not shown are available from the author. All errors are my
own. This paper uses data from SHARELIFE release 1, as of November 24th 2010 and
SHARE release 2.5.0, as of May 24th 2011. The SHARE data collection has been pri-
marily funded by the European Commission through the 5th framework programme
(project QLK6-CT-2001- 00360 in the thematic programme Quality of Life), through
the 6th framework programme (projects SHARE-I3, RII-CT-2006 062193, COMPARE,
CIT5-CT-2005-028857, and SHARELIFE, CIT4-CT-2006-028812) and through the 7th
framework programme (SHARE-PREP, 211909 and SHARE-LEAP, 227822).
Correspondence to: Department of Economics and Management M. Fanno,
University of Padova, via del Santo 33, 35123 Padova, Italy. Tel.: +39 049 8274050.
E-mail address: marco.bertoni@unipd.it
the life course.1 Understanding the long reach of events hap-
pening in critical periods of human development is of uttermost
importance for both economists and policy makers to identify the
proximate causes of successful lives (Layard et al., 2013).
In particular, exposure to childhood starvation has been widely
studied by economists as a case of extremely severe and usually
temporary material deprivation that might anyhow have per-
manent negative consequences. Recent research has shown that
hunger episodes experienced in the earliest phases of life (Meng
and Qian, 2009; Lindeboom et al., 2010; Neelsen and Stratmann,
2011) and until early adolescence (Havari and Peracchi, 2011;
Pinger et al., 2011; Portrait et al., 2011; Kesternich et al., 2013) have
1
On the one hand, Cunha and Heckman (2007), propose a formal dynamic model
of investment in skills across the life cycle, while Currie (2009), provides a thorough
review of the empirical studies documenting a link between family background,
childhood health and human capital accumulation. On the other hand, wide evi-
dence is available on the effects of childhood shocks on late-life outcomes. For
instance, Ichino and Winter-Ebmer (2004), Akbulut-Yuksel (2009), Kesternich et al.
(2014), and Jrges (2013), nd negative effects of exposure to World War II on edu-
cation, health and labour market outcomes, while Lindeboom et al. (2006), Angelini
and Mierau (2012), and van den Berg et al. (2010, 2014), relate exposure to negative
macroeconomic conditions at birth to longevity, childhood health and the cognitive
and health decline due to adverse shocks later in life.

http://dx.doi.org/10.1016/j.jhealeco.2014.12.006
0167-6296/ 2014 Elsevier B.V. All rights reserved.
 M. Bertoni / Journal of Health Economics 40 (2015) 4053
persistent negative effects on education, late-life health, and labour
market outcomes. Hypothesized pathways are both due to biologi-
cal effects (the fetal origin hypothesis see Almond and Currie, 2011)
and to behavioural reactions (see Kesternich et al., 2013).
On the other hand, little is known yet about the link between
childhood hunger and subjective wellbeing later in life. This is quite
surprising given the increasing effort devoted by policy makers
and economists to the analysis of self-reported happiness, with
the aim of considering a denition of wellbeing that goes beyond
strictly economic measures and encompasses multiple dimen-
sions of quality of life (Layard, 2005, 2006). Nicolas Sarkozys and
David Camerons speeches inviting to make happiness the new
policy maximand instead of GDP, the United Nations World Hap-
piness Report series and the OECDs Better Life Index are only some
examples of policy efforts in this sense, while Frey and Stutzer
(2002), Dolan et al. (2008), Powdthavee (2010), and Graham (2012),
provide thorough reviews of the academic work carried out on
subjective wellbeing measures as a tool for welfare analysis. The
little evidence about the long reach of childhood hunger on happi-
ness later in life available so far, presented by Havari and Peracchi
(2011), documents that the experience of extreme material depri-
vation due to war-driven hunger episodes has negative scarring
effects on mental wellbeing decades later, but the topic is still
under-researched. From a broader perspective, Frijters et al. (2011),
and Layard et al. (2013), analyze life-cycle models of subjective
wellbeing, that link a large set of childhood conditions to subjective
wellbeing later in life.
This paper contributes to the literature by taking a different
viewpoint on the relationship between childhood hunger and late-
life wellbeing. A standard assumption in the analysis of satisfaction
with life is the absence of heterogeneity in the way people inter-
pret what being satised means. However, as underlined by Clark
et al. (2005), if people systematically attached different labels to
the same concept of wellbeing, a crucial issue of heterogeneity in
reporting styles would limit the use of subjective measures of wel-
fare. In psychometrics, this problem is known as Differential Item
Functioning (Holland and Wainer, 1993).
Deaton (2008), highlights in this sense that the evaluation of
subjective wellbeing is a relative one: people compare their sit-
uation with a subjective benchmark, a shifting standard that
depends on ones past experiences. According to the satisfac-
tion treadmill model of hedonic adaptation (Kahneman, 1999;
Frederick, 2007), the experience of extreme deprivation leads
people to develop lower aspirations regarding the level of life
achievements to consider as satisfying. As a consequence, when
asked to evaluate a given situation, individuals exposed to hunger
episodes in the past may assign it a better judgement than peo-
ple not exposed, even if that situation represents the same level
of latent happiness for both groups. This point is raised also
by Sen (1985a, 1985b), within the capabilities approach to wel-
fare analysis. He considers subjective evaluations as an unreliable
basis to formulate normative judgements, precisely because of
adaptive aspirations. In his view, the experience of negative con-
ditions may push individuals to accommodate their desires to
contextual constraints (Comin et al., 2004). As a consequence,
for instance, poor people could report to be satised with their
lives even if their latent wellbeing is low, because of deformed
aspirations.
To the best of my knowledge, no empirical evidence is available
on the relation between exposure to negative childhood conditions,
and hunger in particular, and individual reporting scales. This paper
contributes to ll this gap. On top of the importance of this question
to qualify how hedonic adaptation to negative life events works,
modelling response scales allows me to evaluate the association
between exposure to hunger episodes in childhood and late-life
41
wellbeing levels without issues of rescaling bias. Following Sens
approach, if people exposed to hunger adopt lower reference points
to dene a satisfactory life, estimates of the association between
hunger and subjective wellbeing that do not take rescaling into
account will be biased towards nding a positive correlation.
The empirical difculty related with this exercise is the sepa-
rate identication of the effects of childhood hunger on reporting
scales and on latent wellbeing. To this end, I rely on an anchoring
vignettes approach (see King et al., 2004).2 Anchoring vignettes can
be considered as the Rosetta stones of subjective evaluations. In a
vignettes questionnaire, respondents rst evaluate their own situ-
ation in a given domain, then they rate a series of descriptions of
situations of hypothetical persons, the anchoring vignettes, on the
same rating scale used for the self-assessment. Under the assump-
tion that systematic variation in vignette ratings only reects
differences in reporting styles, I can exploit the vignettes to model
individual reporting scales and, in turn, to lter subjective evalua-
tions from reporting heterogeneity.3
I use data on the population aged 50+ of eleven European coun-
tries from the Survey of Health, Ageing and Retirement in Europe,
SHARE. There are two main advantages of using SHARE for this anal-
ysis: the second wave of the survey contains a set of vignettes on life
satisfaction, while the third wave collects retrospective informa-
tion on peoples life histories, including data on exposure to periods
of hunger in childhood and on family background.4
My main empirical result is that the experience of childhood
starvation leads people to shift downwards their subjective refer-
ence points for the evaluation of satisfaction with life, i.e., to give a
higher rating to the same latent level of wellbeing. I also estimate
a long run negative association between hunger exposure in child-
hood and late-life happiness, but contrarily to the extant literature I
am the rst to derive this result from a model that takes differences
in reporting styles into account. Comparing my estimates with the
ones from a model that does not allow for reporting heterogeneity,
I conclude that failure to address this issue leads to underestimate
the negative relationship between childhood hunger and late-life
wellbeing.
Beyond contributing to the economic and psychological liter-
ature on hedonic adaptation, showing an example of satisfaction
treadmill, my ndings are especially relevant for the economic lit-
erature that wants to model the determinants of life satisfaction
over the life cycle, as they highlight the need to consider the rela-
tionship between past conditions and individual-specic reporting
scales to draw conclusions that are free from rescaling bias.
The paper unfolds as follows. Section 2 presents the data and
some descriptive statistics. The econometric model is described in
2
From an econometric perspective, commonly used techniques to deal with
reporting heterogeneity when longitudinal data are available include also the use
of individual xed effects models, that tease out time-invariant scale effects. Still,
results from Angelini et al. (2011), highlight that the scale people use to rate their
conditions is a time-varying one, and these techniques are not of much help to deal
with cross-sectional data.
3
This approach was introduced in social sciences by King et al. (2004), in the
context of political efcacy. Kapteyn et al. (2007), and Angelini et al. (2011), are illus-
trative examples concerning disability conditions, while Bago dUva et al. (2008),
Kok et al. (2012), and Peracchi and Rossetti (2012), study self-reported health and
depression, and Rice et al. (2011, 2012), deal with health system responsiveness. In
the context of life satisfaction, Angelini et al. (2012, 2015), look at satisfaction with
life in general, and Bonsang and van Soest (2012a, 2012b), look at satisfaction with
social contacts, job and income.
4
These measures have been validated against potential issues of recall bias
in other studies (see for instance Garrouste and Paccagnella, 2012; Havari and
Mazzonna, 2011). However, given the self-reported nature of the information on
hunger episodes, if more pessimistic people are more likely to report that they have
suffered of hunger, then the effects of hunger on reporting scales that I estimate are
lower bounds to the true ones.
42 M. Bertoni / Journal of Health Economics 40 (2015) 4053

Section 3, and Section 4 illustrates the main empirical results. I

discuss the validity of the vignettes approach for this analysis in
Section 5, while Section 6 presents some sensitivities and exten-
sions to the main results. Conclusions follow.

2. The data

This paper exploits the Survey of Health, Ageing and Retirement

in Europe, SHARE, that collects detailed information on household
composition, socio-economic status, health and wellbeing of the
population aged 50+ of several European countries.
On the one hand, I draw information on adulthood conditions,
on self-reported wellbeing and on anchoring vignettes from the
COMPARE subsample of the second wave of SHARE, that was col-
lected in 2006/07. On the other hand, I rely on data on childhood
environment contained in SHARELIFE, the third wave of the sur-
vey, that was carried out in 2008/09. Unfortunately, about 30% of
the initial COMPARE sample is lost in SHARELIFE because of panel
attrition. To test for non-random attrition, Table A1 in the Appendix
compares mean baseline values of several relevant variables com-
puted in the full COMPARE sample and in the nal sub-sample, for
the full sample and separately for each country. Mean values of
most variables look alike across samples, and I cannot detect any
selection pattern that is common across countries. The nal sample
is composed of 4950 individuals who took part in both the COM-
PARE and the SHARELIFE surveys, were born between 1920 and
1956 and were residing in Belgium, the Czech Republic, Denmark,
France, Germany, Greece, Italy, the Netherlands, Poland, Spain and
Sweden at the time of the wave 2 interview.5
The measure of subjective wellbeing considered in this study
is respondents self-assessed satisfaction with life as a whole. This Fig. 1. Self-assessment and anchoring vignette questions for life satisfaction in the
measure has attracted researchers and policy makers interest as a SHARE survey. Notes: Monetary values were PPP-adjusted across countries.
simple way to capture a subjective and global cognitive assessment Source: SHARE wave 2 questionnaire.

of peoples life, and has been widely used in empirical studies on

subjective wellbeing (Clark, 2014).6
Data on life satisfaction were collected in SHARE within
the COMPARE questionnaire module. After completion of the
main interview, in eleven countries a representative subset of
respondents7 were asked to complete on their own a paper-
and-pencil questionnaire containing self-assessment questions on
satisfaction with several health domains and with life as a whole. In
addition, for each domain brief descriptions of some relevant fea-
tures of the lives of hypothetical persons (the anchoring vignettes)
were also included after the self-assessment question. Respon-
dents were asked to rate the conditions of the person described in
each vignette with their own preferences, and the same response
scale was provided for both the self-assessment question and the
vignettes.
Fig. 1 reports the self-assessment question and the two
anchoring vignettes about life satisfaction included in the SHARE
questionnaire. The vignettes portray the situations of two

Fig. 2. Self-assessment and vignette evaluations for life satisfaction.

Source: SHARE.
5
I drop a small minority of individuals with missing values on any of the variables
included in the analysis. Information on life satisfaction, anchoring vignettes, and
hunger episodes is missing for less than 1% of the sample. ctitious characters named John and Carrie. Respondents were
6
Dolan et al. (2011), and the OECD (2013), compare several survey measures of asked to answer to all three questions on the same 5-point ordinal
subjective wellbeing. They distinguish between evaluative measures, such as the life scale going from very satised to very dissatised reported in
satisfaction question used in this study, measures of experienced wellbeing, such as
Fig. 1.8 Fig. 2 presents the distribution of the answers to these ques-
the Experience Sampling or the Day Reconstruction Method (Kahneman et al., 2004),
requiring people to complete emotional time diaries, and eudemonic measures of tions in the nal sample. There is substantial variation in vignette
self-realization, purpose and meaning in life. While all of these provide relevant evaluations, suggesting that reporting heterogeneity cannot be
information about subjective wellbeing, evaluative measures have been used most
frequently so far because they are easier to understand and can be accommodated
within general purpose surveys without posing a high burden on the respondent.
7 8
Bonsang and van Soest (2012a), show that the COMPARE sample is indeed rep- As in Angelini et al. (2015), I have recoded this scale as an ordinal variable going
resentative of the full SHARE sample. from 1 (very dissatised) to 5 (very satised).
 M. Bertoni / Journal of Health Economics 40 (2015) 4053 43

Table 1
Descriptive statistics on the estimation sample (N = 4950).

.1
Mean Std. dev.

.08
Self-assessed life satisfaction 3.915 0.766
Johns vignette 2.655 0.819

.06
Carries vignette 3.500 0.846

.04
Female 0.545 0.498
Born before 1930 0.099 0.299

.02
Born 19301934 0.105 0.307
Born 19351939 0.141 0.348
Born 19401944 0.211 0.408

0
Born 19451949 0.182 0.386
1920 1930 1940 1950 1960 1970
Born after 1949 0.262 0.440 Year
BE 0.123 0.329
DE SE NL ES
CZ 0.109 0.312
IT FR DK GR
DE 0.134 0.341
BE CZ PL
DK 0.135 0.342
ES 0.069 0.254
FR 0.050 0.218
GR 0.075 0.264 Fig. 3. Childhood hunger episodes, by country and year.
IT 0.103 0.305 Source: SHARE.
NL 0.067 0.250
PL 0.081 0.272
histories that are comparable across several European countries,
SE 0.053 0.225
Hunger 0.057 0.231
but there is also a potential threat of recall bias, that is common
Childhood SES 0.000 1.000 in retrospective studies (see Smith, 2009). Nonetheless, validation
Troubled parents 0.666 0.472 studies carried out by Garrouste and Paccagnella (2012), and Havari
Mother absent 0.033 0.178 and Mazzonna (2011), show that the state-of-the-art elicitation
Father absent 0.093 0.290
methods used in SHARELIFE, based on life history calendars, greatly
Siblings 0.831 0.375
Rural area 0.437 0.496 reduced the incidence of recall bias.
War exposure 0.452 0.498 Amongst early life conditions, my main interest lies in the
Relocation for war 0.034 0.181 experience of hunger episodes in childhood, i.e., from birth until
Dispossession 0.027 0.163
age 15. The SHARLIFE questionnaire explicitly asks respondents
Lived in orphanage 0.014 0.116
Lived with foster parents 0.011 0.104
to recall whether they experienced a period when they suffered
from hunger, and if so to indicate the exact years when the
In a couple 0.780 0.414
hunger episode started and stopped. In previous studies, Havari
Primary education 0.269 0.443
Secondary education 0.490 0.500 and Peracchi (2011), Pinger et al. (2011), Halmdienst and Winter-
Higher education 0.241 0.428 Ebmer (2013), Kesternich et al. (2014), and Attanasio et al. (2014),
With ADL limitations 0.082 0.275 validated the reliability of this self-reported measure against his-
With IADL limitations 0.129 0.336 torical events. I replicate these ndings in Fig. 3, where I report
Retired 0.510 0.500
Employed 0.288 0.453
the distribution of people in a hunger episode that started in child-
Wealth 6.335 3.153 hood in my sample, by country and year. Hunger episodes are most
Income 5.616 2.345 common during World War II years, from 1939 until 1945, but fur-
Notes: The table reports descriptive statistics for outcome variables, childhood con- ther patterns related to historical events can be traced out in the
ditions and adulthood conditions in the upper, middle and lower panel, respectively. data. For instance, among Spanish people hunger episodes start to
be reported already from the mid-1930s, in coincidence with the
Spanish Civil War, and spikes in hunger prevalence that correspond
easily ignored. Moreover, a global ordering of vignette evalua-
to the Greek famine of 1941/42, the Dutch famine of 1944 and the
tions (Murray et al., 2003) is detectable, as Carries condition is
German famine of 1945/48 are also present (see Pinger et al., 2011,
generally evaluated as describing a higher level of life satisfaction
for further details).10
than Johns.
I also exploit SHARELIFE to derive information on other child-
The second wave of SHARE also provides information on gender,
hood conditions. First, following Mazzonna (2011), I use principal
year and country of birth, educational levels (primary, secondary or
component analysis to extract a single index of childhood socio-
post-secondary qualications), on marital and employment status,
economic status from the following vector of commonly used
on health indicators such as suffering of one or more limitations
proxies (see Brunello et al., 2012): the number of books at home
in Activities of Daily Living (ADL) or in Instrumental Activities of
at age 10, occupation of the main breadwinner in the household at
Daily Living (IADL), on annual household income and on household
age 10, the number of rooms per person, the presence of an inside
net wealth.9 Descriptive statistics for these variables are shown in
toilet and of running water in the accommodation where the indi-
Table 1.
vidual was living at age 10. I also develop indicators for living in a
I derive information on childhood from the third wave of SHARE,
rural area at age 10,11 for whether the father or the mother were
SHARELIFE, that collects retrospective data on the entire lives of
absent from the household at age 10, for having troubled parents,
respondents, including a wealth of questions on early life con-
ditions. SHARELIFE has the advantage of providing data on life
10
Kesternich et al. (2013), further validate this self-reported measure by looking
at post-WWII Germany and showing that the regional and temporal variation in
9
Financial variables are measured in PPP terms at German prices of 2006, and self-reported hunger episodes closely mirrors differences in centralized food supply
divided by household size. As in Angelini et al. (2015), I transform nancial variables rations between German regions during that period of time.
11
using the arcsinh() function, allowing for zeroes and for negative values that show Neelsen and Stratmann (2011), examine the Greek famine of 194142, and show
up for wealth. that hunger incidence was more severe in urban than rural areas.
44 M. Bertoni / Journal of Health Economics 40 (2015) 4053
that smoked, drank or had mental health problems, and for hav-
ing any siblings while in childhood. Like Kesternich et al. (2014), I
also consider an indicator for exposure to war events in childhood,
that varies by cohort and country,12 but I am not able to distin-
guish between broad exposure to war and living in specic combat
areas, as they instead do. Finally, like Bohacek and Myck (2011), and
Halmdienst and Winter-Ebmer (2013), I also consider a set of other
relevant childhood events, including having lived in an orphanage
or with foster parents, being relocated during childhood because of
war, being dispossessed of the familys house, land, business or of
other properties during childhood. Descriptive statistics for these
variables are also reported in Table 1.
3. Empirical methodology
I exploit anchoring vignettes to estimate whether the expe-
rience of childhood hunger shifts the subjective thresholds
individuals use to report wellbeing on a discrete scale. This rst step
allows me to lter subjective evaluations from differential item
functioning, letting me evaluate the association between childhood
starvation and wellbeing later in life without issues of reporting
heterogeneity. Comparison with results derived in a setup that does
not consider the rst modelling step will shed light on the relevance
of rescaling bias.
I use the Heterogeneous Thresholds Ordered Probit model,
Hopit, introduced by King et al. (2004). Let the latent level of life
satisfaction of individual i, Yi , be modelled as follows:
Yi = Xi + i , i N(0, 1) (1) of the persons described in vignettes, so that the cutoffs described
In (1), Xi is a vector of individual observable variables that includes
exposure to hunger and the other childhood covariates described in
the previous section, while i is an error term that is independent of
Xi and follows a standard normal distribution.13 Yi is not observed
in the data, as individuals are asked to report their life satisfaction
on a 5-point ordinal scale. The following rule relates the observed
life satisfaction level Yi to Yi :
j1 j
Yi = j if i < Yi i , j = 1, . . ., 5 (2)
To model subjective response scales, thresholds  i are allowed to
depend on the same vector of individual covariates Xi included in
(1), according to the following specication:
i0 = ; i5 = +;
i1 =  1 Xi + i
(3)
j j1
i = i + exp( j Xi ), j = 2, 3, 4
i N(0, 2 )
In (3), the exponential specication is only needed to grant mono-
tonicity of the individual-specic thresholds. On the other hand,
since individuals could have a tendency to use high or low cutoffs
in all their assessments even conditional on the observables in Xi ,
an individual-specic random effect, i , is introduced linearly in
the thresholds equations, and allows for unobserved correlation
12
Like Kesternich et al. (2014), I consider as exposed to war in childhood all indi-
viduals born until 1945 in all countries except for Sweden and Denmark, that were
not affected by WWII, and Germany, where war ended in 1948, and who were
younger than 15 when the war started, i.e., in 1939 in all countries except for Spain,
where the civil war started in 1936.
13
Location and scale normalization are achieved by setting the constant term to
0 and the variance of the error term i to 1.
between the self-assessment and the vignettes evaluations carried
out by the same individual (see Kapteyn et al., 2007).14
Crucially, if the same variables entered in the equations for
the level and the cutoffs and only information about the self-
assessment Yi was available, parameters in and  1 would not
be identied, while parameters in  j , j = 2, 3, 4 would be identi-
ed only via the exponential functional form restriction.15 In the
Hopit model, identication exploits the fact that each individual
is also asked to rate vignette questions on the same measurement
scale used for the self-assessment, and relies on two assumptions.
The rst one, vignette equivalence, states that there are no differ-
ences in the perception of the level of life satisfaction described
by each vignette that are consistently related with respondents
characteristics, so that any systematic variation in responses to
anchoring questions only reects differences in reporting scales.
Hence, under vignette equivalence, perception of the level of sat-
, does not vary over
isfaction described by the kth vignette, Zki
individuals except for a random error, vki :
Zki = k + vki , vki N(0, v2 ) (4)
Vignettes are evaluated on the same 5-point ordinal scale used for
the self-assessment, and the observed rating of vignette k, Zki , is
as follows:
linked to Zki
j1 j
Zki = j if i < Zki i , j = 1, . . ., 5 (5)
More specically, the second assumption of the model, response
consistency, posits that individuals apply the same subjective
reporting scale to evaluate both their own conditions and the ones
in Eq. (3) apply to both Eqs. (2) and (5), linking the two components
of the model. Under vignette equivalence, the individual-specic
thresholds  i , characterized by the parameters in , are identied
by the variation in vignette evaluations. Thanks to response con-
sistency, these cutoffs pertain also to the self-assessment equation,
and the vector is identied by the variation in the self-reports.16
The validity of the two assumptions has been widely debated in
the literature, with mixed ndings. I will provide empirical evi-
dence in favour of the identication argument while discussing the
empirical results.
Before presenting the estimation outcomes, it is worth dis-
cussing the interpretation of the coefcients for childhood hunger
in the model discussed so far. While episodes of starvation in child-
hood pre-date wellbeing evaluation later in life, so that no issue of
reverse causation arises, causal interpretation may be jeopardized
by the presence of confounding factors. Table A2 in the Appendix
reports marginal effects of a Probit model that relates the set of
childhood conditions described in Section 2 to the occurrence of
childhood hunger. Starvation is more frequent among individu-
als from a low socio-economic background, and is associated with
several childhood events that could also have a direct impact on
late-life outcomes. On the one hand, to lessen concerns related with
omitted variables bias, in the next section I will show that results
are stable when I progressively introduce demographic controls,
the controls for family background and the ones for exposure to
negative childhood experiences included in Table A2. On the other
14
van Soest and Vonkova (2014), show that the inclusion of the random effect
strongly improves the ability of the Hopit model to reproduce the rankings of self-
assessment and vignette evaluations observed in the data.
15
Pudney and Shields (2000), suggest to reach identication of the model by
excluding some variables from the level equation, while Greene and Hensher (2010),
and Peracchi and Rossetti (2013), discuss identication via functional form restric-
tions.
16
The full model is estimated via maximum likelihood. The GLLAMM routine in
STATA is used to carry out the estimation.
 M. Bertoni / Journal of Health Economics 40 (2015) 4053
hand, I acknowledge that inclusion of this wide set of controls may
reduce the potential sources of variation in exposure to hunger,
and that differences along unobservable dimensions could still be
a relevant issue. Hence, I prefer to interpret coefcients related
to hunger as describing robust empirical associations rather than
causal effects.17 I will further comment on the interpretation of the
results while presenting the estimation outcomes.
4. Main empirical results
The empirical results described in this section are based on the
Hopit model presented in Section 3. I also compare these nd-
ings with a baseline model that does not allow for interpersonal
variation in reporting scales, akin to a standard Ordered Probit
model, to assess the relevance of rescaling bias. Irrespectively of
the set of covariates included, a formal likelihood ratio test strongly
rejects the baseline model vis--vis the heterogeneous threshold
one (p-value = 0.000), implying that the presence of heterogeneity
in reporting styles is always statistically relevant.
The various panels of Table 2 present the coefcients related
to hunger when more childhood controls are gradually included in
the model. In each panel, results for the baseline model with homo-
geneous cutoffs are shown in column 1, while column 2 reports
parameter estimates for the self-assessment equation in the Hopit
model, and columns 3, 4, 5 and 6 report estimates for the parame-
ters in the equation for threshold 1, 2, 3 and 4, respectively.
The key empirical nding is that exposure to hunger signif-
icantly affects the scale used to report life satisfaction. Looking
across all panels of Table 2, column 6 shows that there is a negative
and signicant shift in the upper discretizing threshold for people
who suffered of starvation early in life, i.e., they use the highest
point on the ordinal scale to rate the same level of satisfaction with
life that would receive a lower evaluation from people not exposed
to hunger, and the result is robust to the inclusion of controls for
other childhood conditions.18
I interpret this result in the light of the satisfaction tread-
mill model of hedonic adaptation (see Kahneman, 1999; Frederick,
2007). Since no absolute scale to evaluate wellbeing exists, people
carry out subjective evaluations with respect to individual-specic
reference points, which are assumed to depend on ones past
experiences. According to the satisfaction treadmill, exposure to
extreme deprivation in childhood may lead people to evaluate the
same life achievements more positively, as they could have devel-
oped lower aspirations for what having a satisfactory life means.19
This positive rescaling does not mean that people experienc-
ing hunger are more satised with their lives. Indeed, I nd that
the opposite is true. As shown in column 2 of Table 2, childhood
17
Other papers have exploited the sources of exogenous variation described in
Section 2, famines and wars, as instruments for childhood hunger (see Havari and
Peracchi, 2011; Pinger et al., 2011). However, no IV-Hopit model allowing instru-
mental identication of the effect of hunger on the scale used to report subjective
wellbeing is available so far. The derivation of such models is beyond the purposes
of this paper, and is left for future research.
18
In Panel B, the coefcient for hunger in the third cutoff is also negative and
signicant.
19
Since the hunger indicator I am using is self-reported, this coefcient is a lower
bound to the true coefcient on reporting scales if more pessimistic people are more
prone to report of having suffered from hunger. As a consequence, even the correc-
tion for reporting bias on the levels of subjective wellbeing is a lower bound to the
true one. Furthermore, Lewbel (2007), proves that, if a non-differential measure-
ment error assumption holds, potential misclassication problems in exposure to
hunger due to recall bias will lead to attenuation bias. Evidence of attenuation bias
for the effects of self-reported hunger in the SHARE data is present in Pinger et al.
(2011), who use famine IVs to correct for endogeneity. This further stresses that the
hunger coefcients I estimate are lower bounds to the true ones.
45
Fig. 4. Model predictions and counterfactual simulations. Notes: The upper bar
reports the distribution of life satisfaction predicted by the Hopit model presented
in Table 2C. The middle (lower) bar reports the counterfactual distributions of life
satisfaction that would hold if all respondents were given the reporting scale they
would have experienced had they (had they not) been exposed to hunger episodes
in childhood, leaving anything else unchanged.
Source: SHARE.
hunger is negatively associated to happiness levels later in life, and
this relationship is strongly signicant. This nding is robust to
the inclusion of covariates and is consistent with other studies,
but is derived in a more general framework that takes reporting
heterogeneity into account. Comparing results from the baseline
(column 1) and the Hopit model (column 2), it appears in all three
panels that failure to consider reporting heterogeneity will result
in estimates of the association between childhood hunger and life
satisfaction that are biased downwards. The p-value of a test for
joint signicance of hunger exposure in the four threshold equa-
tions is reported at the bottom of each panel: the null hypothesis
is strongly rejected in all cases, conrming the statistical relevance
of rescaling bias. This result is fully in line with the adaptive aspira-
tions hypothesis of Sen (1985a, 1985b), who argues that subjective
measures of wellbeing will only partially reect the welfare effects
of life events in presence of rescaling effects due to the adjustment
of aspirations.
To quantitatively assess the relevance of rescaling, I turn to a
counterfactual simulation based on the model presented in Panel C
of Table 2, that includes all childhood controls and that I will con-
sider as the baseline for all further analysis.20 Results are shown
in Fig. 4. The upper graph plots the distribution of life satisfac-
tion that is predicted by the Hopit model. I derive this gure by
rst computing for each respondent the predicted values of latent
life satisfaction, Y i , and of her specic cutoff points,  i , given her
individual characteristics, and then by plotting the resulting dis-
tribution of life satisfaction measured on the 5-point ordinal scale.
The middle and lower graphs, instead, show the counterfactual dis-
tributions of life satisfaction that would be observed if I assigned
to each individual the cutoff points she would have used had
she, or had she not, experienced childhood hunger, respectively,
while leaving all her other characteristics and her self-assessment
equation unaltered. The share of the sample reporting to be very
satised with life increases by a non-negligible 12% using the
reporting scales that would hold under exposure to hunger with
20
Results for the other two cases are very similar, both qualitatively and quanti-
tatively.
46 M. Bertoni / Journal of Health Economics 40 (2015) 4053
Table 2
Childhood hunger and self-reported wellbeing.
(1) (2)
Baseline Hopit
Panel A. Demographic controls
Hunger 0.298*** 0.381***
(0.075) (0.089)
Cutoffs (p-value) 0.001
Panel B. A + family background
Hunger 0.295*** 0.348***
(0.0787) (0.0903)
Cutoffs (p-value) 0.001
Panel C. B + other childhood experiences
Hunger 0.281*** 0.353***
(0.0781) (0.0929)
Cutoffs (p-value) 0.001
Notes: The table reports coefcients for exposure to childhood hunger from a Hopit model for life satisfaction. Controls included in Panel A are country, gender and cohort
dummies. Panel B adds childhood SES, having troubled parents, mother or father absent at age 10, having siblings, and living in a rural area. Panel C adds exposure to war,
relocation for war, dispossession, living in an orphanage, living with foster parents. The p-value reported in the bottom line of each panel refers to a test for joint signicance
of the hunger coefcients in the four cutoff equations. Observations: 4950.
*
Statistical signicance at 10% level of condence.
**
Statistical signicance at 5% level of condence.
***
Statistical signicance at 1% level of condence.
respect to the complementary case, while the exact same share of
people would report to be neither satised, nor dissatised.21
The full set of parameters estimated from the model that
includes all controls is presented in Table A3 in the Appendix.
First, it is reassuring to see that the same patterns described for
hunger hold for socio-economic status in childhood as well: kids
that start from a disadvantaged background are less satised with
life later on, but tend to rate the same situations more posi-
tively. In my view, this nding strengthens an interpretation in
terms of a satisfaction treadmill. On the other hand, other child-
hood conditions have limited relevance for both reporting scales
and life satisfaction, once differences in socio-economic status
and exposure to hunger episodes are taken into account. Signif-
icant differences on wellbeing levels and on reporting styles are
found across people born in different countries, and the patterns
are similar to those reported by Angelini et al. (2014). Finally,
the parameter associated with Johns vignette ( 1 ) is more nega-
tive than the one associated with Carries ( 2 ), consistently with
the global ordering of the two vignettes described above, and the
standard deviation of the individual random effects is statistically
signicant.
5. Discussing the identifying assumptions
Before showing some sensitivities and extensions to the main
results, it is crucial to discuss the validity of response consistency
and vignette equivalence.
Response consistency states that individuals adopt the same
subjective scale to evaluate satisfaction with their own life and the
situations described in each vignette. Several tests for this assump-
tion were proposed in the literature. Kapteyn et al. (2011), test
this hypothesis by conducting a survey experiment that relies on
longitudinal data on health conditions and self-reported health
status. They construct individual-specic vignettes illustrating the
21
Predictions obtained from the baseline model, that does not consider reporting
heterogeneity, are similar to the ones from the Hopit model, but the goodness of t
is worse: in the baseline model, an even larger share of respondents ends up in the
Satised With Life category. On the other hand, it is not surprising to see that the
top and bottom graph are very similar, as only a small fraction of the sample was
exposed to childhood hunger.
(3) (4) (5) (6)
Cutoff 1 Cutoff 2 Cutoff 3 Cutoff 4
0.0603 0.0323 0.088 0.197***
(0.103) (0.0679) (0.0569) (0.0531)
0.0582 0.0434 0.100* 0.175***
(0.105) (0.0681) (0.057) (0.0535)
0.0545 0.0400 0.0948 0.195***
(0.109) (0.0716) (0.0591) (0.0553)
past situation of each respondent, deriving the relevant informa-
tion from her answers to questions posed in previous waves of the
panel, and ask her to rate the vignettes. Comparing current vignette
evaluations with past self-assessments, they nd mixed support
for response consistency, depending on the domain of interest. A
growing number of studies suggests instead to validate response
consistency using objective measures of the concept of interest. For
instance, van Soest et al. (2011), consider self-perceived drinking
problems, and show that peoples evaluation of their own drinking
problems and of vignettes describing a drinking level equal to their
own one are strongly aligned, supporting response consistency.22
Angelini et al. (2014), however, highlight that such tests can hardly
be implemented when it comes to evaluating satisfaction with life
in general, as this is a multidimensional concept that cannot be
unquestionably measured using a single objective indicator. They
instead show that, in the SHARE data, people whose personal sit-
uation in each of the domains described in a vignette is analogous
to the situation depicted in the vignette give equal evaluations of
satisfaction with their own life and of that of the person described
in the vignette, as implied by response consistency.
Even when response consistency holds, identication requires
also the vignette equivalence assumption, stating that, except for
random errors, the latent level of life satisfaction of the person
described in a vignette is perceived equally by all respondents,
irrespectively of their personal characteristics. To give an exam-
ple, according to vignette equivalence, the fact that the person
described in a vignette has grandchildren should not lead to differ-
ences in the average perception of the latent level of satisfaction of
the vignette between respondents with and without grandchildren.
To provide evidence in favour of this assumption, I re-estimate
the baseline model after splitting the sample along the follow-
ing observed dimensions, which are mentioned in the vignettes
and may be related to the perception of their latent satisfaction
level: gender, age, having grandchildren, having chronic health
conditions, wealth level, carrying out recreational activities. Table 3
shows the coefcients for the vignette levels ( 1 and  2 ) estimated
22
A similar test was proposed by Datta Gupta et al. (2010), and Bago dUva et al.
(2011), for health, cognitive functions and mobility limitations. Their results do not
unambiguously point against or in favour of response consistency.
 M. Bertoni / Journal of Health Economics 40 (2015) 4053 47

Table 3
Estimated vignette levels across subsamples dened by observable characteristics.

Gender Born before 1940 Grandchildren

Male Female Yes No Yes No

1 2.056 1.962 1.98 1.986 2.069 2.099

95% c.i. (1.740, 2.372) (1.681, 2.243) (1.484, 2.476) (1.716, 2.256) (1.785, 2.353) (1.766, 2.432)
2 0.88 0.755 0.9 0.731 0.894 0.877
95% c.i. (0.574, 1.186) (0.482, 1.028) (0.412, 1.388) (0.469, 0.993) (0.618, 1.170) (0.554, 1.200)

Obs. 2252 2698 1709 3241 2989 1961

Chronic health conditions Wealth above country median Recreational activities

Yes No Yes No Yes No

1 1.823 2.215 2.336 1.892 2.278 1.877

95% c.i. (1.503, 2.143) (1.916, 2.514) (2.019, 2.653) (1.594, 2.190) (1.960, 2.596) (1.578, 2.176)
2 0.643 0.972 1.159 0.663 0.996 0.749
95% c.i. (0.331, 0.955) (0.684, 1.260) (0.853, 1.465) (0.373, 0.953) (0.691, 1.301) (0.457, 1.041)

Obs. 2168 2782 2474 2476 2527 2423

Notes: The table reports the coefcient for the rating levels of each vignette question from a Hopit model for life satisfaction like the one presented in Table 2C, together with
their 95% condence intervals. Estimation is carried out separately on each of the subsamples listed in the table.

in the split samples, with their 95% condence intervals. Since the
coefcients are very similar across the split samples and condence
intervals always overlap, no systematic perception difference in
the latent satisfaction level of the vignettes along these observable
dimensions is detected, providing strong support to the validity of
vignette equivalence.23 Several other tests for this assumption have
been proposed in the literature: I discuss them in the Appendix,
where I also show that they are satised in my sample.
6. Sensitivities and extensions
This section proposes some extensions and robustness checks
to the main empirical results. First, it is interesting to understand
whether the long-run associations between childhood hunger and
late-life wellbeing levels and reporting scales are mediated by
adulthood conditions that depend on hunger episodes and are com-
monly known to determine satisfaction with life. Evidence from
SHARE presented by Havari and Peracchi (2011), and Pinger et al.
(2011), shows that childhood hunger is related to lower educa-
tional achievement and worse health and labour market outcomes
in adulthood. In turn, using the SHARE data Angelini et al. (2014),
show that these characteristics are linked with both the level and
the scale used to report life satisfaction. I thus introduce a set of
adulthood outcomes as controls in the baseline model,24 and eval-
uate by how much results change with respect to the baseline. The
sign, magnitude, and signicance of the coefcients of hunger on
both reporting scales and latent wellbeing are unchanged,25 sug-
gesting that the role of the mediating variables is not central.
23
I thank one referee for suggesting this test.
24
The variables considered are taken from the second wave of SHARE and are
described in Section 2. Although Angrist and Pischke (2008), refer to mediators
as bad controls, a similar strategy is exploited, for instance, by Giuliano and
Spilimbergo (2014), to assess the effects of growing up in a recession on beliefs
in adulthood, by Frijters et al. (2011), and Layard et al. (2013), in the context
of modelling life-cycle dynamics of subjective wellbeing, and by Halmdienst and
Winter-Ebmer (2013), to seek for mediators of the effects of early-life shocks on
health later in life.
25
The coefcient for hunger in the baseline model is equal to 0.278 (s.e.: 0.0795),
the one in the level equation for the Hopit model to 0.369 (s.e.: 0.0960), and the
one in the equation for Cutoff 4 in the Hopit model to 0.206 (s.e.: 0.055). The other
hunger coefcients are not signicant. Counterfactual simulations obtained from
this model (not shown) are also in line with the ones presented in Fig. 4.
Next, I consider potential issues of dynamic selection. Looking at
the Dutch potato famine of 1846/47, Lindeboom et al. (2010), show
that bad nutritional conditions in infancy negatively affect mortal-
ity age. On the other hand, all results I have reported so far are
conditional on surviving until age 50+. As a consequence, if more
optimistic people survive longer after an episode of hardship, or
if survivors have characteristics that make them more prone to
report higher levels of life satisfaction, sample selection may have
biased my results towards nding positive rescaling. To understand
whether survival is linked to reporting styles and life satisfaction
levels, I estimate a Probit regression for the probability of deceas-
ing between the second and the third waves of SHARE, conducted
in 2006/07 and in 2008/09, respectively, on vignette evaluation,
self-reported satisfaction with life and a set of controls measured
at wave 2.26 Results are presented in Table A4 in the Appendix,
and do not suggest that reporting styles and life satisfaction lev-
els are correlated with death probabilities. Although these results
are based on surviving in old age, and not on surviving to hunger
episodes in childhood, they are still reassuring about concerns of
dynamic selection on the basis of reporting styles and life satis-
faction levels. Furthermore, results are fully comparable with the
baseline also when I drop from the sample individuals born before
1930 or before 1935, amongst whom dynamic selection should
be more serious.27 As a nal check, in line with Kesternich et al.
(2014), I regress both an indicator for survival of the parents of
the respondent until the SHARE interview and the average age
of death of the parents on an indicator for whether the individ-
ual was exposed to hunger in childhood, an indicator for whether
the individual reports that his or her parents had mental health
problems, and an interaction term between the two, controlling
for country of birth dummies, family background indicators and
war exposure. For both outcomes, the coefcient related with
the interaction term is small in magnitude and not statistically
signicant: I cannot detect any differential effect of exposure to
26
The sample considered excludes wave 2 respondent for whom no information
is available at wave 3 because of unit non-response. Similar results (not shown) are
obtained when I consider survival until wave 4, conducted in 2011/12. In that case,
however, I have to drop Greece, as it did not take part in the fourth wave of the
survey.
27
Results are available from the author. If recall bias is also stronger for older
individuals, results from the subsample of younger individuals should be less subject
to this issue as well.
48 M. Bertoni / Journal of Health Economics 40 (2015) 4053

Table 4
Heterogeneity in hunger coefcients on self-reported wellbeing across groups.

(1) (2) (3) (4) (5) (6)

Baseline Hopit Cutoff 1 Cutoff 2 Cutoff 3 Cutoff 4

Panel A. Gender
Males (Obs.: 2252) 0.199* 0.228* 0.0954 0.0999 0.102 0.252***
(0.112) (0.136) (0.152) (0.0957) (0.0854) (0.0818)
Females (Obs.: 2698) 0.388*** 0.512*** 0.00362 0.00933 0.0841 0.165**
(0.110) (0.130) (0.164) (0.109) (0.0828) (0.0774)

Panel B. Siblings
Yes (Obs.: 4115) 0.229*** 0.297*** 0.0830 0.0416 0.0986 0.223***
(0.0864) (0.103) (0.117) (0.0766) (0.0655) (0.0617)
No (Obs.: 835) 0.473** 0.573** 0.142 0.126 0.0513 0.0209
(0.187) (0.223) (0.271) (0.166) (0.137) (0.131)

Panel C. SES
High (Obs.: 2473) 0.0684 0.0315 0.0466 0.139 0.0399 0.255**
(0.144) (0.186) (0.202) (0.121) (0.104) (0.102)
Low (Obs.: 2477) 0.346*** 0.454*** 0.126 0.0574 0.149** 0.138**
(0.0939) (0.111) (0.131) (0.0858) (0.0721) (0.0678)

Notes: The table reports coefcients for exposure to childhood hunger from a Hopit model for life satisfaction like the one presented in Table 2C. Estimation is carried out
separately for males and females in Panel A, for individuals with siblings and single children in Panel B, and for individuals above and below the median level of socio-economic
status in Panel C. Other childhood covariates included are described in Section 2. A test for joint signicance of the hunger coefcients in the four cutoff equations is rejected
at the 5 percent level of condence in all cases except for individuals without siblings.
*
Statistical signicance at 10% level of condence.
**
Statistical signicance at 5% level of condence.
***
Statistical signicance at 1% level of condence.

hunger on survival probabilities and age of death for parents who,
according to their offspring, did or did not have mental health prob-
lems, weakening concerns of dynamic selection linked with mental
wellbeing.
I next explore potential sources of heterogeneity in the associa-
tion between hunger and wellbeing, after correcting for differential
item functioning.28 I rst ask whether there are gender differences
in the association between childhood hunger and late-life wellbe-
ing. Evidence coming mainly from developing countries29 shows
that the impacts of negative childhood shocks are larger for girls
than for boys, and that this is partly explained by differential com-
pensation mechanisms carried out by parents, who tend to favour
survival of sons vs. daughters. Panel A of Table 4 reports results
when I split the sample by gender. The association between hunger
and late-life wellbeing is stronger for girls, and only marginally
signicant for boys. According to Graham and Oswald (2010), the
dynamics of adaptation to life events depend on hedonic capital,
i.e., the stock of psychological resources available to an individual.
In their theoretical model, hedonic capital depends on a collection
of stock-like variables affecting individual wellbeing, like interper-
sonal relationships, personality traits, and social status, which help
to smooth wellbeing when a shock hits. In this context, the pres-
ence of siblings may contribute to the generation of hedonic capital:
sharing the same negative experiences and encouraging each other,
children with siblings should be able to cope with starvation bet-
ter than single children. In turn, higher income may help to buffer
the negative inuence of hunger episodes. I provide evidence in
favour of these two hypothesis by splitting the sample between
single children and individuals with siblings, and between respon-
dents with an index of socio-economic status below and above
the median level in the sample. Results presented in Panel B of
Table 4 show that the correlation between starvation and happiness
is stronger for single children, and Panel C of Table 4 conrms that
people from higher socio-economic background suffer less from
hunger episodes in childhood, consistently with a buffering effect
of income.30
I provide further robustness checks in Table A5 in the Appendix.
Panel A excludes from the sample people who migrated from coun-
try of birth, an outcome that may be linked to the experience of
hunger, and results are still in line with the ones presented in
Table 3. Finally, in Panel B and Panel C I respectively drop peo-
ple from countries not affected by World War II (i.e., Sweden and
Denmark) and people born after 1949, as episodes of hunger were
less common within these groups. Again, the patterns detected are
comparable with the baseline.
7. Conclusions
Policy makers and economists are keen on developing a deeper
understanding of factors promoting happiness throughout life.
Using data from eleven European countries, this paper studies the
association between childhood hunger, an episode of extremely
severe deprivation that has been widely analyzed by economists,
and self-reported wellbeing later in life.
Using anchoring vignettes, I contribute to the literature by
showing that people exposed to hunger in childhood adopt lower
standards to evaluate satisfaction with life, as predicted by the
satisfaction treadmill model of hedonic adaptation. This is a new
nding, that sheds light on the determinants of the reference points
used to evaluate subjective wellbeing, on which little is known so
far, and questions the standard assumption of no systematic het-
erogeneity in the way people interpret the discrete scales on which
they are commonly asked to report happiness. I am in turn able
to provide the rst empirical evidence on the negative long-run
association between childhood hunger and the levels of adulthood
wellbeing that is not affected by rescaling issues, and I show that

28 30
The main focus is on the self-assessment equation, and results are purely The incidence of hunger is of course lower amongst people coming from well-
descriptive. off families, but exogenous events generate variability in exposure to hunger even
29
For instance, see Maccini and Yang (2009), for a review. within the well-off.
 M. Bertoni / Journal of Health Economics 40 (2015) 4053 49

Notes: The table reports attrition rate and mean values of several variables measured at baseline (wave 2), for the full wave 2 COMPARE sample and for the selected wave 3 sample considered in the analysis, for each country
estimates that fail to consider reporting heterogeneity are biased

vignette
Carries
towards nding a positive correlation.

3.4
3.4
3.4
3.4
3.3
3.3
3.4
3.5
3.3
3.3
3.2
3.2
3.8
3.8
3.3
3.3
3.6
3.7
3.7
3.6
3.5
3.6

3.5
3.5
These results raise awareness about the importance of consid-
ering rescaling when subjective indicators are used to measure the
welfare consequences of life events. Failure to do so could lead

vignette
to confound differences in latent wellbeing and heterogeneity in

Johns
reporting scales. If these two forces ran in opposite directions, as

2.9
2.8
2.4
2.3
2.7
2.7
2.5
2.4
2.4
2.4
2.4
2.4

2.7
2.7
2.5
2.5
2.8
2.8
2.6
2.7

2.7
2.7
3.0
3.0
in the case analyzed in this paper, rescaling could have worrisome

Self-assessed life
implications for conclusions drawn from subjective wellbeing mea-
sures. As shown in the paper, anchoring vignettes are a simple
survey instrument that can help to overcome scale biases. This

satisf.
study also provides further support to the validity of the assump-

3.9

4.2
4.3
4.1
4.1
3.8
3.8
3.6
3.6
3.8
3.8
4.3
4.4
3.6
3.6
3.9

3.7
3.7
3.7
3.7

3.9
3.9
4.0

4.0
tions behind this methodology, widening its applicability, and
helping to gain a better comprehension of the determinants of
subjective wellbeing.

% retired

52.7
58.4
57.4
33.8
36.8
34.3
33.2
53.6
56.4
56.7
56.8
47.8
47.1
39.8
39.8
51.6
52.3
67.3
68.4
55.4
51.8

51.3
54.0

51.0
Appendix.

% employed
A.1. Other tests for the vignette equivalence assumption

27.3
27.8
37.3
38.8
40.7
36.2
26.2
25.3
18.2
17.6
27.5
30.8
43.2

33.3
31.7
21.1
19.7
27.2
26.1
18.8
21.7

28.8
45.0

29.0
Several test for the vignette equivalence hypothesis are present
in the literature. First, Murray et al. (2003), and Rice et al. (2011),
highlight that a minimal requirement for vignette equivalence to

couple
% in a

80.3

78.4
79.1
84.9
83.6
81.4
80.9
83.6
85.7
70.8
69.2
81.9

71.3
69.1

75.2
70.1
70.5
75.9
76.7

77.8
81.0

82.0

75.0

78.0
hold is that individuals systematically order the vignettes in the
same way.31 Kristensen and Johansson (2008), further suggest that,
in a cross-country study, vignette equivalence is unlikely to hold if
% with ADL
limitations

results from models estimated pooling all countries and separat-

ing groups of countries that share different sets of values and social
6.6
5.7
8.2
6.1
3.3
2.8
8.8
8.2
8.4
8.9
9.8
8.4
5.9
5.2
5.8
4.6
10.3
11.3
7.5
6.6
26.2
23.8

8.8
8.2
norms are very different from one another. In my sample, only 8%
of the respondents rate vignettes inconsistently with the global
ordering, with little cross-country heterogeneity. Results are fully
education

comparable with the baseline even when I replicate the analysis

% higher
Attrition between the full COMPARE sample (wave 2) and the subsample of survivors in SHARELIFE (wave 3).

dropping individuals who rate vignettes inconsistently, suggesting

29.9
33.4
33.4
33.5
27.8
28.2
13.4
12.4
9.2
8.3
25.3
26.8
39.5
41.7
21.1
19.4
23.4
22.4
11.2
12.8
17.6
19.9

23.5
24.1
that the ndings are not driven by violations of vignette equiv-
alence that would result in inconsistent vignette ordering. As in
Rice et al. (2011), and Angelini et al. (2014), I also re-estimate the
(income)

baseline model separately for three groups of countries that share

10.13
10.24
10.31
10.36
10.34
9.41
9.35
9.74

10.26
10.34

10.25
9.51
9.45

9.48
9.51

9.86
9.87
10.07

9.70

10.20

10.01
10.03

9.06
9.06
Log

similar values according to the Inglehart-Wezel values map32 :

ex-communist countries (Czech Republic and Poland), catholic
countries (Belgium, France, Italy and Spain) and protestant coun-
64.86
64.55
65.54
65.13
61.63
62.22
64.18
64.35
64.98
64.84
64.28

64.35
64.13
64.63
64.53
65.51
65.87
64.31
64.15
62.93

64.25
64.04

61.50

64.40
tries (Denmark, Germany, the Netherlands and Sweden), leaving
Age

Germany as the baseline country in each group. The estimated

vignette levels and the country dummies from the pooled and split
samples are in line with each other in terms of both ordering and
% female

magnitude,33 conrming the validity of vignette equivalence in my

52.7

54.1
52.5
51.7
52.3
52.9
55.3
53.6
53.3
55.1
54.8
54.3
54.7
52.2
54.8
53.6
54.1
58.5
58.6
56.4
57.3

54.2
54.5
52.0

data.34
Attrition
rate (%)

A.2. Tables
38.1

40.2

28.9

24.2

29.8

27.5

25.3

22.5

37.4

27.5

30.7
33.0

See Tables A1A5.

N. obs

1103
683
440
263
482
323
478
340
666
505
356
250
926
671
498
372
812
629
850
532
527
382

7138
4950
Wave 2
Wave 3
Wave 2
Wave 3
Wave 2
Wave 3
Wave 2
Wave 3
Wave 2
Wave 3
Wave 2
Wave 3
Wave 2
Wave 3
Wave 2
Wave 3
Wave 2
Wave 3
Wave 2
Wave 3
Wave 2
Wave 3

Wave 2
Wave 3
Sample

31
Still, Bago dUva et al. (2011), highlight that interpersonal differences in the lev-
els of vignette evaluations are not inconsistent with equal ranking of the vignettes.
and for the pooled sample.

32
See Kristensen and Johansson (2008). I drop Greece for this analysis because it
is an orthodox country.
33
All ndings also hold when I split countries according to the tertiles of the HDI
Czech Republic

index, as suggested by Rice et al. (2011).

Netherlands

34
Bago dUva et al. (2011), and Peracchi and Rossetti (2013), propose formal sta-
Denmark
Germany

tistical test for the joint validity of the two assumptions. Still, both tests are based
Belgium
Country

Sweden
Table A1

Greece

Poland
France

on the further assumptions of correct model specication and no omitted variables,

Spain

Total
Italy

and consequently they may reject the null hypothesis for reasons that differ from
failure of the two identifying conditions.
50 M. Bertoni / Journal of Health Economics 40 (2015) 4053

Table A2
Probit model for childhood hunger.

Female 0.00253 Rural area 0.0146***

(0.00301) (0.00359)
Childhood SES 0.00740*** War exposure 0.0252***
(0.00211) (0.00908)
Troubled parents 0.0112*** Relocation for war 0.0397**
(0.00314) (0.0155)
Mother absent 0.00357 Dispossession 0.115***
(0.00706) (0.0311)
Father absent 0.0396*** Lived in orphanage 0.0730**
(0.0118) (0.0365)
Siblings 0.000124 Lived with foster parents 0.0109
(0.00424) (0.00746)

Cohort dummies Yes

Country dummies Yes
Observations 4950

Notes: The table reports marginal effects from a Probit regression. Dependent variable: suffering of hunger in childhood.
*
Statistical signicance at 10% level of condence.
**
Statistical signicance at 5% level of condence.
***
Statistical signicance at 1% level of condence.

Table A3
Childhood conditions and self-reported wellbeing.

(1) (2) (3) (4) (5) (6)

Baseline Hopit Cutoff 1 Cutoff 2 Cutoff 3 Cutoff 4

Hunger 0.281*** 0.353*** 0.0545 0.0400 0.0948 0.195***

(0.0781) (0.0929) (0.109) (0.0716) (0.0591) (0.0553)
Childhood SES 0.104*** 0.163*** 0.0310 0.00607 0.00787 0.0317**
(0.0201) (0.0253) (0.0334) (0.0212) (0.0153) (0.0129)
Troubled parents 0.0569 0.0995** 0.0479 0.00474 0.0181 0.0352
(0.0378) (0.0469) (0.0589) (0.0372) (0.0282) (0.0249)
Mother absent 0.0802 0.0865 0.165 0.150 0.0920 0.0109
(0.104) (0.127) (0.169) (0.0945) (0.0809) (0.0678)
Father absent 0.0383 0.0459 0.00322 0.0462 0.111** 0.0678
(0.0649) (0.0799) (0.107) (0.0664) (0.0464) (0.0424)
Siblings 0.150*** 0.149** 0.143* 0.0280 0.0489 0.0415
(0.0479) (0.0586) (0.0747) (0.0465) (0.0355) (0.0309)
Rural 0.0646* 0.0116 0.0986* 0.117*** 0.0349 0.0143
(0.0367) (0.0453) (0.0599) (0.0371) (0.0274) (0.0236)
War exposure 0.0278 0.0357 0.208* 0.112* 0.0191 0.0164
(0.0636) (0.0789) (0.113) (0.0674) (0.0474) (0.0406)
Relocation for war 0.000518 0.164 0.172 0.0730 0.131* 0.101
(0.101) (0.132) (0.142) (0.0914) (0.0794) (0.0658)
Dispossession 0.0178 0.0368 0.0917 0.0296 0.0910 0.0810
(0.109) (0.135) (0.182) (0.118) (0.0767) (0.0698)
Lived in orphanage 0.0606 0.0527 0.384 0.130 0.0749 0.0720
(0.152) (0.189) (0.313) (0.168) (0.111) (0.0971)
Lived with foster parents 0.0409 0.0397 0.104 0.0316 0.131 0.0713
(0.174) (0.227) (0.296) (0.162) (0.142) (0.111)
Female 0.0862** 0.147*** 0.0653 0.0510 0.0133 0.0530**
(0.0343) (0.0426) (0.0547) (0.0340) (0.0256) (0.0221)
Born before 1930 0.206*** 0.104 0.207* 0.0701 0.0656 0.0740
(0.0719) (0.0881) (0.122) (0.0747) (0.0531) (0.0479)
Born 19301934 0.0919 0.0961 0.206 0.0747 0.0965* 0.0307
(0.0766) (0.0956) (0.130) (0.0790) (0.0570) (0.0497)
Born 19351939 0.113 0.261*** 0.280** 0.112 0.0414 0.0138
(0.0720) (0.0911) (0.124) (0.0766) (0.0545) (0.0463)
Born 19401944 0.0396 0.0955 0.0840 0.0687 0.0246 0.00338
(0.0673) (0.0845) (0.123) (0.0720) (0.0514) (0.0425)
Born after 1949 0.172*** 0.205*** 0.0784 0.0761 0.00965 0.00847
(0.0519) (0.0636) (0.0892) (0.0541) (0.0388) (0.0323)
SE 0.598*** 0.635*** 0.546*** 0.0243 0.113 0.311***
(0.102) (0.135) (0.174) (0.0997) (0.0763) (0.0692)
NL 0.291*** 0.555*** 0.652*** 0.312*** 0.0147 0.0413
(0.0838) (0.120) (0.140) (0.0940) (0.0604) (0.0536)
 M. Bertoni / Journal of Health Economics 40 (2015) 4053 51

Table A3 (Continued )

(1) (2) (3) (4) (5) (6)

Baseline Hopit Cutoff 1 Cutoff 2 Cutoff 3 Cutoff 4

ES 0.179** 0.213** 0.760*** 0.0893 0.467*** 0.161***

(0.0827) (0.103) (0.137) (0.0838) (0.0689) (0.0530)
IT 0.479*** 0.321*** 0.886*** 0.221*** 0.200*** 0.0286
(0.0728) (0.0914) (0.125) (0.0771) (0.0543) (0.0496)
FR 0.197** 0.151 0.639*** 0.0350 0.0101 0.0745
(0.0898) (0.117) (0.148) (0.0882) (0.0637) (0.0644)
DK 0.572*** 0.132 0.324* 0.0188 0.0893 0.161***
(0.0808) (0.103) (0.193) (0.108) (0.0609) (0.0490)
GR 0.479*** 0.529*** 0.773*** 0.264*** 0.0104 0.608***
(0.0797) (0.0971) (0.137) (0.0879) (0.0558) (0.0603)
BE 0.0536 0.218** 0.442*** 0.0860 0.284*** 0.286***
(0.0689) (0.0870) (0.129) (0.0783) (0.0544) (0.0440)
CZ 0.455*** 0.631*** 0.0269 0.0796 0.0464 0.131***
(0.0713) (0.0894) (0.149) (0.0899) (0.0502) (0.0450)
PL 0.366*** 0.487*** 0.614*** 0.229*** 0.204*** 0.167***
(0.0787) (0.0964) (0.139) (0.0864) (0.0595) (0.0506)
Constant 4.181*** 0.590*** 0.204*** 0.784***
(0.189) (0.0934) (0.0642) (0.0542)
Baseline model cutoffs 3.638*** 2.117*** 0.953*** 0.862***
(0.104) (0.0895) (0.0847) (0.0850)
1 1.988*** 2.070***
(0.0897) (0.109)
2 0.793*** 0.884***
(0.0855) (0.106)
v 0.0194 0.013
(0.0195) (0.0198)
 0.454*** 0.406***
(0.0183) (0.0182)

Notes: The table reports full estimation outcomes for the Hopit model for life satisfaction introduced in Table 2C. The p-value reported in the bottom line refers to a test for
joint signicance of the hunger coefcients in the four cutoff equations. Observations: 4950.
*
Statistical signicance at 10% level of condence.
**
Statistical signicance at 5% level of condence.
***
Statistical signicance at 1% level of condence.

Table A4
Mortality across SHARE waves 2 and 3, vignette evaluation and life satisfaction.

(1) (2)

Johns vignette 0.002 0.001

(0.002) (0.002)
Carries vignette 0.001 0.001
(0.002) (0.002)
Life satisfaction 0.001
(0.002)

Wave 2 covariates Yes Yes

Country dummies Yes Yes
Cohort dummies Yes Yes

Observations 5459 5459

Notes: The table reports marginal effects from Probit regressions. Dependent variable: deceased in SHARE wave 3. The regression controls for country and cohort xed effects,
gender, education, marital status, employment status, income, wealth and health, all measured at wave 2.
*
Statistical signicance at 10% level of condence.
**
Statistical signicance at 5% level of condence.
***
Statistical signicance at 1% level of condence.

Table A5
Hunger effects on self-reported wellbeing sensitivities.

(1) (2) (3) (4) (5) (6)

Baseline Hopit Cutoff 1 Cutoff 2 Cutoff 3 Cutoff 4

Panel A. Excluding migrants

Hunger 0.312*** 0.394*** 0.0590 0.0279 0.0892 0.208***
(0.0799) (0.0946) (0.111) (0.0736) (0.0604) (0.0568)
Observations 4873
Cutoffs (p-value) 0.001

Panel B. War countries only

Hunger 0.295*** 0.351*** 0.0403 0.0562 0.0989* 0.181***
(0.0787) (0.0944) (0.111) (0.0725) (0.0598) (0.0559)
Observations 4017
Cutoffs (p-value) 0.003
52 M. Bertoni / Journal of Health Economics 40 (2015) 4053
Table A5 (Continued )
(1) (2)
Baseline Hopit
Panel C. 19201949 cohorts
Hunger 0.258*** 0.315***
(0.0798) (0.0960)
Observations 3652
Cutoffs (p-value) 0.001
Notes: Each panel reports coefcients for exposure to childhood hunger from a Hopit model for life satisfaction like the one presented in Table 2C. Panel A excludes people
who migrated from country of birth. Panel B excludes people from Sweden and Denmark. Panel C excludes individuals born after 1949. Other childhood covariates included
are described in Section 2. The p-value reported in the bottom line of each panel refers to a test for joint signicance of the hunger coefcients in the four cutoff equations.
*
Statistical signicance at 10% level of condence.
**
Statistical signicance at 5% level of condence.
***
Statistical signicance at 1% level of condence.
References
Akbulut-Yuksel, M., 2009. Children of war: the long-run effects of large-scale phys-
ical destruction and warfare on children. IZA Discussion Paper 4407.
Almond, D., Currie, J., 2011. Killing me softly: the fetal origins hypothesis. The Journal
of Economic Perspectives 25 (3), 153172.
Angelini, V., Cavapozzi, D., Paccagnella, O., 2011. Dynamics of reporting work disabil-
ity in Europe. Journal of the Royal Statistical Society: Series A 174 (3), 621638.
Angelini, V., Cavapozzi, D., Corazzini, L., Paccagnella, O., 2012. Age, health and
life satisfaction among older Europeans. Social Indicators Research 105 (2),
293308.
Angelini, V., Cavapozzi, D., Corazzini, L., Paccagnella, O., 2014. Do Danes and Italians
rate life satisfaction in the same way? Using vignettes to correct for individual-
specic scale biases. Oxford Bulletin of Economics and Statistics 76 (5), 643666.
Angelini, V., Mierau, J.O., 2012. Childhood health and the business cycle: evidence
from western Europe. HEDG Working Papers n. 12/28.
Angrist, J.D., Pischke, J.S., 2008. Mostly Harmless Econometrics: An Empiricists Com-
panion. Princeton University Press, Princeton, New Jersey.
Attanasio, O., Brugiavini, A., Trevisan, E., Weber, G., 2014. The consequences of
nancial hardship (and recessions) on income and welfare. In: Brugiavini, A.,
Weber, G. (Eds.), Longer-term Consequences of the Great Recession on the Lives
of Europeans. Oxford University Press, Oxford.
Bago dUva, T., Lindeboom, M., ODonnell, O., van Doorslaer, E., 2008. Does reporting
heterogeneity bias the measurement of health disparities? Health Economics
17 (3), 351375.
Bago dUva, T., Lindeboom, M., ODonnell, O., van Doorslaer, E., 2011. Slipping
anchor?: testing the vignettes approach to identication and correction of
reporting heterogeneity. Journal of Human Resources 46 (4), 875906.
Bohacek, R., Myck, M., 2011. Long shadows of history: persecution in central Europe
and its labor market consequences. IZA Discussion Paper 6130.
Bonsang, E., van Soest, A., 2012a. Satisfaction with social contacts of older Europeans.
Social Indicators Research 105 (2), 273292.
Bonsang, E., van Soest, A., 2012b. Satisfaction with job and income among older indi-
viduals across European countries. Social Indicators Research 105 (2), 227254.
Brunello, G., Weber, G., Weiss, C.T., 2012. Books are forever: early life conditions,
education and lifetime earnings in Europe. IZA Discussion Paper n. 6386.
Clark, A., 2014. Is happiness the best measure of well-being? In: Oxford Handbook
of Well-Being and Public Policy. Oxford University Press, Oxford (Chapter 16).
Clark, A., Etil, F., Postel-Vinay, F., Senik, C., van der Straeten, K., 2005. Heterogeneity
in reported well being: evidence from twelve European countries. The Economic
Journal 115 (502), C118C132.
Comin, F., Bagolin, I., Porsse, M., 2004. Adaptive preferences: a problem or a good
guide? In: Proceedings of the 4th International Conference on the Capability
Approach: Enhancing Human Security, 57 September 2004, University of Pavia,
Italy.
Cunha, F., Heckman, J.J., 2007. The technology of skill formation. The American
Economic Review 97 (2), 3147.
Currie, J., 2009. Healthy, wealthy, and wise: socioeconomic status, poor health in
childhood, and human capital development. Journal of Economic Literature 47
(1), 87122.
Datta Gupta, N., Kristensen, N., Pozzoli, D., 2010. External validation of the use of
vignettes in cross-country health studies. Economic Modelling 27 (4), 854865.
Deaton, A., 2008. Income, health and wellbeing around the world: Evidence From
The Gallup World Poll. The Journal of Economic Perspectives 22 (2), 5372.
Dolan, P., Layard, R., Metcalfe, R., 2011. Measuring subjective wellbeing for public
policy: recommendations on measures. CEP Special Paper n. 23 March 2011.
Dolan, P., Peasgood, T., White, M., 2008. Do we really know what makes us happy?
A review of the economic literature on the factors associated with subjective
well-being. Journal of Economic Psychology 29 (1), 94122.
Frederick, S., 2007. Hedonic treadmill. In: H-Baumeister (Encyc)-45348.qxd., pp.
419420.
Frey, B.S., Stutzer, A., 2002. What can economists learn from happiness research?
Journal of Economic Literature 40 (2), 402435.
Frijters, P., Johnston, D.W., Shields, M.A., 2011. Destined for (un)happiness: does
childhood predict adult life satisfaction? IZA Discussion Paper 5819.
(3) (4) (5) (6)
Cutoff 1 Cutoff 2 Cutoff 3 Cutoff 4
0.0452 0.0647 0.105* 0.193***
(0.113) (0.0736) (0.0606) (0.0570)
Garrouste, C., Paccagnella, O., 2012. Data quality: three examples of consistency
across SHARE and SHARELIFE data. In: SHARELIFE Methodology. Mannheim
Research Institute for the Economics of Ageing (MEA), Mannheim.
Giuliano, P., Spilimbergo, A., 2014. Growing up in a recession. Review of Economic
Studies 81 (2), 787817.
Graham, C., 2012. The Pursuit of Happiness: An Economy of Well-being. Brookings
Institution Press, Washington, D.C.
Graham, L., Oswald, A.J., 2010. Hedonic capital, adaptation and resilience. Journal of
Economic Behavior and Organization 76 (2), 372384.
Greene, W.H., Hensher, D.A., 2010. Modelling Ordered Choices: A Primer. Cambridge
University Press, Cambridge.
Halmdienst, N., Winter-Ebmer, R., 2013. Long-run effects of childhood shocks on
health in late adulthood. NRN working papers 2013-01.
Havari, E., Mazzonna, F., 2011. Can we trust older peoples statements on their child-
hood circumstances? Evidence from SHARELIFE. SHARE Working Paper Series
05-2011.
Havari, E., Peracchi, F., 2011. Childhood circumstances and adult outcomes: evidence
from SHARELIFE. EIEF Working Paper 1115.
Holland, P.W., Wainer, H., 1993. Differential Item Functioning. Erlbaum, Hillsdale,
NJ.
Ichino, A., Winter-Ebmer, R., 2004. The long-run educational cost of World War Two.
Journal of Labor Economics 22, 5786.
Jrges, H., 2013. Collateral damage: the German food crisis, educational attain-
ment and labor market outcomes of German post-war cohorts. Journal of Health
Economics 32 (1), 286303.
Kahneman, D., 1999. Objective happiness. In: Kahneman, D., Diener, E., Schwartz, N.
(Eds.), Well-being: The Foundations of Hedonic Psychology. Russell Sage, New
York, pp. 325.
Kahneman, D., Krueger, A.B., Schkade, D.A., Schwarz, N., Stone, A.A., 2004. A survey
method for characterizing daily life experience: the day reconstruction method.
Science 306, 17761780.
Kapteyn, A., Smith, J.P., van Soest, A., 2007. Vignettes and self-reports of work dis-
ability in the United States and the Netherlands. The American Economic Review
97 (1), 461473.
Kapteyn, A., Smith, J.P., van Soest, A., 2011. Anchoring vignettes and response con-
sistency. RAND Working Papers 840.
Kesternich, I., Siinger, B., Smith, J.P., Winter, J.K., 2014. The effects of World War
II on economic and health outcomes across Europe. Review of Economics and
Statistics 96 (1), 103118.
Kesternich, I., Siinger, B., Smith, J.P., Winter, J.K., 2013. Individual behavior as a
pathway between early-life shocks and adult health. Evidence from hunger
episodes in post-war Germany. Rand Working Papers 1015.
King, G., Murray, C.J.L., Salomon, J.A., Tandon, A., 2004. Enhancing the validity
and cross-cultural comparability of measurement in survey research. APSR 98,
191207.
Kok, R., Avendano, M., Bago dUva, T., Mackenbach, J., 2012. Can reporting het-
erogeneity explain differences in depressive symptoms across Europe? Social
Indicators Research 105 (2), 191210.
Kristensen, N., Johansson, E., 2008. New evidence on cross-country differences in
job satisfaction using anchoring vignettes. Labour Economics 15 (1), 96117.
Layard, R., 2005. Happiness: Lessons from a New Science. Penguin, London.
Layard, R., 2006. Happiness and Public Policy: A Challenge to the Profession. The
Economic Journal 116 (510), C24C33.
Layard, R., Clark, A., Cornaglia, F., Powdthavee, N., 2013. What predicts a successful
life? A life-course model of well-being. IZA Discussion Paper 7682.
Lewbel, A., 2007. Estimation of average treatment effects with misclassication.
Econometrica 75 (2), 537551.
Lindeboom, M., Portrait, F., van den Berg, G.J., 2006. Economic conditions early in
life and individual mortality. The American Economic Review 96 (1), 290302.
Lindeboom, M., Portrait, F., van den Berg, G.J., 2010. Long-run effects on longevity of
a nutritional shock early in life: the Dutch potato famine of 18461847. Journal
of Health Economics 29 (5), 617629.
Maccini, S., Yang, D., 2009. Under the weather: health, schooling, and economic
consequences of early-life rainfall. The American Economic Review 99 (3),
10061026.
 M. Bertoni / Journal of Health Economics 40 (2015) 4053
Mazzonna, F., 2011. The long-lasting effects of family background: a European cross-
country comparison. Munich Centre for the Economics of Aging Working Paper
Nr. 245-2011.
Meng, X., Qian, N., 2009. The long term consequences of famine on survivors: evi-
dence from a unique natural experiment using Chinas great famine. NBER
Working Papers 14917.
Murray, C.J.L., Ozaltin, E., Tandon, A., Salomon, J.A., Sadana, R., Chatterji, S., 2003.
Empirical evaluation of the anchoring vignettes approach in health surveys.
In: Murray, C.J.L., Evans, D.B. (Eds.), Health Systems Performance Assessment:
Debates, Methods and Empiricism. World Health Organisation, Geneva, pp.
369399.
Neelsen, S., Stratmann, T., 2011. Effects of prenatal and early life malnutrition: evi-
dence from the Greek famine. Journal of Health Economics 30, 479488.
OECD, 2013. OECD Guidelines on Measuring Subjective Well-being. OECD Publish-
ing, Paris.
Peracchi, F., Rossetti, C., 2012. Heterogeneity in health responses and anchoring
vignettes. Empirical Economics 42 (2), 513538.
Peracchi, F., Rossetti, C., 2013. The heterogeneous thresholds ordered response
model: identication and inference. Journal of the Royal Statistical Society 176
(3), 703722.
Pinger, P., Schoch, J., van den Berg, G.J., 2011. Instrumental variable estimation of the
causal effect of hunger early in life on health later in life. IZA Discussion Paper
6110.
Portrait, F., Teeuwiszen, E., Deeg, D., 2011. Early life undernutrition and chronic
diseases at older ages: the effects of the Dutch famine on cardiovascular diseases
and diabetes. Social Science and Medicine 73, 711718.
Powdthavee, N., 2010. The Happiness Equation: The Surprising Economics of our
Most Valuable Asset. Icon Books, London.
53
Pudney, S., Shields, M., 2000. Gender, race, pay and promotion in the British nursing
profession: estimation of a generalized ordered probit model. Journal of Applied
Econometrics 15 (4), 367399.
Rice, N., Robone, S., Smith, P., 2011. Analysis of the validity of the vignette approach
to correct for heterogeneity in reporting health system responsiveness. Euro-
pean Journal of Health Economics 12, 141162.
Rice, N., Robone, S., Smith, P., 2012. Vignettes and health system responsiveness
in cross-country comparative analysis. Journal of the Royal Statistical Society:
Series A 175 (2), 337369.
Sen, A.K., 1985a. Commodities and Capabilities. North Holland, Amsterdam.
Sen, A.K., 1985b. The Standard of Living: The Tanner Lectures. Cambridge University
Press, Cambridge.
Smith, J.P., 2009. Reconstructing childhood health histories. Demography 46 (2),
387403.
van den Berg, G.J., Deeg, D., Lindeboom, M., Portrait, F., 2010. The role of early-
life conditions in the cognitive decline due to adverse events later in life. The
Economic Journal 120 (548), F411F428.
van den Berg, G.J., Deeg, D., Lindeboom, M., Scholte, R., 2014. Does the size
of the effect of adverse events at high ages on daily-life physical function-
ing depend on the economic conditions around birth? IZA Discussion Paper
8075.
van Soest, A., Delaney, L., Harmon, C., Kapteyn, A., Smith, J.P., 2011. Validating the
use of anchoring vignettes for the correction of response scale differences in
subjective questions. Journal of the Royal Statistical Society Series A 174 (3),
575595.
van Soest, A., Vonkova, H., 2014. Testing the specication of parametric models by
using anchoring vignettes. Journal of the Royal Statistical Society Series A 177
(1), 115133.
 Journal of Health Economics 40 (2015) 5468

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

Impacts of the Affordable Care Act dependent coverage provision on

health-related outcomes of young adults
Silvia Barbaresco a , Charles J. Courtemanche b, , Yanling Qi a
a
Georgia State University, United States
b
Georgia State University and National Bureau of Economic Research, United States

a r t i c l e i n f o a b s t r a c t

Article history: The rst major insurance expansion of the Affordable Care Act a provision requiring insurers to allow
Received 7 August 2014 dependents to remain on parents health insurance until turning 26 took effect in September 2010.
Received in revised form We estimate this mandates impacts on numerous outcomes related to health care access, preventive
17 December 2014
care utilization, risky behaviors, and self-assessed health. We estimate difference-in-differences models
Accepted 22 December 2014
with 2325 year olds as the treatment group and 2729 year olds as the control group. For the full sam-
Available online 30 December 2014
ple, the dependent coverage provision increased the probabilities of having health insurance, a primary
care doctor, and excellent self-assessed health, while reducing body mass index. However, the mandate
JEL classication:
I12
also increased risky drinking and did not lead to any signicant increases in preventive care utilization.
I13 Subsample analyses reveal particularly large gains for men and college graduates.
I18 2015 Elsevier B.V. All rights reserved.

Keywords:
Affordable Care Act
Health insurance
Dependent coverage
Risky behaviors
Preventive care

1. Introduction Many states already had some form of dependent coverage

The Patient Protection and Affordable Care Act (ACA) of March
2010 aimed to achieve nearly universal coverage in the United
States through a combination of mandates, subsidies, Medic-
aid expansions, and health insurance exchanges (Gruber, 2011).
Although the majority of the ACAs provisions just took effect in
2014, one important component of the law a dependent cover-
age provision was implemented on September 23rd, 2010. This
provision allows dependents to remain on a parents private health
insurance plan until the start of the rst plan year after they turn 26
years old. Previously, private insurers often dropped non-student
dependents at age 19 and student dependents at age 23 (Anderson
et al., 2012, 2014).
Corresponding author at: Department of Economics, Andrew Young School of
Policy Studies, Georgia State University, Atlanta, GA 30303, United States.
Tel.: +1 404 413 0082.
E-mail addresses: sbarbaresco1@student.gsu.edu (S. Barbaresco),
ccourtemanche@gsu.edu (C.J. Courtemanche), yqi3@gsu.edu (Y. Qi).
mandate before the ACA, but the state laws are typically weaker.
Most state laws have an age threshold below 26 or require addi-
tional criteria, such as being a full-time student, living with ones
parents, or not being married. Moreover, state laws do not apply
to self-funded benet programs, and more than half of private
sector workers with employer-provided health insurance are in
self-funded plans (Monheit et al., 2011). Perhaps because of these
limitations, Monheit et al. (2011) and Levine et al. (2011) nd that
state dependent coverage mandates only lead to small increases
in dependent coverage that are offset by a decline in young adults
holding their own policies. In contrast, the ACA provision applies
to all young adults under age 26 and all private plans. It there-
fore has the potential to dramatically affect young adults across the
country, including in states with a pre-existing dependent coverage
provision.
The ACA dependent coverage expansion provides a unique
opportunity to study the impacts of a health insurance intervention
specic to young adults, the age group with the highest uninsured
rate (Levine et al., 2011). Prior to the ACA, the uninsured rate was
29% among individuals ages 1824 and 27% among those 2534,
compared to 19% for 3544 year olds and 14% for 4564 year olds

http://dx.doi.org/10.1016/j.jhealeco.2014.12.004
0167-6296/ 2015 Elsevier B.V. All rights reserved.
 S. Barbaresco et al. / Journal of Health Economics 40 (2015) 5468
(DeNavas-Walt et al., 2010). Since any attempt to obtain univer-
sal coverage necessarily involves large coverage expansions among
young adults, it is important to understand the effects of insur-
ance on this group. It is unclear the extent to which results from
other contexts such as Medicaid, Medicare, or the Massachusetts
health care reform of 2006 are applicable. Young adults are gen-
erally healthier than the populations covered by these programs,
and therefore may experience smaller gains from health insurance.
Alternatively, young adults may be relatively poor and therefore
respond strongly to reduced out-of-pocket costs of medical care.1
Given the short amount of time since its implementation,
researchers are only beginning to study the impacts of the ACA
dependent coverage provision. Cantor et al. (2012) and Sommers
and Kronick (2012) show that the mandate increased health
insurance coverage for young adults across all racial groups and
regardless of employment status. Sommers et al. (2013) nd that
the provision increased insurance coverage among young adults,
while reducing delays in getting care and care foregone because of
cost. Akosa Antwi et al. (2013) again nd an increase in insurance
coverage, but they also present evidence of labor market conse-
quences such as young adults shifting from full-time to part-time
jobs. Akosa Antwi et al. (2014) show that the mandate increased
young adults utilization of inpatient care, particularly for mental
illness. Chua and Sommers (2014) do not nd any evidence that
the provision affected health care use, but they do nd a reduc-
tion in out-of-pocket medical expenses and increases in excellent
self-reported physical and mental health.
These papers all share a common general research design: com-
paring changes in outcomes among the treated age range 1925
to those of other young adults. The age range used for the control
group varies across these studies, with some including individuals
up to 34 years old (Sommers and Kronick, 2012; Sommers et al.,
2013; Chua and Sommers, 2014). Slusky (2013) questions the valid-
ity of this approach, arguing that different age groups are often
subject to different economic shocks. He runs placebo tests using
data from before the mandate and articial treatment dates, nd-
ing that the same specication estimates signicant effects more
often than could be attributed to chance. He suggests narrowing the
age bandwidths of the treatment and control groups as a possible
solution.
We contribute to this literature on the ACA dependent cover-
age provision in four ways. First, we consider a number of new
outcomes. Using data from the Behavioral Risk Factor Surveil-
lance System (BRFSS), we investigate 18 outcomes related to health
care access, utilization of preventive care, risky health behaviors,
and self-assessed health. The health care access measures include
having insurance, a primary care doctor, and any foregone care
because of cost. Our preventive care measures are dummies for
recent u vaccinations, well-patient checkups, and pap tests. The
health behavior outcomes reect smoking, drinking, body mass
index, exercise, and pregnancy. The self-assessed health variables
relate to overall, mental, and physical health as well as health-
related functional limitations. Of these outcomes, only insurance
coverage, foregone care because of cost, and self-assessed physi-
cal and mental health are studied in other papers in the literature.
To our knowledge we are the rst to investigate the ACA depend-
ent coverage provisions impact on preventive care or health
behaviors. Moreover, although Chua and Sommers (2014) examine
self-assessed physical and mental health, their measures and ours
1
Aside from age, the ACA dependent coverage mandate is also a unique coverage
expansion in that it represents an expansion of private rather than public insur-
ance, and that, since it only affects those whose parents have insurance, the treated
population may be of higher socioeconomic status than that of other interventions.
55
are meaningfully different. They use dummies for self-reporting
excellent physical and mental health, so their estimates only cap-
ture changes at the upper end of the health distribution. In contrast,
we utilize ve measures that should together capture changes at
various parts of the distribution. A dummy for excellent overall
health reects the high end, a dummy for very good or excellent
health reects a somewhat lower portion, and three more severe
outcomes number of days of the past 30 not in good physical
health, not in good mental health, and with health-related limita-
tions reect an even lower portion. This distinction will prove
critical to the results.
Our second contribution is to push further than prior studies
toward addressing the methodological concerns raised by Slusky
(2013), both by using narrow age ranges for the treatment and
control groups and by validating these selections through placebo
testing. Our treatment group consists of individuals ages 2325,
slightly below the dependent coverage provisions age cutoff, and
our control group consists of those slightly above the cutoff at ages
2729. We run placebo tests checking for effects of articial inter-
ventions in the pre-treatment period. Our classications perform
well in the placebo tests, whereas the wider age ranges commonly
used in the literature prove more problematic.
Another contribution is that we use over three full years of
post-treatment data (2011 through 2013, plus a few months after
implementation at the end of 2010). To our knowledge, none of the
prior papers in the ACA dependent coverage provision literature
have used more than one full year of post-treatment data, which
leaves the estimates susceptible to confounding from temporary
age-specic shocks and uctuations. If estimated effects persist
with three years of post-treatment data, we can be more condent
that they are not driven by transitory movements in unobserved
characteristics.
Finally, we contribute to the literature by testing for heteroge-
neous effects. Of the outcomes included in our paper, heterogeneity
in the effects of the ACA dependent coverage provision has only
previously been evaluated for insurance coverage (Akosa Antwi
et al., 2013; Sommers et al., 2013) and cost being a barrier to
care (Sommers et al., 2013). We will nd important heteroge-
neous effects on other outcomes as well, such as self-assessed
health. Moreover, although Akosa Antwi et al. (2013) and Sommers
et al. (2013) evaluate whether effects differ by certain demographic
characteristics, neither paper tests for heterogeneous effects by
socioeconomic status.2 We will nd that the effects of the depend-
ent coverage provision vary considerably by education level.
Our difference-in-differences results from the full sample sug-
gest that the ACA dependent coverage provision improved health
care access for young adults, had little effect on preventive care use,
had mixed effects on risky health behaviors, and improved self-
assessed health at the high end of the distribution. Specically, we
document improvements in four of the eighteen outcomes: health
insurance coverage, access to a primary care doctor, excellent self-
assessed health, and body mass index. However, we nd evidence
of an increase in risky drinking, and no clear effects in either direc-
tion on the remaining thirteen outcomes.
We evaluate heterogeneity in the effects of the mandate through
subsample analyses, nding the greatest improvements in out-
comes for men and college graduates. The increase in health
insurance coverage was greater for men than women, and only
2
Sommers et al. (2013) note that testing for heterogeneity by educational attain-
ment is difcult because many individuals in their treatment group 1925 year
olds are still in the process of completing their education. Another advantage
of using a narrow age range for the treatment group 2325 year olds is that
excluding the prime college ages largely ameliorates this concern.
56 S. Barbaresco et al. / Journal of Health Economics 40 (2015) 5468
men experienced statistically signicant gains in any outcomes
beyond health insurance: primary care access, exercise, and overall
self-assessed health. Stratifying by education reveals that the insur-
ance expansions were similar for college graduates and non-college
graduates. However, only college graduates experienced signicant
gains in any other outcomes besides insurance specically, pri-
mary care access, cost being a barrier to care, body mass index
(BMI), obesity, and overall self-assessed health. Young adults with
different education levels therefore appear to respond differently
to exogenously obtaining health insurance.
2. Health insurance and health-related outcomes
The most obvious theoretical implication of health insurance is
that by lowering the effective price of health care, health insur-
ance should increase its utilization. However, increased health
care utilization does not necessarily improve health. Diminish-
ing marginal returns suggest that health care can only improve
health up to a certain level (e.g. Grossman, 1972). Whether
the additional consumption of medical care induced by insur-
ance generates substantial gains in health therefore depends
on the initial level of health capital. Since the uninsured can
often obtain essential needs by paying directly or receiving char-
ity care, these individuals need not have low baseline levels
of health. Moreover, the marginal returns to health care differ
for different outcomes. Risky health behaviors such as smoking,
excessive drinking, and overeating might be particularly dif-
cult to improve through health care, as they require lifestyle
changes. Medical professionals ability to inuence health behav-
iors is generally limited to providing accountability, information,
strategies, and sometimes drugs to make behavioral changes eas-
ier.
Another relevant issue when evaluating the impact of health
insurance on health is that obtaining insurance could induce indi-
viduals to take more health risks, since the provision of health
insurance decreases the nancial losses associated with sick-
ness. This concept is known as ex ante moral hazard (Ehrlick
and Becker, 1972). Theoretically, ex ante moral hazard could both
increase risky behaviors and reduce investments in preventive
care.
Finally, exogenous provision of health insurance could lead to
income effects for individuals who used to purchase their own
insurance policy but now are able to receive free or subsidized
coverage, or for the newly-insured if their out-of-pocket medical
expenses drop. The available evidence from natural experiments
suggests that additional income increases health care utilization
(Acemoglu et al., 2013), either increases BMI or has no effect
(Lindahl, 2005; Schmeiser, 2009; Cawley et al., 2010), increases
smoking along the intensive but not extensive margin (Apouey
and Clark, 2015), and increases drinking (Apouey and Clark, 2015).
The income effect may therefore improve health via medical care
but worsen health via risky behaviors. Accordingly, evidence of
incomes causal effect on overall health is mixed, with Lindahl
(2005) and Frijters et al. (2005) nding that it improves self-
assessed health, Apouey and Clark (2015) nding that it improves
mental health but not overall health, and Snyder and Evans (2006)
showing that it raises mortality risk among seniors.
In sum, the effects of insurance on preventive health care
utilization, risky health behaviors, and overall health status are
theoretically ambiguous. Insurance may improve these outcomes
through direct price effects, worsen them through ex ante moral
hazard, or affect them in either direction through income effects.
The net effects could differ for different outcomes. For instance,
direct price effects might dominate for primary care utilization
but moral hazard might dominate for risky behaviors. Empirical
analysis is necessary to resolve this ambiguity.
Causally interpretable evidence generally conrms the predic-
tion that insurance increases health care utilization for U.S. adults.
Manning et al. (1987) analyzed the randomized RAND Health Insur-
ance Experiment, nding that lower copayments increased doctor
visits. Medicaid and Medicare expansions have been shown to
increase utilization of primary and hospital care (Currie and Gruber,
1996a; Dafny and Gruber, 2005; Finkelstein et al., 2012; Taubman
et al., 2014; Lichtenberg, 2002; Card et al., 2008). Other evidence
suggests that the Massachusetts universal coverage initiative of
2006 increased preventive services while reducing emergency
room utilization, avoidable hospitalizations, and medical needs
unmet because of cost (Miller, 2011, 2012; Kolstad and Kowalski,
2012; Van der Wees et al., 2013). More directly relevant to our
study population, Anderson et al. (2012, 2014) exploit the sharp
drops in coverage on parents insurance at ages 19 and 23 to show
that losing coverage reduced young adults emergency room and
hospital visits. Finally, as mentioned previously, Akosa Antwi et al.
(2014) show that the ACA dependent coverage provision increased
hospital admissions, although Chua and Sommers (2014) nd no
signicant effects on survey measures of hospital, primary care, or
prescription drug utilization.
The evidence of health insurances effect on health is mixed.
The RAND experiment only found that better insurance cover-
age improved health for certain subgroups (Brook et al., 1983).
Medicaid expansions increase self-reported overall, physical, and
mental health and reduce mortality, but have no statistically
detectable effects on laboratory-measured health outcomes (Currie
and Gruber, 1996b; Finkelstein et al., 2012; Sommers et al., 2012;
Baicker et al., 2013). Card et al. (2009) nd a reduction in the
mortality rate among recently hospitalized Medicare recipients,
but Finkelstein and McKnight (2008) nd no signicant effect of
Medicare on the mortality rate of seniors in general. Evidence
suggests that the Massachusetts reform improved self-assessed
overall, physical, and mental health, while decreasing functional
limitations, joint disorders, and mortality (Van der Wees et al.,
2013; Courtemanche and Zapata, 2014; Sommers et al., 2014). As
mentioned previously, Chua and Sommers (2014) nd that the ACA
dependent provision increased the probabilities of self-reporting
excellent physical and mental health.
Evidence on the causal effects of health insurance on risky health
behaviors is also mixed.
Brook et al. (1983) nd no evidence that insurance affected
smoking or body weight in the RAND experiment. Dave and
Kaestner (2009) report that Medicare decreased physical activity
while increasing smoking and drinking. Finkelstein et al. (2012)
do not nd any signicant impacts of Medicaid on smoking or
BMI. Courtemanche and Zapata (2014) nd that the Massachusetts
reform reduced body mass index and did not affect smoking or
physical activity.
In sum, there is little prior evidence on the effects of health
insurance on young adults access to care, preventive care uti-
lization, risky health behaviors, or health. Given the theoretical
ambiguities and variation in empirical ndings discussed above,
we cannot assume prior results from other contexts such as
Medicaid and Medicare generalize. For instance, young adults rel-
atively high baseline levels of health might lead them to have
relatively inelastic demand for health care or a low marginal effect
of health care on health. On the other hand, young adults demand
for health care could be relatively elastic given their generally
low income and wealth levels. Moreover, one might expect young
adults to be the most susceptible to ex ante moral hazard since this
is often the life stage in which opportunities to engage in particular
risky behaviors (e.g. binge drinking) are introduced.
 S. Barbaresco et al. / Journal of Health Economics 40 (2015) 5468
3. Data
Our main data source is the BRFSS, a telephone survey con-
ducted by state health departments in conjunction with the U.S.
Centers for Disease Control and Prevention to collect information
on health and health behaviors. The survey is conducted monthly
through a random digit dialing method that selects a representative
sample of respondents from the non-institutionalized population
of adults at least 18 years old. The BRFSS provides several advan-
tages for our analyses. First, it contains a wide range of appropriate
outcome variables. Second, it includes demographic characteristics
as well as state, month, and year identiers that allow us to con-
struct the treatment variable and jointly control for many different
factors. Next, it contains a much larger number of observations
than other datasets with the necessary variables. Finally, the BRFSS
includes a number of pre-treatment waves that allow for detailed
testing of differential trends in the outcomes between treatment
and control groups.
Our primary analysis sample consists of the 20072013 waves,
which include the year the ACA dependent coverage mandate took
effect plus three years on both sides. One reason we exclude the
years before 2007 is to limit our sample to years of relatively poor
economic performance. This reduces the possibility of confounding
from differential impacts of macroeconomic shocks on the health-
related outcomes of different age groups. However, robustness
checks and placebo tests will utilize data as far back as 2001. We
do not use any waves before 2001 because the BRFSS made major
changes to the survey in that year. Many of the questions used to
construct our outcome variables are either not available in earlier
years or differ in non-trivial ways.
Most of our analyses use ages 2325 as the treatment group
and ages 2729 as the control group. Following much of the prior
literature, 26 year olds are excluded because their treatment sta-
tus is ambiguous: they may still be covered by the ACA mandate
depending on their birthdate and the start date of their parents
insurance plan year (Akosa Antwi et al., 2013). Although the prior
literature uses 1925 as the treatment group, we prefer 2325
for two reasons.3 First, prior to the ACA, insurers most commonly
dropped non-student dependents from parents plans at age 19, but
most commonly dropped student dependents at age 23. Excluding
1922 year olds therefore results in a cleaner treatment group, i.e.
a higher proportion of the treatment group actually being affected
by the treatment. Accordingly, Akosa Antwi et al. (2014) show that
the ACA dependent coverage provisions impact on having insur-
ance was more than twice as large for 2325 year olds as for 1922
year olds. Second, Slusky (2013) shows that the models from prior
papers with ages 1925 as the treatment group lead to poor placebo
test results for insurance and labor market outcomes. He suggests
narrowing the age bandwidth as a potential solution. Indeed, we
will show that wider age ranges lead to problematic placebo test
results for our outcomes as well, and that our narrower age range
performs better.
We utilize eighteen different health-related dependent vari-
ables. The rst three relate to health care access: dummy variables
reecting whether the respondent has any health insurance,
has a primary care physician, and had any medical care needed
but not obtained because of cost in the previous year. Unfortu-
nately, the BRFSS does not include more detailed questions on
health insurance, such as the source of coverage. The next three
3 9
Studies in the literature utilize somewhat different control groups. Cantor et al.
(2012) use 2730 year olds; Sommers and Kronick (2012), Sommers et al. (2013),
and Chua and Sommers (2014) use 2634 year olds; Akosa Antwi et al. (2013) use
1618 and 2729 year olds; and Akosa Antwi et al. (2014) use 2729 year olds.
57
outcomes dummies for having a u vaccination (shot or spray),
a well-patient doctor check-up visit (e.g. physical), and a pap
test (for women) in the previous year reect preventive care
utilization.4 The next category of variables relates to risky health
behaviors: a dummy for whether the individual currently smokes,
number of alcoholic drinks in the past 30 days, a dummy for being
a risky drinker (more than 30 drinks total or at least one occasion
with four or more drinks for women, more than 60 drinks total
or at least one occasion with ve or more drinks for men),5 body
mass index (BMI = weight in kg/height in m2 ),6 a dummy for obese
(BMI 30), a dummy for whether an unmarried female respondent
is pregnant (the only proxy for risky sexual activity available in
the BRFSS), and a dummy for obtaining any recreational exercise
in the past 30 days.7 Finally, we include several variables related
to self-assessed health status: a dummy for whether overall health
is very good or excellent, a dummy for whether overall health
is excellent, and days of the last 30 not in good mental health,
not in good physical health, and with health-related functional
limitations. Although self-assessed health is subjective, research
has repeatedly found it to be correlated with objective measures of
health such as mortality (e.g. Idler and Benyamini, 1997; DeSalvo
et al., 2006; Phillips et al., 2010). Self-assessed health is also a
global measure of health that captures the full range of possible
diseases and limitations (Idler and Benyamini, 1997).8
We also utilize a wide array of control variables. These include
dummy variables for each year of age, gender, race/ethnicity,
marital status, education, household income category, number of
children in the household, whether the respondent reports her
primary occupation as student, and whether the respondent is
unemployed. Additionally, we control for monthly state unem-
ployment rate, obtained from the Bureau of Labor Statistics. As
mentioned previously, we are concerned about different impacts
of the recession on different age groups, so controlling for several
variables related to economic conditions at both the individual and
aggregate levels could potentially be important. We also control for
whether the respondents state had any dependent coverage man-
date covering her age * marital status * student status group in the
survey year based on information from the National Conference
of State Legislatures (2010).9 Additionally, in the u vaccination
regressions we control for interactions of the age xed effects with
4
Other preventive care variables typically studied in the literature, such as mam-
mograms and prostate exams, are not relevant for our study population of young
adults.
5
The dummy for risky drinker is created to come as close as the BRFSS data
will allow to the National Institute on Alcohol Abuse and Alcoholisms de-
nition of at-risk drinking: more than 7 drinks per week total or at least one
occasion with three or more drinks for women, and more than 14 drinks per
week total or at least one occasion with four or more drinks for men. See
http://pubs.niaaa.nih.gov/publications/womensfact/womensfact.htm.
6
Body mass index is based on self-reported height and weight, which are prone
to measurement error (Cawley, 2004). Researchers have repeatedly found that this
measurement error does not affect the signs and signicance of regression estimates
with BMI as a dependent variable, though it may slightly attenuate the magnitude
of the estimates (e.g. Lakdawalla and Philipson, 2002; Courtemanche et al., 2014,
forthcoming).
7
Unfortunately, the more detailed BRFSS questions on physical activity are only
available in odd numbered survey years and changed dramatically in 2011, so they
are not useful for our analyses.
8
Moreover, other commonly-used measures of health are not practical in our
context. Mortality rates are likely too low among young adults to estimate effects of
coverage expansions with meaningful precision, while measures of avoidable hos-
pitalizations confound insurances impact on health with the reduction in effective
prices.
Note that not everyone coded as a 1 for state mandate is actually treated by
such a mandate. Additional qualiers beyond age, student status, and marital status
exist in some states, while young adults whose parents employers self-insure are
also not covered by state mandates.
58 S. Barbaresco et al. / Journal of Health Economics 40 (2015) 5468

the number of positive inuenza tests in the country during the par- Table 1
Sample sizes for different outcomes.
ticular u season (a proxy for severity of the u season). Flu seasons
in the post-treatment years were much more severe than those in Outcome variable Total Treatment Control
the pre-treatment years, so adding these interactions prevents the (2325) (2729)
estimates from being confounded by differential responses to u Health care access
season severity by young adults of different ages.10 Any health insurance coverage 126,702 53,057 73,645
Finally, we include a dummy for whether the respondent is part Any primary care doctor 118,392 49,520 68,872
Cost prevented care in past year 107,831 45,041 62,790
of a cell phone only component of the sample, added in 2011
(this variable is 0 for all respondents before 2011). The fact that Preventive care utilization
Flu vaccination in past year 118,394 49,502 68,892
individuals who only used cell phones were not explicitly included
Well-patient checkup in past year 107,931 45,085 62,846
in the sample until 2011 raises the question of whether our sample Pap test in past year (women only)a 26,919 10,799 16,120
makeup meaningfully changed at about the same time the post-
Risky health behaviors
treatment period began. To address this issue, we not only control
Currently smokes cigarettes 125,616 52,607 73,009
for cell phone only users but also utilize the BRFSS sampling Alcoholic drinks in past 30 days 120,958 50,521 70,437
weights in all analyses. We found that these weights eliminate Risky drinker in past 30 days 120,037 50,110 69,927
any sharp changes in sample demographic characteristics in 2011. Body mass index 120,373 50,529 69,844
Additionally, this issue would only bias our regression estimates if Obese 120,373 50,529 69,844
Any exercise in past 30 days 122,720 51,337 71,383
the relationship between the outcomes of landline and cell phone Pregnancy (unmarried women only) 39,499 19,610 19,889
users is different among 2325 year olds than among 2729 year
Self-assessed health
olds, and in a way that is not captured by the controls. It is not
Overall health very good or excellent 126,662 53,102 73,560
obvious why this would be the case. Accordingly, we have veried Overall health excellent 126,662 53,102 73,560
(results available upon request) that dropping individuals who only Days of last 30 not in good mental health 124,773 52,386 72,387
use cell phones from our sample has very little effect on the coef- Days of last 30 not in good physical health 124,861 52,387 72,474
cient estimates, though it does generally increase the standard Days of last 30 with health-related 125,365 52,615 72,750
limitations
errors due to the reduced sample size.
a
After excluding observations with missing data for any of the The pap test variable is only available in even-numbered years, reducing the
sample size for that outcome.
control variables, Table 1 reports the sample sizes for the regres-
sions for each dependent variable, along with the numbers of
Table 2
individuals in the treatment and control groups. The sample sizes
Pre-treatment means and standard deviations for control variables.
differ slightly across dependent variables for two reasons. First,
each health-related variable is missing for a different number Control variable Treatment Control (Ages
of respondents. Second, the health-related variables have differ- (Ages 2325) 2729)

ent reection periods; some apply to the present (e.g. current Age dummies (age = 23 is omitted)
smoker), while others refer to a 30-day period (e.g. number of alco- Age = 24 0.349 (0.477)
Age = 25 0.322 (0.467)
holic drinks in the past 30 days) and others to a one-year period
Age = 27 0.310 (0.462)
(e.g. any well-patient doctor visit in the past year). We are con- Age = 28 0.343 (0.475)
cerned that short-run estimates would be misleading for variables Age = 29 0.347 (0.476)
with a long reection period.11 We therefore drop respondents sur- Female 0.505 (0.500) 0.508 (0.500)
Race/ethnicity dummies (non-Hispanic white is omitted)
veyed during this period of ambiguity; e.g. for well-patient doctor
Non-Hispanic black 0.112 (0.316) 0.116 (0.320)
visit in the past year we drop October 2010 through September Hispanic 0.224 (0.417) 0.209 (0.407)
2011, while for drinks in the past 30 days we drop only October Other than black, Hispanic, or white 0.087 (0.282) 0.077 (0.266)
2010.12 Currently married 0.305 (0.460) 0.564 (0.496)
Table 2 lists the control variables and compares the pre- Education dummies (less than high school degree is omitted)
High school degree but no further 0.283 (0.450) 0.257 (0.437)
treatment (January 2007 through September 2010) summary
Some college but no four-year degree 0.299 (0.458) 0.271 (0.444)
statistics of the treatment and control groups. Individuals in the College graduate 0.303 (0.459) 0.364 (0.481)
Household income dummies (less than $10,000 is omitted)
Between $10,000 and $15,000 0.068 (0.252) 0.049 (0.216)
Between $15,000 and $20,000 0.102 (0.303) 0.077 (0.267)
10
Specically, for the pre-treatment years 2007, 2008, and 2009, there were Between $20,000 and $25,000 0.116 (0.321) 0.097 (0.296)
23,753, 39,827, and 27,682 positive inuenza test results in the corresponding u Between $25,000 and $35,000 0.144 (0.351) 0.129 (0.335)
seasons 20062007, 20072008, and 20082009. For the post- or during-treatment Between $35,000 and $50,000 0.166 (0.372) 0.165 (0.371)
years 2010, 2011, 2012, and 2013, there were 157,449, 55,403, 27,012, and 75,342 Between $50,000 and $75,000 0.143 (0.350) 0.187 (0.390)
number of inuenza test results in the corresponding u seasons 20092010, $75,000 and over 0.186 (0.389) 0.240 (0.427)
20102011, 20112012, and 20122013 (CDC, 2014). The large 20092010 u sea- Number of children in household dummies (0 is omitted)
son number largely reects the swine u pandemic, but two of the three subsequent One child 0.230 (0.421) 0.235 (0.424)
seasons were still relatively strong. Our results suggest that younger young adults Two children 0.159 (0.366) 0.233 (0.423)
respond more strongly to u season severity than older young adults; therefore, Three children 0.055 (0.229) 0.110 (0.313)
omitting these interactions would lead to biased estimates. Four children 0.018 (0.133) 0.038 (0.192)
11
For example, suppose a respondent is surveyed in November 2010, the second Five or more children 0.008 (0.090) 0.016 (0.124)
month of the post-implementation period. The respondent would be classied as Cell phone only 0.703 (0.457)a 0.678 (0.467)a
post-treatment, but her answer about well-patient doctor visits in the past year Student 0.109 (0.312) 0.054 (0.226)
would reect only two months of the post-treatment period and ten months of the Unemployed 0.111 (0.314) 0.093 (0.290)
pre-treatment period. State unemployment rate 7.032 (2.615) 7.186 (2.666)
12
For u vaccinations in the past year, we only drop October 2010 through Pre-ACA state mandate 0.220 (0.415) 0.033 (0.179)
December 2010, as opposed to dropping a full year. We feel a shorter reection
Notes: BRFSS sampling weights are used. Means are reported, with standard devia-
period is appropriate in this case because u vaccinations are typically administered
tions in parentheses.
in the fall. For instance, if someone surveyed in March 2011 reports being vaccinated a
Indicates the summary statistics are from 2011 to 2013, since the variable is 0
in the past year, that vaccine almost certainly occurred during the post-treatment
for all respondents in all prior years.
period (October 2010 or later).
 S. Barbaresco et al. / Journal of Health Economics 40 (2015) 5468
treatment group are less likely to be married, have a college degree,
earn a high income, and have children in the household, and they
are more likely to be students or employed.
Table 3 reports the pre- and post-treatment sample means of
the outcome variables for the treatment and control groups, and
calculates the simple difference-in-difference of means. Prior to
the ACA dependent coverage provision, the uninsured rate was
higher for young adults in the treatment group than those in the
control group. The treatment group had lower rates of health care
utilization and health care access than the control group; higher
drinking and unmarried pregnancy rates but healthier levels of
risky drinking, BMI, obesity, and exercise; and broadly similar lev-
els of smoking and self-assessed health. Comparing changes in
the post- and pre-treatment means for the treatment and control
groups, the difference-in-differences are positive and signicant
for any insurance, primary care doctor, excellent health, and risky
drinker; negative and signicant for body mass index and obesity;
and insignicant for the other outcomes including all those in the
preventive care category.
Simple difference-in-differences estimates account for xed dif-
ferences in unobservable characteristics between the treatment
and control group, but are still susceptible to bias from time-
varying observables and unobservables. Figs. 13 show that at a
rst glance the pre-ACA trends for the treatment and control groups
appear generally similar for most outcomes, providing preliminary
evidence that changes over time in observables and unobserv-
ables may not be substantially different for 2325 year olds and
2729 year olds. We next turn to regression analyses that adjust
for changes in observables. Later, we will also conduct more formal
tests of the assumption of common trends in unobservables.
4. Average effects of the ACA dependent coverage mandate
4.1. Baseline model
We estimate the effects of the ACA dependent coverage provi-
sion on the eighteen health-related outcomes using reduced-form
difference-in-differences regressions. While it is tempting to esti-
mate instrumental variables models using the mandate as an
instrument for having insurance coverage, we are not condent
that the exclusion restriction would hold in such models because
there are several other mechanisms through which the mandate
could affect health-related outcomes besides the extensive margin
of health insurance coverage. Other possible mechanisms include
the intensive margin of coverage (switching from high deductible
catastrophic coverage to more comprehensive coverage), income
effects, and peer effects.
Our baseline regression is of the form
Yigst = 0 + 1 (Treatg Postt ) + 2 X igst + g + t + s + igst (1)
where Yigst is the health-related outcome for individual i of age g
living in state s in time t, expressed in a month/year combination.13
Treatg is a dummy variable for whether age g is in the treated age
range 2325 as opposed to the control age range 2729. Postt indi-
cates whether period t is after the implementation of the provision
(October 2010 or later). 1 is the difference-in-differences coef-
cient and it captures the difference between the effects of the
mandate on the treatment and control groups. X igst is a vector
13
Even though most of our outcomes are binary or non-negative count, we esti-
mate linear models because they typically give reliable estimates of average effects
(Angrist and Pischke, 2008). In unreported regressions (available upon request), we
verify that the average treatment effects are very similar using probit regressions
for the binary outcomes and negative binomial regressions for the count outcomes.
59
of the aforementioned control variables for sex, race, marital sta-
tus, education, income, children, cell phone survey, student status,
individual and state unemployment, and state dependent cover-
age mandate. We also include xed effects for each year of age,
month/year of time (e.g. January of 2007), and state, denoted by
g , t , and  s , respectively. igst is the error term.14 We do not
separately include Treatg and Postt in the model because Treatg is
perfectly collinear with the age xed effects while Postt is perfectly
collinear with the month/year xed effects.
We report heteroskedasticity-robust standard errors clustered
at the level of treatment: age. Following convention when there
are a small number of clusters (six in our case), for hypothesis
testing we use a t-distribution with degrees of freedom equal to
the number of clusters minus one. The critical values used in our
hypothesis tests are therefore considerably more stringent than
those using the standard normal distribution. It is possible that
even using stringent critical values might not be sufcient to elim-
inate the tendency to over-reject when the number of clusters is
small (Cameron et al., 2008). However, the placebo tests in the next
section will reject the null hypothesis even fewer than the expected
number of times, suggesting that our hypothesis tests are suf-
ciently conservative. One of our robustness checks will also address
this issue.
The key identifying assumption in a difference-in-differences
model is common counterfactual trends between the treatment
and control groups; i.e. in the absence of the intervention the
treatment and control groups would have experienced the same
changes in outcomes. Slusky (2013) argues that this assumption is
problematic when studying the impact of the ACA dependent cov-
erage provision on labor market-related outcomes (e.g. sources of
health insurance coverage, employment status, and work hours)
since cyclical uctuations in the economy have different effects
on different age groups. Since economic uctuations are related
to health,15 Sluskys concern could also apply to health-related
outcomes. As discussed previously, this is one of our main reasons
for using narrow age bandwidths of 2325 and 2729.
4.2. Robustness checks
We also estimate several variations of (1) as robustness checks.
First, we run regressions including only the demographic controls
(the sex, age, race, children, and marital status dummies) and xed
effects, excluding the economic controls since they may be endoge-
nous to the dependent coverage provision. Obtaining access to
parents insurance could potentially inuence a young adults deci-
sions about employment and education, which would then affect
income. Including covariates related to employment, education,
and income might therefore control away part of the causal effect
of the policy.
Our next several robustness checks vary the time period
included in the sample. In order to verify that the results are
not driven by our chosen length of the pre-treatment period, we
consider two alternatives: starting the sample in 2004 and 2001.
Additionally, we run regressions dropping March 2010 through
December 2010, as these months are somewhat ambiguous with
respect to their treatment status. We drop MarchSeptember
because the ACA was passed in March, so some insurance
14
In unreported regressions (available upon request) we have veried the results
remain virtually identical if we replace the state xed effects with xed effects for
each state-by-year combination.
15
Research generally shows that recessions are associated with improvements in
health and health behaviors (e.g. Ruhm, 2000, 2002, 2005), although recent evidence
suggests that the countercyclical nature of health observed in prior recessions may
not have been present during our sample period (Ruhm, 2013; Tekin et al., 2013).
60 S. Barbaresco et al. / Journal of Health Economics 40 (2015) 5468

Table 3
Means and standard deviations for outcome variables.

Outcome variable Pre-treatment period Post-treatment period Difference-in-differences

Treatment Control (Ages Treatment Control (Ages

(Ages 2325) 2729) (Ages 2325) 2729)

Health care access

Any health insurance coverage 0.680 (0.466) 0.753 (0.431) 0.709 (0.454) 0.708 (0.455) 0.073 (0.018)***
Any primary care doctor 0.564 (0.496) 0.641 (0.480) 0.519 (0.500) 0.558 (0.497) 0.038 (0.010)**
Cost prevented care in past year 0.241 (0.427) 0.216 (0.411) 0.240 (0.427) 0.235 (0.424) 0.020 (0.014)

Preventive care utilization

Flu vaccination in past year 0.225 (0.418) 0.246 (0.431) 0.239 (0.426) 0.265 (0.441) 0.006 (0.009)
Well-patient checkup in past year 0.521 (0.500) 0.545 (0.498) 0.524 (0.499) 0.529 (0.499) 0.019 (0.011)
Pap test in past year 0.693 (0.461) 0.724 (0.447) 0.614 (0.487) 0.647 (0.478) 0.002 (0.013)

Risky health behaviors

Currently smokes cigarettes 0.260 (0.432) 0.249 (0.432) 0.257 (0.437) 0.254 (0.435) 0.009 (0.012)
Alcoholic drinks in past 30 days 17.359 (43.926) 13.883 (34.703) 19.481 (43.947) 16.841 (40.916) 0.836 (0.889)
Risky drinker in past 30 days 0.775 (0.418) 0.807 (0.394) 0.749 (0.434) 0.769 (0.422) 0.013 (0.005)*
Body mass index 26.404 (5.807) 27.253 (6.031) 26.167 (6.019) 27.192 (6.142) 0.177 (0.050)**
Obese 0.222 (0.415) 0.262 (0.440) 0.197 (0.398) 0.252 (0.434) 0.014 (0.003)***
Any exercise in past 30 days 0.810 (0.392) 0.799 (0.401) 0.819 (0.385) 0.799 (0.401) 0.009 (0.005)
Pregnancy 0.048 (0.215) 0.043 (0.203) 0.044 (0.205) 0.040 (0.195) 0.001 (0.004)

Self-assessed health
Overall health very good/excellent 0.607 (0.488) 0.610 (0.488) 0.608 (0.488) 0.589 (0.492) 0.022 (0.011)
Overall health excellent 0.255 (0.436) 0.257 (0.437) 0.250 (0.433) 0.236 (0.425) 0.017 (0.003)***
Days not in good mental health 4.050 (7.638) 3.844 (7.680) 4.410 (8.067) 4.165 (8.063) 0.040 (0.162)
Days not in good physical health 2.240 (5.526) 2.303 (5.815) 2.446 (5.999) 2.484 (6.170) 0.025 (0.053)
Days with health-related limitations 1.589 (4.757) 1.664 (5.177) 1.727 (5.131) 1.739 (5.332) 0.063 (0.104)

Notes: Standard errors, heteroskedasticity-robust and clustered by age, are in parentheses. BRFSS sampling weights are used. Means are reported, with standard deviations
in parentheses.
*** Indicates the difference-in-difference is signicant at the 1% level; ** 5% level; * 10% level.

Health Insurance Coverage Personal Doctor Visits Cost Barrier to Care

.7

.3
.8

.25
.65
.75

.2
.6
.7

.15
.55

.1
.65

.5

2001 2003 2005 2007 2009 2011 2013 2001 2003 2005 2007 2009 2011 2013 2001 2003 2005 2007 2009 2011 2013

Age23~25 Age27~29 Age23~25 Age27~29 Age23~25 Age27~29

Fit line 95% CIs Fit line 95% CIs Fit line 95% CIs

Flu Shots Patient Checkups Pap Tests

.3

.8
.56

.75
.25

.54

.7
.2

.52

.65
.15

.5

.6

2001 2003 2005 2007 2009 2011 2013 2001 2003 2005 2007 2009 2011 2013 2001 2003 2005 2007 2009 2011 2013

Age23~25 Age27~29 Age23~25 Age27~29 Age23~25 Age27~29

Fit line 95% CIs Fit line 95% CIs Fit line 95% CIs

Fig. 1. Trends in access to care and preventive care variables by age group.
 S. Barbaresco et al. / Journal of Health Economics 40 (2015) 5468 61

Smokers Alcoholic Drinks per Month Risky Drinkers

.22 .24 .26 .28 .3 .32

.85
25
20

.8
.75
15
10

.7
2001 2003 2005 2007 2009 2011 2013 2001 2003 2005 2007 2009 2011 2013 2001 2003 2005 2007 2009 2011 2013

23~25 27~29 23~25 27~29 23~25 27~29

Fit line 95% CIs Fit line 95% CIs Fit line 95% CIs

Body Mass Index Obesity Any Exercise

.78 .8 .82 .84 .86

28

.3
.25
27
26

.2
.15
25

2001 2003 2005 2007 2009 2011 2013 2001 2003 2005 2007 2009 2011 2013 2001 2003 2005 2007 2009 2011 2013

23~25 27~29 23~25 27~29 23~25 27~29

Fit line 95% CIs Fit line 95% CIs Fit line 95% CIs

Pregnancy
.02 .03 .04 .05 .06 .07

2001 2003 2005 2007 2009 2011 2013

23~25 27~29
Fit line 95% CIs

Fig. 2. Trends in health behavior variables by age group.

plans may have complied preemptively prior to the dependent
coverage provisions ofcial implementation in September. We
drop OctoberDecember because, even though the mandate was
implemented in September, insurers did not have to comply until
the start of the next plan year, which is often January.16
Our nal robustness check addresses the potential concern that
standard errors may be understated because of autocorrelation
given the small number of clusters. We collapse the data into one
observation for each year of age in the pre-treatment period and
one observation for each year of age in the post-treatment period,
for a total of twelve observations. We then estimate

16
Y gt = 0 + 1 (Treatg Postt ) + 2 Treatg + 3 Postt + 4 X gt + gt (2)
Akosa Antwi et al. (2013) include two treatment variables to separately model
the effects of the mandate during the implementation period and after full imple-
mentation. We have considered this specication in unreported regressions and the where the lines above variables indicate averages across all individ-
estimated post-implementation effects remain very similar. uals of age g in time period (pre- or post-treatment) t, weighted by
62 S. Barbaresco et al. / Journal of Health Economics 40 (2015) 5468

Overall Health - Very Good & Excellent Overall Health - Excellent Days Not in Good Mental Health
.58 .6 .62 .64 .66 .68

.3

5
.28

4.5
.26

4
.24

3.5
.22

3
2001 2003 2005 2007 2009 2011 2013 2001 2003 2005 2007 2009 2011 2013 2001 2003 2005 2007 2009 2011 2013

Age23~25 Age27~29 Age23~25 Age27~29 Age23~25 Age27~29

Fit line 95% CIs Fit line 95% CIs Fit line 95% CIs

Days Not in Good Physical Health Days with Functional Limitations

1.8 2 2.2 2.4 2.6 2.8

2
1 1.2 1.4 1.6 1.8

2001 2003 2005 2007 2009 2011 2013 2001 2003 2005 2007 2009 2011 2013

Age23~25 Age27~29 Age23~25 Age27~29

Fit line 95% CIs Fit line 95% CIs

Fig. 3. Trends in self-assessed health variables by age group.

the individual BRFSS sampling weights. Since the small sample size the probability of having a primary care doctor by 2.03.4 per-
prevents all the control variables from being separately included, centage points and decreased the probability of having any care
X is a single variable that summarizes the inuence of all the con- needed but foregone because of cost by 1.62.3 percentage points.
trols. X is computed by regressing outcome Y on the controls using The effect on primary care doctor access is statistically signicant
the individual-level pre-treatment data, then predicting Y for the in all specications, but the effect on care foregone because of cost
whole sample based on the coefcient estimates, then aggregating is never signicant.
in the same manner described above. Despite this improved access, we do not nd any evidence of
increased preventive care utilization. We estimate a total of eigh-
teen models across the three preventive care measures, and none
4.3. Results
of these models reveal a statistically signicant positive effect of
the dependent coverage provision. The estimated effects on u
Table 4 presents the results for the baseline model and robust-
vaccinations and pap tests are negative in most specications and
ness checks. In addition to reporting estimated treatment effects
occasionally statistically signicant. The estimates for well-patient
and standard errors, for the baseline regressions we also report (in
checkup are all positive but never signicant.
brackets) the treatment effects expressed in standard deviations
We nd mixed evidence regarding the dependent coverage
of the dependent variables to provide some comparability of effect
provisions impacts on risky health behaviors. No signicant esti-
sizes across the different outcomes. Full regression output, includ-
mates are observed for smoking, pregnancy, or alcoholic drinks per
ing results for the controls, for selected outcomes is available in
month. However, the mandate statistically signicantly increased
Appendix Table A1.
the probability of risky drinking (excessive drinks per month or
The results suggest sizeable improvements in health care access
any binge drinking) in all specications, with magnitudes ranging
along at least some dimensions. We estimate that the ACA depend-
from 0.8 to 1.4 percentage points. The dependent coverage expan-
ent coverage provision statistically signicantly increased the
sion therefore appears to affect drinking at only the high end of
insurance coverage rate of 2325 year olds by between 5.5 and
the distribution, which is consistent with an ex ante moral hazard
6.7 percentage points, depending on the model. This is somewhat
explanation since mild to moderate drinking generally does not
larger than the around 35 percentage point increase estimated by
increase the need for medical services. In contrast, the dependent
previous studies that use the broader treated age range of 1925
coverage provision appears to improve weight-related behaviors.
(Cantor et al., 2012; Sommers and Kronick, 2012; Akosa Antwi et al.,
2013; Sommers et al., 2013).17 Additionally, the mandate increased

17
This discrepancy is consistent with Akosa Antwi et al.s (2014) nding that the
mandates impact on the probability of having any coverage was around twice as
large for 2325 year olds than 1922 year olds (4 compared to 2 percentage points).
Alternatively, estimates using the treated age range 1925 could be biased down-
ward given the problems documented in our placebo tests and those of Slusky
(2013).
 S. Barbaresco et al. / Journal of Health Economics 40 (2015) 5468 63

Table 4
Difference-in-difference regression estimates of effects of ACA dependent coverage mandate.

Outcome variable Baseline model Demographic Start in 2004 Start in 2001 Drop 3/1012/10 Collapsed data
controls only

Health care access

Any health insurance 0.061 (0.017)** [0.130] 0.067 (0.018)** 0.059 (0.013)*** 0.055 (0.012)*** 0.064 (0.016)*** 0.061 (0.015)***
Any primary doctor 0.032 (0.010)** [0.065] 0.034 (0.010)** 0.020 (0.006)** 0.021 (0.006)** 0.033 (0.009)** 0.029 (0.011)**
Cost prevented care 0.019 (0.014) [0.044] 0.019 (0.014) 0.022 (0.015) 0.023 (0.015) 0.020 (0.015) 0.016 (0.011)

Preventive care utilization

Flu vaccination 0.014 (0.007) [0.033] 0.011 (0.008) 0.017 (0.009) 0.018 (0.008)* 0.014 (0.008) 0.020 (0.006)**
Well-patient checkup 0.013 (0.011) [0.026] 0.015 (0.010) 0.011 (0.010) 0.011 (0.010) 0.017 (0.010) 0.011 (0.006)
Pap test 0.004 (0.015) [0.009] 0.003 (0.014) 0.019 (0.010) 0.025 (0.015)** 0.015 (0.015) 0.002 (0.008)

Risky health behaviors

Currently smokes 0.003 (0.007) [0.007] 0.006 (0.010) 0.001 (0.004) 0.008 (0.005) 0.001 (0.006) 0.005 (0.007)
Drinks per month 0.120 (0.906) [0.003] 0.468 (0.887) 0.429 (0.604) 0.597 (0.590) 0.083 (0.840) 0.011 (0.929)
Risky drinker 0.011 (0.003)** [0.026] 0.008 (0.004)* 0.009 (0.003)** 0.009 (0.007)** 0.014 (0.003)*** 0.009 (0.003)**
Body mass index 0.098 (0.029)** [0.017] 0.175 (0.045)** 0.124 (0.062) 0.169 (0.061)** 0.173 (0.074)* 0.118 (0.033)** *
Obese 0.009 (0.008) [0.022] 0.014 (0.005)** 0.010 (0.007) 0.011 (0.008) 0.013 (0.006)* 0.010 (0.004)**
Any exercise 0.003 (0.004) [0.008] 0.008 (0.007) 0.005 (0.003) 0.004 (0.004) 0.001 (0.005) 0.007 (0.003)**
Pregnancy 0.003 (0.005) [0.014] 0.002 (0.005) 0.004 (0.004) 0.002 (0.004) 0.003 (0.005) 0.002 (0.004)

Self-assessed health
Very good/exc. Health 0.015 (0.011) [0.031] 0.018 (0.010) 0.016 (0.011) 0.015 (0.008) 0.011 (0.010) 0.014 (0.009)
Excellent health 0.014 (0.005)** [0.032] 0.014 (0.003)*** 0.013 (0.004)** 0.014 (0.005)** 0.014 (0.006)* 0.015 (0.004)***
Days not good mental 0.081 (0.158) [0.010] 0.064 (0.144) 0.050 (0.144) 0.036 (0.116) 0.156 (0.156) 0.084 (0.127)
Days not good phys. 0.059 (0.068) [0.011] 0.045 (0.046) 0.022 (0.079) 0.014 (0.076) 0.075 (0.028)** 0.028 (0.063)
Days health limitations 0.122 (0.099) [0.025] 0.102 (0.093) 0.073 (0.109) 0.065 (0.094) 0.201 (0.107) 0.101 (0.086)

Notes: *** indicates signicant at the 1% level; ** 5% level; * 10% level. Standard errors, heteroskedasticity-robust and clustered by age, are in parentheses. All regressions
include the controls plus age, state, and time xed effects. BRFSS sampling weights are used. For the baseline regression, effect sizes in standard deviations of the dependent
variable (for the treatment group in the pre-treatment period) are in brackets.

The mandate reduces BMI in all six specications, with magnitudes
ranging from 0.098 to 0.175. All but one of the six estimates for
BMI are signicant, with the remaining one being nearly signi-
cant. The effect on obesity is also negative in all six models, though
it is only signicant in three. The effect on probability of having any
exercise is positive in all specications but only signicant in one.
It is possible that our inability to measure exercise in greater detail
e.g. calories burned per day from physical activity prevents the
emergence of further signicant results. It is also possible that the
reduction in BMI is coming via reduced caloric intake, which we
are unable to measure in the BRFSS.
It is theoretically conceivable that insurance coverage could
increase risky drinking but reduce weight. Health care access
may be more helpful for losing weight than reducing drinking.
Gains in information and accountability may both be greater for
weight control than drinking: dieting strategies can be complicated
and benet greatly from professional advice, and accountability is
greater for weight since patients are weighed at each visit. Addi-
tionally, the ex ante moral hazard effect could be stronger for risky
drinking than weight-related behaviors. Binge drinking has a non-
trivial chance of resulting in immediate medical needs, either from
alcohol poisoning, drunk driving accidents, or other injuries.18 In
contrast, expenditures to treat diseases associated with obesity
typically occur years down the road. Perhaps uninsured young
adults assume that they will be insured by time these downside
risks are realized, in which case ex ante moral hazard would not
apply. In short, the direct price effect could dominate for BMI, while
the ex ante moral hazard effect could dominate for drinking. Income
effects may play a role as well, especially for alcohol consumption
given the aforementioned evidence of a positive causal effect of
income on drinking (Apouey and Clark, 2015).
Turning to the self-assessed health outcomes, the mandate
increased the probability of young adults reporting excellent over-
all health by 1.31.5 percentage points and very good/excellent
health by 1.11.8 percentage points. However, only the estimates
for excellent health are signicant, as the standard errors for very
good/excellent health are larger. We do not nd any evidence of
effects on the variables representing more severe health problems:
days not in good mental health, not in good physical health, and
with health-related functional limitations. The lack of effects on our
mental and physical health outcomes is particularly interesting in
light of Chua and Sommers (2014) nding that the ACA dependent
coverage provision increased the probabilities of reporting excel-
lent mental and physical health. Chua and Sommers mental and
physical health variables emphasize changes at the high end of
the health distribution and may therefore correspond more closely
to our variable for excellent overall health than our physical and
mental health variables, which focus on not good health. In other
words, both our results and those of Chua and Sommers are con-
sistent with the provisions effects on mental and physical health
being concentrated in the high end of the health distribution.
Finally, we provide a brief discussion of the relative magnitudes
of the effects on different outcomes by comparing the treatment
effects expressed in standard deviations of the dependent vari-
ables. Not surprisingly, the largest effect of 0.13 standard deviations
is on the probability of having any health insurance coverage.
The next largest statistically signicant effect is on primary care
doctor access (0.065 standard deviations), then excellent health
(0.032 standard deviations), then risky drinker (0.026 standard
deviations), then nally BMI (0.017 standard deviations). The
largest statistically insignicant effects are on u vaccinations
(0.033 standard deviations) and very good/excellent health (0.031
standard deviations).

18
5. Placebo tests
In the US, approximately 80,000 cases of alcohol poisoning and 10,322 alcohol-
impaired driving crashes occur annually, with these incidents disproportionately
involving young adults (CDC, 2012; NHTSA, 2014). 599,000 alcohol-related injuries We next provide a series of placebo tests to evaluate whether the
occur annually among 1824 year old college students (NIAAA, 2013). previous results can credibly be interpreted as causal effects of the
64 S. Barbaresco et al. / Journal of Health Economics 40 (2015) 5468
ACA dependent coverage provision. Following Slusky (2013), we
estimate variants of equation (1) that test for effects of articially-
timed treatments during pre-treatment years. We estimate
models for three different seven-year windows of pre-treatment
data (to match the seven years used in our main 20072013 analy-
ses): 20032009, 20022008, and 20012007. Since the rst month
after the implementation of the actual dependent coverage man-
date was the 46th month (October 2010) of our 20072013 sample,
in each placebo test sample we date the implementation of the
articial intervention to the 46th month (e.g. October 2006 for
the 20032009 sample). We estimate (1) for each of the eighteen
dependent variables in each of the three placebo test samples.
Table 5 reports the coefcient estimates of interest from these
placebo tests. We run three tests for each of the eighteen depend-
ent variables, though a test is not possible for checkups using
20012007 data since the checkup question was not asked until
2005. This leaves a total of 53 regressions. Given the large num-
ber of estimates, we would expect some signicant results even for
valid models. Specically, approximately 01 estimates should be
signicant at the 1% level, about 23 at the 5% level, and about 5 at
the 10% level. We obtain numbers even smaller than these. No esti-
mated treatment effects are signicant at the 1% level, 2 (3.8%)
are signicant at the 5% level, and 3 (5.7%) are signicant at the
10% level. Moreover, we do not obtain more than one placebo test
rejection for any outcome. In other words, it is not clear that there
are any outcomes for which our baseline difference-in-differences
model is inappropriate.
In the interest of contributing to the broader debate in the liter-
ature about the appropriateness of different age bandwidths when
using difference-in-differences models to estimate the effects of
the ACA dependent coverage provision, we also run the same set
of placebo tests for the most common age ranges used in the liter-
ature: treatment group 1925 and control group 2634 (Sommers
and Kronick, 2012; Sommers et al., 2013; Chua and Sommers,
2014). We obtain 4 placebo test rejections (7.5%) at the 1% level,
7 (13.2%) at the 5% level, and 11 (20.8%) at the 10% level. The full
table of results is available upon request.
6. Heterogeneity
Having established our baseline results and assessed the
validity of our model, we next turn to an examination of
heterogeneity in the treatment effects. We considered strati-
cations by sex, race/ethnicity, education, and state pre-ACA
dependent coverage law status, but we did not observe any sta-
tistically signicant differences in effects across the subgroups
for race/ethnicity and pre-ACA law, so we only report the
results for the stratications by sex and education. For education,
we stratify into two groups: college graduates and non-college
graduates.19
Theoretically, the ACA dependent coverage provision could have
heterogeneous effects on health-related outcomes for three rea-
sons. First, there could be heterogeneous effects on the probability
of having insurance coverage. In the pre-treatment portion of sam-
ple, females were more likely to have insurance than males (76%
versus 67%), and college graduates were much more likely to
have insurance than non-college graduates (88% versus 64%). One
might therefore expect larger gains in coverage among males and
non-college graduates. On the other hand, young adults of high
19
Further stratication by education led to estimates that were too imprecise to be
useful. Note that we do not include a separate category for current students because
our sample only includes those 23 and older, so the proportion of our respondents
reporting student as their primary occupation is low.
socioeconomic status may be more likely to have parents with
employer-provided coverage, so the gains in coverage could poten-
tially be larger for college graduates.
A second possible source of heterogeneity is that, even if the
gains in health insurance are the same among all groups, differ-
ent groups could respond differently to receiving coverage. For
instance, Grossman (1972) argues that education enables individ-
uals to become more efcient producers of health. More education
may therefore better equip individuals to make the most out of
the newly-acquired insurance (e.g. more easily nd providers who
accept the insurance, ask better questions at doctors appoint-
ments, or better follow medical advice). Alternatively, the price
elasticity of medical care could be strongest among low-income
individuals, in which case the effects of obtaining insurance on
health care utilization and health could be largest for non-college
graduates. The price elasticity of medical care could also differ by
sex. For instance, evidence suggests that females are more risk
averse than males (e.g. Jianakoplos and Bernasek, 1998). One might
therefore expect females to be more likely to obtain medical care
regardless of its price, whereas males might only utilize care if
the cost is minimal; i.e. males might have stronger price elastic-
ities. Indeed, in our pre-treatment data uninsured females had
higher rates of primary care doctor access, u vaccination, and
well-patient checkups than uninsured males.
Third, as discussed at the beginning of Section 4, the dependent
coverage provision could affect health-related outcomes through
mechanisms besides the extensive margin of insurance coverage
particularly the intensive margin of coverage and there could be
heterogeneous effects along these dimensions. For instance, sup-
pose part of the reason females and college graduates had lower
pre-ACA uninsured rates was because they were more likely to
privately purchase a bare-bones, catastrophic plan if they did not
have access to employer-provided coverage. In that case, the ACA
dependent coverage provision may lead to larger gains along the
intensive margin of coverage for women and college graduates,
leading to larger improvements in health-related outcomes among
these groups.
The rst two columns of Table 6 report the results for females
and males. Males experienced a 2.9 percentage point larger gain
in health insurance coverage than females, and the difference
is signicant at the 1% level. Moreover, only males experienced
statistically signicant favorable effects on any outcomes besides
insurance coverage. Specically, males rates of primary care
doctor access, having any exercise, reporting very good/excellent
health, and reporting excellent health increased substantially by
4.6, 1.9, 2.9, and 3.1 percentage points, respectively. These effects
are all signicantly different from zero, and three of the four (all
but very good/excellent health) are also statistically different
from the corresponding effects on females. The only statistically
signicant result for females (besides insurance coverage) is an
adverse effect on days with health-related limitations. In sum,
the results suggest that males experienced larger improvements
in health-related outcomes from the ACA dependent coverage
provision than females, and that there appear to be multiple
reasons for this heterogeneity. Gains in insurance coverage were
larger for males, consistent with them having a higher pre-ACA
uninsured rate. Responses to obtaining insurance coverage also
appear to have been stronger for males, perhaps indicating a larger
price elasticity of demand for medical care.
The last two columns of Table 6 report the results stratifying by
college degree attainment. Both groups experienced similar gains
in insurance coverage as a result of the ACA dependent cover-
age provision. However, statistically signicant improvements in
outcomes besides health insurance are only observed for college
graduates. The mandate led to large and signicant gains for college
 S. Barbaresco et al. / Journal of Health Economics 40 (2015) 5468 65

Table 5
Placebo regressions.

Outcome variable 20032009 20022008 20012007

Treatment 10/06 Treatment 10/05 Treatment 10/04

Health care access

Any health insurance coverage 0.002 (0.005) 0.002 (0.008) 0.009 (0.007)
Any primary care doctor 0.008 (0.014) 0.002 (0.007) 0.019 (0.011)
Cost prevented care in past year 0.007 (0.011) 0.017 (0.013) 0.013 (0.007)

Preventive care utilization

Flu vaccination in past year 0.013 (0.016) 0.007 (0.008) 0.001 (0.007)
Well-patient checkup in past year 0.002 (0.014) 0.008 (0.014)
Pap test in past year 0.012 (0.014) 0.024 (0.022) 0.027 (0.025)

Risky health behaviors

Currently smokes cigarettes 0.019 (0.007)** 0.006 (0.009) 0.007 (0.008)
Alcoholic drinks in past 30 days 1.648 (1.035) 0.659 (0.584) 1.146 (0.788)
Risky drinker in past 30 days 0.001 (0.006) 0.014 (0.007)* 0.009 (0.011)
Body mass index 0.001 (0.146) 0.023 (0.196) 0.082 (0.145)
Obese 0.002 (0.008) 0.005 (0.011) 0.0005 (0.009)
Any exercise in past 30 days 0.008 (0.008) 0.008 (0.005) 0.004 (0.005)
Pregnancy 0.011 (0.010) 0.005 (0.012) 0.002 (0.007)

Self-assessed health
Overall health very good/excellent 0.011 (0.003)** 0.002 (0.009) 0.004 (0.009)
Overall health excellent 0.004 (0.005) 0.009 (0.005) 0.006 (0.008)
Days not in good mental health 0.064 (0.174) 0.005 (0.232) 0.054 (0.143)
Days not in good physical health 0.041 (0.109) 0.034 (0.121) 0.165 (0.107)
Days with health-related limitations 0.039 (0.084) 0.043 (0.051) 0.017 (0.060)

Notes: *** indicates signicant at the 1% level; ** 5% level; * 10% level. Standard errors, heteroskedasticity-robust and clustered by age, are in parentheses. All regressions
include the controls plus age, state, and time xed effects. BRFSS sampling weights are used.

Table 6
Heterogeneity by sex and education.

Outcome variable Sex Education

Female Male Not college graduate College graduate

Health care access

Any health insurance coverage 0.045 (0.017)** 0.074 (0.016)***+++ 0.067 (0.019)** 0.061 (0.013)***
Any primary care doctor 0.016 (0.009) 0.046 (0.012)**+ 0.025 (0.012) 0.051 (0.006)***++
Cost prevented care in past year 0.019 (0.021) 0.016 (0.013) 0.014 (0.017) 0.034 (0.009)**

Preventive care utilization

Flu vaccination in past year 0.020 (0.012) 0.012 (0.010) 0.022 (0.008)** 0.003 (0.009)
Well-patient checkup in past year 0.013 (0.014) 0.013 (0.016) 0.006 (0.019) 0.035 (0.016)
Pap test in past year 0.004 (0.015) 0.007 (0.021) 0.008 (0.028)

Risky health behaviors

Currently smokes cigarettes 0.011 (0.011) 0.004 (0.016) 0.001 (0.008) 0.002 (0.006)
Alcoholic drinks in past 30 days 0.117 (0.441) 0.359 (1.559) 0.068 (1.171) 0.398 (0.985)
Risky drinker 0.009 (0.012) 0.015 (0.014) 0.016 (0.004)*** 0.007 (0.007)
Body mass index 0.133 (0.153) 0.018 (0.160) 0.001 (0.050) 0.254 (0.096)**
Obese 0.010 (0.010) 0.005 (0.012) 0.004 (0.009) 0.017 (0.004)***
Any exercise in past 30 days 0.010 (0.007) 0.019 (0.004)***++ 0.001 (0.006) 0.010 (0.005)
Pregnancy 0.003 (0.005) 0.005 (0.006) 0.001 (0.006)

Self-assessed health
Overall health very good or excellent 0.001 (0.022) 0.029 (0.009)** 0.007 (0.009) 0.029 (0.017)
Overall health excellent 0.003 (0.009) 0.031 (0.005)***++ 0.002 (0.006) 0.037 (0.012)**
Days of last 30 not in good mental health 0.100 (0.196) 0.083 (0.160) 0.259 (0.154) 0.323 (0.193)+++
Days of last 30 not in good physical health 0.109 (0.081) 0.011 (0.167) 0.211 (0.145) 0.262 (0.166)
Days of last 30 with health-related limitations 0.347 (0.110)** 0.102 (0.206) 0.265 (0.161) 0.149 (0.096)

Notes: +++ difference between effects on subgroups is signicant at the 1% level; ++ 5% level; +10% level. See other notes for Table 5.

graduates in the following outcomes: primary care doctor access
(5.1 percentage points), cost being a barrier to care (reduction of 3.4
percentage points), BMI (reduction of 0.25 units), obesity (reduc-
tion of 1.7 percentage points), and excellent self-reported health
(increase of 3.7 percentage points). Besides insurance, the only sig-
nicant effects for non-college graduates are unfavorable: a 2.2
percentage point reduction in u vaccinations and a 1.6 percentage
point increase in risky drinking. In short, college graduates expe-
rienced greater improvements in health-related outcomes than
non-college graduates, and this appears to be due to heteroge-
neous effects of coverage rather than heterogeneous effects on
coverage. This is consistent with a Grossman-style story in which
education enables individuals to better take advantage of their
health care opportunities. However, the results could also be partly
attributable to greater gains along the intensive margin of cover-
age for college graduates, which we cannot measure in our data.
Regardless of the reason, these results suggest that the mandate
increases SES-based disparities in health.
66 S. Barbaresco et al. / Journal of Health Economics 40 (2015) 5468
7. Discussion
The rst major insurance expansion under the ACA a provi-
sion requiring insurers to allow young adults to remain on their
parents health insurance until turning 26 was implemented in
September 2010. This paper uses data from the BRFSS to examine
the effects of this mandate on various outcomes related to health
care access, preventive care utilization, risky health behaviors, and
self-assessed health. We implement a difference-in-differences
model with individuals slightly below the mandates age cutoff
(ages 2325) as the treatment group and those slightly above the
cutoff (ages 2729) as the control group.
We rst estimate average effects for the entire sample.
The results suggest that the ACA dependent coverage provision
increased health care access but not utilization of preventive care,
had mixed effects on risky health behaviors, and improved health
at the high end of the distribution. Specically, we observe signif-
icant and robust favorable effects on health insurance, access to a
primary care doctor, probability of having excellent self-assessed
health, and BMI. However, we also nd an adverse effect on risky
drinking consistent with ex ante moral hazard and no clear effects
on the other outcomes. We then validate our model through a series
of placebo tests and show that our classications of treatment
and control groups perform better in these tests that the wider
age bandwidths common in the literature. Finally, we conduct
subsample analyses, nding particularly striking improvements in
outcomes for men and college graduates. Men had larger gains
in health insurance coverage than women, and only men experi-
enced statistically signicant gains in any outcomes beyond health
insurance specically primary care access, exercise, and overall
self-assessed health. Insurance expansions were similar for college
graduates and non-college graduates, but only college graduates
experienced signicant gains in any other outcomes: primary care
access, cost being a barrier to care, BMI, obesity, and overall self-
assessed health.
The ACA dependent coverage mandate provides a unique oppor-
tunity to study a health insurance intervention specic to young
adults as opposed to seniors (Medicare), the poor (Medicaid), or the
uninsured population at large (the Massachusetts reform). In gen-
eral, our results suggest that health insurance affects health-related
outcomes of young adults more modestly than prior studies have
observed for these other populations. First, we nd no evidence
of increased preventive care utilization, in contrast to prior results
from both Medicaid (Finkelstein et al., 2012) and the Massachusetts
reform (Kolstad and Kowalski, 2012). Second, we only nd statis-
tically signicant improvements in overall self-assessed health at
the top of the distribution, as reporting of excellent health increases
but there is no clear evidence of an effect on reporting very good
or excellent health. We do not observe any gains in the variables
reecting more severe health conditions: days not in good physical
health, days not in good mental health, and days with functional
limitations. This contrasts the clear gains in these same outcomes
observed for both Medicaid (Finkelstein et al., 2012) and the Mas-
sachusetts reform (Van der Wees et al., 2013; Courtemanche and
Zapata, 2014). Interestingly, Chua and Sommers (2014) nd that
the ACA dependent coverage provision increased the probabilities
of self-reporting excellent physical and mental health. Combining
their results with ours suggests that physical and mental health did
improve, but only at the high end of the distribution.
While our results suggest that health insurance expansions
for young adults are less impactful than those for other age
groups, it is still important to emphasize that we do observe
some improvements in important outcomes, including health care
access, excellent self-assessed health, and BMI. One might have ini-
tially worried that a coverage expansion for young adults would not
lead to any health improvements given the generally good baseline
health of this age group.
An important contribution of our paper is that we provide, to our
knowledge, the rst empirical investigation of ex ante moral hazard
that focuses specically on young adults. We nd evidence consis-
tent with ex ante moral hazard in only one domain: risky drinking
(binge drinking or excessive number of drinks per month). In con-
trast, we nd evidence that the dependent coverage improved
weight-related behaviors while not affecting smoking and preg-
nancies. Our results therefore suggest that ex ante moral hazard is
domain-specic.
Another interesting result is that, since the improvement in
health is concentrated among college graduates, the ACA depend-
ent coverage provision appears to increase SES-based disparities
in health. This is contrary to the usual impacts of public poli-
cies to expand health insurance. Medicaid has been shown to
improve at least some health outcomes (Currie and Gruber, 1996a,
1996b; Finkelstein et al., 2012; Sommers et al., 2012), implying
reduced income-based disparities in health. The Massachusetts
reform also appears to have reduced income-based disparities, as
Courtemanche and Zapata (2014) found the largest gains in self-
assessed health among low-income individuals.
Several caveats to our analyses provide directions for future
research. First, since we study eighteen different dependent vari-
ables, we might expect one or two results to emerge as signicant
at conventional levels simply by chance. We did not employ multi-
ple hypothesis test adjustments in this paper because, even though
such adjustments control the Type I error rate (probability of falsely
rejecting any null hypotheses), they do so at the cost of substan-
tially increasing the Type II error rate (probability of failing to reject
false null hypotheses).20 However, future research should revisit
our questions using different data to see if any of our ndings could
be attributable to chance rather than genuine causal effects of the
mandate.
Next, we focus on estimating the ACA dependent cov-
erage provisions effects on 2325 year olds, ignoring
possible effects on 1922 year olds because of the greater
difculty in nding a suitable control group and the
weaker ex ante expectations of signicant effects. Further under-
standing whether benets accrue to young adults besides 2325
year olds is obviously important in order to fully evaluate the
policy.
Further research is also necessary to understand the mech-
anisms through which the mandate improves health. Increased
health care utilization is an obvious possibility, but early evi-
dence on the ACA provisions impact on health care consumption
is mixed. Akosa Antwi et al. (2014) report a rise in hospitaliza-
tions using administrative data, but Chua and Sommers (2014) nd
no evidence of changes in survey-based measures of hospital care,
primary care, or prescription drug utilization, while we nd no sig-
nicant increases in preventive care. Another possible explanation
is that self-assessments of health improve due to a warm glow
20
For instance, the simple Bonferroni correction involves multiplying all p-values
by the number of hypotheses being tested, which is eighteen in our case. This would
make it virtually impossible to reject any null hypothesis in regressions that already
demand quite a bit of the data by including xed effects and clustering at an aggre-
gated level. It is not clear to us that it would be preferable to, for example, fail to
reject ve false null hypotheses for the sake of not rejecting one true null hypothesis.
This seems especially true in cases such as ours, where null results are an important
part of the story. Moreover, we view our analyses as testing for eighteen distinct
effects, some of which are more plausible theoretically than others, as opposed to
testing for one effect that may manifest itself through eighteen different measures. It
is not clear why, for instance, we should inate the p-values in the health insurance
regressions merely because we also study smoking, pregnancies, etc.
 S. Barbaresco et al. / Journal of Health Economics 40 (2015) 5468 67

from the peace of mind of having insurance. Finkelstein et al. (2012)
proposed this as an explanation for their nding from the Oregon
Medicaid experiment that most of the gains in self-assessed health
appeared to occur before changes in utilization.
Finally, and critically, our results should not be interpreted as
providing a full accounting of the benets of expanding insurance
coverage among young adults. The primary purpose of insur-
ance is to protect individuals from nancial risk, and gains along
this dimension may be especially substantial for young adults
given their relatively low income and wealth levels. Moreover,
expanding coverage among young adults is an important compo-
nent of the overall strategy behind the ACA since it is necessary
to offset the additional costs of insuring older and sicker indi-
viduals under community rating. In other words, the costs and
benets of the different components of the ACA need to be evalu-
ated together, as the different pieces of the reform are designed to
work synergistically.
Acknowledgments
We thank the editor, two anonymous referees, Yaa Akosa
Antwi, Kitt Carpenter, Jim Marton, Melinda Pitts, Anne Roy-
alty, Chris Ruhm, Kosali Simon, Erdal Tekin, and audiences at
the American Society of Health Economists Biennial Conference,
University of Virginia, and Vanderbilt University for valuable
comments.
Appendix.

Table A1
Full regression output for selected dependent variables.

Control variable Insurance Smoker Excellent health

Treated * Post 0.061 (0.017)* 0.003 (0.007) 0.014 (0.005)*

Age = 24 0.004 (0.002)* 0.007 (0.001)** 0.002 (0.001)*
Age = 25 0.007 (0.003) 0.022 (0.001)** 0.004 (0.001)**
Age = 27 0.012 (0.010) 0.030 (0.002)** 0.016 (0.002)**
Age = 28 0.016 (0.011) 0.039 (0.002)** 0.026 (0.002)**
Age = 29 0.029 (0.011)* 0.032 (0.002)** 0.018 (0.003)**
Female 0.070 (0.006)** 0.056 (0.005)** 0.020 (0.008)
Non-Hispanic black 0.012 (0.010) 0.126 (0.009)** 0.006 (0.007)
Hispanic 0.117 (0.007)** 0.183 (0.004)** 0.019 (0.005)**
Other than black, Hispanic, or white 0.011 (0.010) 0.025 (0.013) 0.010 (0.004)*
Currently married 0.069 (0.008)** 0.110 (0.008)** 0.027 (0.005)**
High school degree but no further 0.110 (0.012)** 0.091 (0.017)** 0.039 (0.009)**
Some college but no 4-year degree 0.171 (0.017)** 0.161 (0.014)** 0.051 (0.008)**
College graduate 0.251 (0.019)** 0.310 (0.020)** 0.111 (0.009)**
Between $10,000 and $15,000 0.048 (0.014)* 0.007 (0.006) 0.008 (0.018)
Between $15,000 and $20,000 0.070 (0.018)* 0.021 (0.011) 0.007 (0.013)
Between $20,000 and $25,000 0.032 (0.009)* 0.007 (0.009) 0.0001 (0.009)
Between $25,000 and $35,000 0.051 (0.010)** 0.020 (0.013) 0.026 (0.013)
Between $35,000 and $50,000 0.120 (0.009)** 0.036 (0.014) 0.044 (0.010)**
Between $50,000 and $75,000 0.169 (0.011)** 0.063 (0.016)** 0.058 (0.011)**
$75,000 and over 0.179 (0.012)** 0.057 (0.015)* 0.108 (0.012)**
One child in household 0.021 (0.007)* 0.035 (0.009)** 0.012 (0.003)**
Two children in household 0.031 (0.005)** 0.044 (0.012)* 0.010 (0.008)
Three children in household 0.020 (0.008) 0.055 (0.010)** 0.025 (0.011)
Four children in household 0.017 (0.021) 0.071 (0.023)* 0.038 (0.016)
Five or more children in household 0.065 (0.023)* 0.071 (0.018)* 0.007 (0.012)
Cell phone only 0.013 (0.006) 0.007 (0.006) 0.011 (0.007)
Student 0.006 (0.016) 0.035 (0.008)** 0.013 (0.009)
Unemployed 0.164 (0.017)** 0.100 (0.010)** 0.027 (0.006)**
State unemployment rate 0.004 (0.003) 0.009 (0.004)* 0.002 (0.002)
Pre-ACA state mandate 0.017 (0.010) 0.001 (0.010) 0.015 (0.004)*

Notes: ** indicates signicant at the 1% level; * 5% level. Standard errors, heteroskedasticity-robust and clustered by age, are in parentheses. All regressions also include the
age, state, and time xed effects. BRFSS sampling weights are used. Separate variables for treated and post are not included because they are subsumed by the age and
time xed effects.
68 S. Barbaresco et al. / Journal of Health Economics 40 (2015) 5468
References
Acemoglu, D., Finkelstein, A., Notowidigdo, M.J., 2013. Income and health spend-
ing: evidence from oil price shocks. Review of Economics and Statistics 95,
10791095.
Anderson, M., Dobkin, C., Gross, T., 2012. The effect of health insurance coverage
on the use of medical services. American Economic Journal: Economic Policy 4,
127.
Anderson, M., Dobkin, C., Gross, T., 2014. The effect of health insurance on emergency
department visits: evidence from an age-based eligibility threshold. Review of
Economics and Statistics 96, 189195.
Angrist, J.D., Pischke, J.S., 2008. Mostly Harmless Econometrics: An Empiricists Com-
panion. Princeton University Press, Princeton, NJ.
Akosa Antwi, Y., Asako, S.M., Simon, K., 2013. Effects of federal policy to insure
young adults: evidence from the 2010 Affordable Care Act dependent coverage
mandate. American Economic Journal: Economic Policy 5, 128.
Akosa Antwi, Y., Asako, S.M., Simon, K., 2014. Access to Health Insurance and the Use
of Inpatient Medical Care: Evidence from the Affordable Care Act Young Adult
Mandate. National Bureau of Economic Research Working Paper #20202.
Apouey, B., Clark, A.E., 2015. Winning big but feeling no better? The effect of lottery
prizes on physical and mental health. Health Economics (forthcoming).
Baicker, K., Taubman, S.L., Allen, H.L., Bernstein, M., Gruber, J.H., Newhouse, J.P.,
Schneider, E.C., Wright, B.J., Zaslavsky, A.M., Finkelstein, A.N., 2013. The Oregon
experiment effects of Medicaid on clinical outcomes. New England Journal of
Medicine 368, 17131722.
Brook, R.H., Ware, J.E., Rogers, W.H., Keeler, E.B., Davies, A., Donald, C., Goldberg,
G.A., Lohr, K., Masthay, P.C., Newhouse, J.P., 1983. Does free care improve adults
health? Results from a randomized controlled trial. New England Journal of
Medicine 309, 14261434.
Cameron, A.C., Gelback, J.B., Miller, D.L., 2008. Bootstrap-based improvements for
inference with clustered errors. Review of Economics and Statistics 90, 414427.
Cantor, J.C., Monheit, A.C., DeLia, D., Lloyd, K., 2012. Early impact of the affordable
care act on health insurance coverage of young adults. Health Services Research
47, 17731790.
Card, D., Dobkin, C., Maestas, N., 2008. The impact of nearly universal insurance cov-
erage on health care utilization: evidence from Medicare. American Economic
Review 98, 22422258.
Card, D., Dobkin, C., Maestas, N., 2009. Does Medicare save lives? Quarterly Journal
of Economics 124, 597636.
Cawley, J., 2004. The impact of obesity on wages. Journal of Human Resources 39,
451474.
Cawley, J., Moran, J., Simon, K., 2010. The impact of income on the weight of elderly
Americans. Health Economics 19, 979993.
Centers for Disease Control and Prevention, 2012. Binge Drinking is Big-
ger Problem than Previously Thought, Available at: http://www.cdc.gov/
media/releases/2012/p0110 binge drinking.html (accessed 13.12.14).
Centers for Disease Control and Prevention, 2014. Past Weekly Surveil-
lance Reports, Available at: http://www.cdc.gov/u/weekly/pastreports.htm
(accessed 12.12.14).
Chua, K.P., Sommers, B.D., 2014. Changes in health and medical spending among
young adults under health reform. Journal of the American Medical Association
311, 24372439.
Courtemanche, C., Zapata, D., 2014. Does universal coverage improve health? The
Massachusetts experience. Journal of Policy Analysis and Management 33,
3669.
Courtemanche, C., Heutel, G., McAlvanah, P., 2015. Impatience, incentives, and obe-
sity. Economic Journal (forthcoming).
Courtemanche, C., Pinkston, J., Stewart, J., 2014. Adjusting Body Mass for Measure-
ment. National Bureau of Economic Research Working Paper #19928.
Currie, J., Gruber, J., 1996a. Health insurance eligibility, utilization of medical care,
and child health. Quarterly Journal of Economics 111, 431466.
Currie, J., Gruber, J., 1996b. Saving babies: the efcacy and cost of recent changes in
the Medicaid eligibility of pregnant women. Journal of Political Economy 104,
12631296.
Dave, D., Kaestner, R., 2009. Health insurance and ex ante moral hazard: evidence
from Medicare. International Journal of Health Care Finance and Economics 9,
367390.
Dafny, L., Gruber, J., 2005. Public insurance and child hospitalizations: access and
efciency effects. Journal of Public Economics 89, 109129.
DeNavas-Walt, C., Proctor, B.D., Smith, J.C., 2010. Income, Poverty, and Health Insur-
ance Coverage in the United States: 2009. U.S. Census Bureau, U.S. Government
Printing Ofce, Washington, DC, pp. 179.
DeSalvo, K.B., Bloser, N., Reynolds, K., He, J., Muntner, P., 2006. Mortality predic-
tion with a single general self-rated health question. A meta-analysis. Journal of
General Internal Medicine 21, 267275.
Ehrlick, I., Becker, G.S., 1972. Market insurance, self-insurance, and self-protection.
Journal of Political Economy 80, 623648.
Finkelstein, A.N., McKnight, R., 2008. What did Medicare do? The initial impact of
Medicare on mortality and out of pocket medical spending. Journal of Public
Economics 92, 16441668.
Finkelstein, A.N., Taubman, S., Wright, B., Bernstein, M., Gruber, J., Newhouse, J.P.,
Allen, H., Baicker, K., The Oregon Health Study Group, 2012. The Oregon health
insurance experiment: evidence from the rst year. Quarterly Journal of Eco-
nomics 127, 10571106.
Frijters, P., Haisken-DeNew, J.P., Shields, M., 2005. The causal effect of income on
health. Evidence from German reunication. Journal of Health Economics 24,
9971017.
Grossman, M., 1972. On the concept of health capital and demand for health. Journal
of Political Economy 80, 223255.
Gruber, J., 2011. The Impacts of the Affordable Care Act: How Reasonable Are the
Projections? National Bureau of Economic Research Working Paper #17168.
Idler, E.L., Benyamini, Y., 1997. Self-rated health and mortality: a review of twenty-
seven community studies. Journal of Health and Social Behavior 38, 2137.
Jianakoplos, N.A., Bernasek, A., 1998. Are women more risk averse? Economic
Inquiry 36, 620630.
Kolstad, J.T., Kowalski, A.E., 2012. The impact of health care reform on hospital and
preventive care evidence from Massachusetts. Journal of Public Economics 96,
909929.
Lakdawalla, D., Philipson, T., 2002. The Growth of Obesity and Technological
Change: A Theoretical and Empirical Investigation. National Bureau of Economic
Research Working Paper # 8946.
Levine, P.B., McKnight, R., Heep, S., 2011. How effective are public policies to increase
health insurance coverage among young adults? American Economic Journal:
Economic Policy 3 (1), 129156.
Lichtenberg, F.R., 2002. The effects of Medicare on health care utilization and out-
comes. Frontiers in Health Policy Research, vol. 5. MIT Press, Cambridge, MA.
Lindahl, M., 2005. Estimating the effect of income on health and mortality using
lottery prizes as an exogenous source of variation in income. Journal of Human
Resources 40 (1), 144168.
Manning, W.G., Newhouse, J.P., Duan, N., Keeler, E.B., Leibowitz, A., Marquis, M.S.,
1987. Health insurance and the demand for medical care: evidence from a ran-
domized experiment. American Economic Review 77 (3), 251277.
Miller, S., 2011. The effect of insurance on outpatient emergency room visits: an
analysis of the 2006 Massachusetts health reform. Journal of Public Economics
96, 893908.
Miller, S., 2012. The impact of the Massachusetts health care reform on health care
use among children. American Economic Review: Papers and Proceedings 102
(3), 502507.
Monheit, A., Cantor, J., DeLia, D., Belloff, D., 2011. How have state policies to expand
dependent coverage affected the health insurance status of young adults?
Health Services Research 46, 251267.
National Conference of State Legislatures, 2010. Covering Young Adults through
their Parents or Guardians Health Policy, Retrieved from: http://www.ncsl.org/
research/health/dependent-health-coverage-state-implementation.aspx
National Highway Trafc Safety Administration, 2014. Trafc Safety Facts
2012: Alcohol-Impaired Driving, Available at: http://www-nrd.nhtsa.
dot.gov/Pubs/811870.pdf (accessed 14.12.14).
National Institute on Alcohol Abuse and Alcoholism, 2013. College Drink-
ing, Available at: http://pubs.niaaa.nih.gov/publications/CollegeFactSheet/
CollegeFactSheet.pdf (accessed 14.12.14).
Phillips, A.C., Der, G., Carroll, D., 2010. Self-reported health, self-reported tness,
and all-cause mortality: prospective cohort study. British Journal of Health Psy-
chology 15, 337346.
Ruhm, C.J., 2000. Are recessions good for your health? Quarterly Journal of Eco-
nomics 115 (2), 617650.
Ruhm, C.J., 2002. Does drinking really decrease in bad times? Journal of Health
Economics 21 (4), 659678.
Ruhm, C.J., 2005. Healthy living in hard times. Journal of Health Economics 24 (2),
341363.
Ruhm, C.J., 2013. Recessions, Healthy No More? National Bureau of Economic
Research Working Paper #19287.
Schmeiser, M.D., 2009. Expanding wallets and waistlines: the impact of family
income on the BMI of women and men eligible for the earned income tax credit.
Health Economics 18, 12771294.
Slusky, D.J G., 2013. Revisiting the Labor Effects of the Affordable Care Act. Princeton
University Working Paper.
Snyder, S.E., Evans, W.N., 2006. The impact of income on mortality: evidence
from the social security notch. Review of Economics and Statistics 88 (3),
482495.
Sommers, B.D., Baicker, K., Epstein, A.M., 2012. Mortality and access to care among
adults after state Medicaid expansions. New England Journal of Medicine 367,
10251034.
Sommers, B.D., Buchmueller, T., Decker, S.L., Carey, C., Kronick, R., 2013. The afford-
able care act has led to signicant gains in health insurance and access to care
for young adults. Health Affairs 32, 165174.
Sommers, B.D., Kronick, R., 2012. The affordable care act and insurance cover-
age for young adults. Journal of the American Medical Association 307 (9),
913914.
Sommers, B.D., Long, S.K., Baicker, K., 2014. Changes in mortality after Massachusetts
health care reform: a quasi-experimental study. Annals of Internal Medicine 160
(9), 585593.
Taubman, S.L., Allen, H.L., Wright, B.J., Baicker, K., Finkelstein, A.N., 2014. Medicaid
increases emergency-department use: evidence from Oregons health insurance
experiment. Science 343, 263268.
Tekin, E., McClellan, C., Minyard, K.J., 2013. Health and Health Behaviors during
the Worst of Times: Evidence from the Great Recession. National Bureau of
Economic Research Working Paper #19234.
Van der Wees, P.J., Zaslavsky, A.M., Ayanian, J.Z., 2013. Improvements in health status
after Massachusetts health care reform. Milbank Quarterly 91, 663689.
 Journal of Health Economics 40 (2015) 6982

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

Medicaid expansions for the working age disabled: Revisiting the

crowd-out of private health insurance
Kathryn L. Wagner
Department of Economics, University of Notre Dame, 434 Flanner Hall, Notre Dame, IN 46556, United States

a r t i c l e i n f o a b s t r a c t

Article history: Disabled individuals under 65 years old account for 15% of Medicaid recipients but half of all Medic-
Received 20 February 2014 aid spending. Despite their large cost, few studies have investigated the effects of Medicaid expansions
Received in revised form for disabled individuals on insurance coverage and crowd-out of private insurance. Using an eligibility
17 December 2014
expansion that allowed states to provide Medicaid to disabled individuals with incomes less than 100% of
Accepted 19 December 2014
the federal poverty level, I address these issues. Crowd-out estimates range from 49% using an ordinary
Available online 29 December 2014
least squares procedure to 100% using two-stage least-squares analysis. This potentially large degree of
crowd-out could have scal implications for the Affordable Care Act which has greatly expanded Medicaid
JEL classication:
H4
eligibility in 2014.
I1 2014 Elsevier B.V. All rights reserved.

Keywords:
Medicaid expansions
Health insurance
Crowd-out
Disability
Public programs

1. Introduction primarily based on crowd-out estimates from Medicaid expansions

A major feature of the healthcare reform enacted in the United
States under the Patient Protection and Affordable Care Act (PPACA)
is the expansion of state Medicaid programs to provide health cov-
erage to all individuals under the age of 65 who have incomes
less than 138% of the federal poverty level (FPL). As of November
2014, only 27 states have expanded their Medicaid programs, but
estimates suggest that more than 25 million uninsured Ameri-
cans would gain coverage if all states were to adopt the legislation
(Kaiser Family Foundation, 2012, 2014; Holahan et al., 2012).1 A
widespread concern of expanding Medicaid is that newly eligi-
ble individuals who have private insurance will choose to drop
their private plans and take-up Medicaid coverage instead due
to Medicaids low cost a phenomenon known as crowd-out.
The Urban Institutes Health Insurance Policy Simulation Model
(HIPSM) assumes a crowd-out rate of 22% to predict changes in
insurance coverage from the PPACA (Buettgens, 2011). This rate is
E-mail address: kwagner5@nd.edu
1
Initially, the PPACA Medicaid expansions were required of all states, but became
optional under a June 2012 Supreme Court ruling.
for children and their parents. Children and parents are expected
to make up a small portion of the new Medicaid eligibles, how-
ever, under the PPACA. If crowd-out for other populations is larger
than 22%, then the predictions from HIPSM will overestimate the
reduction in the uninsured population from the PPACA Medicaid
expansions (Holahan et al., 2012).
One subset of the population that stands to gain from the PPACA
Medicaid expansions is disabled individuals under 65. Medicaid has
typically only covered disabled individuals with very low income
levels, but the PPACA will allow those with higher levels of income
to qualify. The average income eligibility limit for the disabled was
87% of the federal poverty level (FPL) in 2008. More than 3.5 mil-
lion work disabled Americans had incomes between 87 and 138%
FPL in 2011 (Authors calculations using the March CPS 2012) and
would be eligible for coverage if all states adopt the PPACA Medic-
aid expansions.
The disabled under 65 are the most expensive coverage group
under Medicaid and accounted for 43% ($160 billion) of total Med-
icaid payments even though they made up only 15% of recipients
in 2009 (Kaiser Family Foundation: statehealthfacts.org). Despite
their expense, effects from Medicaid eligibility expansions for the
disabled population are not well understood. This paper examines
a Medicaid eligibility expansion that allowed states the option

http://dx.doi.org/10.1016/j.jhealeco.2014.12.007
0167-6296/ 2014 Elsevier B.V. All rights reserved.
70 K.L. Wagner / Journal of Health Economics 40 (2015) 6982
to offer Medicaid to their disabled residents who had monthly
incomes up to 100% of the FPL. Given their different health needs
and cost of care, the disabled population may experience a different
rate of crowd-out than children and parents.
In this paper, I focus on the effects of Medicaid eligibility on
health coverage for working-age individuals who report having a
work disability. As with earlier Medicaid expansion research, OLS
estimates are subject to omitted variables bias and measurement
error. To account for these issues, I use the simulated instrument
originally described in Currie and Gruber (1996a, 1996b) to mea-
sure only the legislative effect of changes in Medicaid eligibility.
Using two large, nationally representative datasets and a two-stage
least-squares (2SLS) analysis, I nd the rate of crowd-out to be
about 100%. This rate is much larger than what has been found
in Medicaid expansions for parents and children. The point esti-
mate of crowd-out is statistically signicant at the 5% level and is
robust to several alternative samples. The condence interval for
this estimate, however, is broad and so we interpret it with caution.
2. Literature review
Gruber and Simon (2008) provide an excellent review of the
crowd-out literature which I will briey summarize below. The
effect of Medicaid eligibility expansions on participation and
crowd-out has been primarily investigated through expansions for
children and their families. The agship paper in this eld is Cutler
and Gruber (1996) who used a 2SLS approach to investigate the
effect of Medicaid eligibility expansions for children and pregnant
women on health insurance coverage with the simulated eligibility
instrument of Currie and Gruber (1996a, 1996b). Cutler and Gruber
(1996) found a large and statistically signicant degree of crowd-
out, but many follow-up papers found smaller amounts. Across the
entire literature, crowd-out measures have varied widely from as
small as 0% to as large as 60% for children and families.2
Subsequent papers explain the smaller size of crowd-out
through several key criticisms of the Cutler and Gruber (1996)
empirical strategy, one of which is that they used the annual from
the March Current Population Survey (CPS) rather than monthly
data from the Survey of Income and Program Participation (SIPP).
In this paper, I analyze data from both the CPS and the SIPP and
nd similar results for the disabled population. Another critique in
this literature is that crowd-out estimates can be sensitive to the
treatment of individuals reporting private and Medicaid coverage
simultaneously (Gruber and Simon, 2008). Populations with over-
lapping coverage can be interpreted in multiple ways, but it has
become standard practice to assume that this group is transition-
ing from private insurance to Medicaid. In my analysis, I nd little
difference in crowd-out measures accounting and not accounting
for the overlap population.
In addition to crowd-out for children and parents, a recent paper
nds evidence of crowd-out among the childless adult population.
Garthwaite et al. (2013) nds that private health insurance cov-
erage sharply increased for childless adults in conjunction with
an increase in their labor supply after their disenrollment from
TennCare, an extension of Tennessees Medicaid program. This
implies that childless adults pursued health coverage through an
employer after losing public coverage and this suggests evidence of
2
See Aizer and Grogger (2003), Blumberg et al. (2000), Card and Shore-Sheppard
(2004), Cutler and Gruber (1996), Dubay and Kenney (1996, 1997), Gruber and
Simon (2008), Ham and Shore-Sheppard (2005), Hamersma and Kim (2013), Hudson
et al. (2005), LoSasso and Buchmueller (2004), Shore-Sheppard (2008), Thorpe and
Florence (1998), and Yazici and Kaestner (2000) for these estimates.
crowd-out.3 The results from Garthwaite et al. (2013) are informa-
tive regarding the PPACA Medicaid expansions since the majority
of those who will become eligible are childless adults. Disabled
individuals, though they make up a smaller portion of the PPACA
eligible population, will gain some eligibility through the expan-
sions. Given their expensive health care needs, the cost of the
PPACA expansions for disabled individuals may be economically
meaningful despite their small population size.
3. Medicaid, eligibility expansions, and eligibility pathways
for the disabled
In scal year 2010, Medicaid spent $389 billion to provide
health coverage to roughly 60 million low-income Americans
(Kaiser Family Foundation: statehealthfacts.org). The program is
means-tested and jointly funded by both federal and state gov-
ernments with the federal government covering at least half of
all Medicaid spending for each state (Kaiser Family Foundation:
statehealthfacts.org).4 States are responsible for the administra-
tion of their own Medicaid programs, but are required to meet
minimum federal requirements in terms of services covered and
eligibility standards. While Medicaid also covers groups such as
children, pregnant women, parents, and the elderly, I will focus on
the coverage of disabled individuals.
There are several pathways through which disabled people can
become eligible for Medicaid coverage. I impute eligibility based
only on three pathways where an individual qualies for cov-
erage by having low income and a disability: the Supplemental
Security Income Program (SSI), the State Supplemental Payment
Program (SSP), and poverty-related coverage. SSI and SSP are pro-
grams that provide cash assistance to disabled individuals and
typically come with Medicaid coverage automatically.5 These path-
ways are accounted for when imputing eligibility, but there were
few changes in the eligibility standards for these programs from
1995 to 2007 and they do not provide much variation.
Of the three main eligibility pathways used to impute eligibil-
ity in this paper, poverty-related Medicaid has experienced the
greatest changes recently and generates most of the variation that
identies the econometric model. The Omnibus Reconciliation Act
of 1986 (OBRA86) authorized this pathway and gave states the
option to increase the Medicaid income eligibility level for the dis-
abled up to 100% of the federal poverty level (FPL). Prior to the
OBRA86, disabled individuals mainly had to qualify under SSI which
required an income level below 74% FPL of income (Social Security
Administration, 2012). While some states opted into the 100% FPL
pathway soon after the passage of the OBRA86, after this initial
round of adoptions, eight states increased the income eligibility
level for the disabled to 100% of the FPL between 1998 and 2003
period (Bruen et al., 2003). I take advantage of the variation in
timing of adopting the poverty-related coverage as well as the mag-
nitude of the income eligibility level during this later enactment
period to identify changes in Medicaid take-up and crowd-out.
3
It should be noted, however, that the results for the TennCare disenrollment are
unique in the literature and several other papers (in particular those on the Oregon
Health Insurance Experiment) do not nd big changes in employment due to health
insurance (Baicker et al., 2013).
4
The federal reimbursement rate of Medicaid expenses in each state is called the
Federal Medical Assistance percent (FMAP). FMAPs range from 50% to 74.73%. States
with lower incomes per capita relative to the US income per capita receive higher
reimbursement rates from the federal government (Baumrucker, 2010).
5
This is not always the case. 209(b) states are allowed to have stricter standards
for eligibility than SSI standards. Some states also require a separate application for
Medicaid in addition to their SSI application.
 K.L. Wagner / Journal of Health Economics 40 (2015) 6982 71

There are other pathways for disabled Medicaid eligibility that Table 1
Medicaid upper income thresholds for the disabled population: 1996 and 2007.
are not used for eligibility imputation in this analysis. One such
pathway is a Medically Needy program through which some states State Income eligibility State Income eligibility
provide Medicaid to disabled individuals with exorbitant medi- limit (%FPL) limit (%FPL)
cal fees. Another pathway is a state buy-in program that allows 1996 2007 1996 2007
disabled individuals with incomes above the eligibility levels to
AK 131 117b MT 74 74
purchase Medicaid as an insurance plan. Buy-in plans required that AL 82 80 NC 74 100
purchasers be employed and pay a premium for coverage which is AR 74 80 ND 74 74
unlike the three pathways used to impute eligibility. To ensure that AZ 74 100a NE 100 100
buy-in programs are not what drives my results, I run an additional CA 99 100 NH 79 77
CO 83 77 NJ 100 100
model that adds buy-in program indicator variables to the main CT 118 94 NM 74 74
model.6 DC 100 100 NV 80 78
The nal pathway excluded from eligibility imputation is cost DE 74 74 NY 88 84
sharing options with Medicare or dual-eligibles (see Appendix A FL 90 88 OH 63 64
GA 74 100 OK 83 100
for more information on these individuals). The benets package
HI 115 115b OR 74 74
for dual-eligibles is not as generous as the package received under IA 74 74 PA 100 100
other pathways of eligibility. As a result I treat individuals who ID 80 78 RI 84 100
qualify for Medicaid coverage under cost sharing as having an IL 48 100 SC 100 100
alternative form of Medicaid and Medicaid participation results in IN 74 74 SD 76 76
KS 74 74 TN 74 74
this paper do not consider Medicare cost sharing recipients.7 See
KY 74 74 TX 74 74
Appendix A for further information on Medicaid eligibility path- LA 74 74 UT 74 100
ways for the disabled. MA 133 133a VA 74 80
MD 74 74 VT 80 80
ME 100 100 WA 78 78
4. Data and methodology MI 76 100 WI 87 84
MN 85 95 WV 74 74
Following the approach from Cutler and Gruber (1996) for MO 74 85 WY 74 74
measuring the impact of Medicaid expansions on take-up and MS 100 100

crowd-out, we estimate the following equation
Coverageist = + Eligibleist 1 + Xist 2 + 2 + t + ist
where i, s, and, t index the individual, state, and year respec-
tively. The dependent variable, Coverageist , is an indicator variable
that equals 1 if an individual has health coverage of a certain
type (Medicaid, Private, or Uninsured), and equals zero otherwise.
The key variable of interest, Eligibleist , is an indicator variable that
equals 1 if a person is eligible for Medicaid coverage and zero oth-
erwise. Xist is a vector of demographic characteristics including age,
education, sex, race, family size, number of children in the house-
Source: Brown et al. (2007), Bruen et al. (2003), Bruen et al. (1999), Coe (2005),
Congressional Research Service (1993), De Nardi et al. (2011), Horvath (1997),
(1) Kaiser Commission on Medicaid and the Uninsured (2010), Mississippi Division of
Medicaid (19912008), Social Security Administration (19912008), Stone (2002,
2011), and state Medicaid websites.
Bold and italicized states are those using the 209(b) option.
a
Arizona offers Medicaid eligibility for those at or below 100% FPL through the
Arizona Health Care Cost Containment System. This system was formed through
a special demonstration waiver (Section 1115). Massachusetts extends Medicaid
eligibility even further to 133% FPL using a Section 1115 demonstration as well.
b
Hawaii and Alaska have higher federal poverty levels than the continental
United States. The upper income limits for these two states are constructed relative
to the poverty level of the 48 remaining states.

hold, and marital status. The variables s and t are state and year
xed effects respectively and the nal term is a random error.
In order to impute eligibility, we must rst determine the
Medicaid standards specic to each state and year. There is no
central database for state Medicaid eligibility policies. Following
Brown et al. (2007) and Coe (2005), I compiled Medicaid eligibil-
ity information from numerous sources.8 With this information,
I constructed the income eligibility limits for disabled Medicaid
applicants in each state for all years between 1996 and 2007. The
income eligibility limit is the highest level of income (as a percent
of the federal poverty level) that a disabled individual can have and
6
With the addition of controls for the presence of a Buy-in program the estimates
of take-up for Medicaid and private insurance is 0.404 (0.140) and 0.445 (0.148)
respectively. The coefcients on the Buy-in indicator variable in these models are
0.002 (0.011) and 0.009 (0.015) suggesting that the presence of a buy-in program
had little effect on the Medicaid and Private insurance take-up rates. State clustered
standard errors are in parentheses.
7
Including cost sharing individuals in the analysis produces Medicaid and private
take-up estimates of 0.339 (0.180) and 0.529 (0.213), respectively. Standard errors
are in parentheses. These estimates suggest a crowd-out point estimate of 156%
which is even larger than the primary specication.
8 9
The specic sources used to construct the upper income threshold of Medicaid
eligibility rules for the aged and disabled were Brown et al. (2005), Bruen et al. (2003,
1999), Coe (2005), Congressional Research Service (1993), De Nardi et al. (2011),
Horvath (1997), Kaiser Commission on Medicaid and the Uninsured (2010), Kassner
and Shirley (2000), Mississippi Division of Medicaid (19912008), Social Security
Administration (19912008), Stone (2002, 2011), and state Medicaid websites.
still qualify for Medicaid coverage. This is constructed by taking the
most generous of the SSI/209(b), poverty-related, and SSP path-
ways available to the disabled in a given state and year. The 1996
and 2007 income eligibility limits by state are shown in Table 1.
Using the eligibility standards for each state, we can impute eli-
gibility for an individual. To do this we use the 1996, 2001, and
2004 panels of the Survey of Income and Program Participation
(SIPP). Each SIPP panel has 12 waves (interviews) except the 2001
panel which only has nine and surveys at least 40,000 households.
Every four months, the SIPP asks respondents a set of core questions
including information on basic demographics (age, sex, race, etc.),
income, and participation in programs such as Medicaid or Supple-
mental Security Income. Respondents are interviewed every four
months, but are asked to report information for each of the four
months in the reference period. Thus the core data in the SIPP are
at the person-month level.9 I limit the analysis, however, to only
observations reported in the fourth reference month due to the
potential for transitions in health coverage or family status only
Given that the SIPP is a panel data set it is possible that a respondent could
have up to 12 interviews in any given pattern. Standard errors throughout the main
analysis are clustered at the state level since the Medicaid expansions vary by the
state. When we cluster at the individual level, however, the standard errors are
similar. The individually clustered standard errors for Medicaid, Private, and Overlap
take-up estimates are (0.147), (0.160), and (0.063) respectively.
72 K.L. Wagner / Journal of Health Economics 40 (2015) 6982
being reported between interviews (seams) rather than between
months. This is consistent with other work on Medicaid using the
SIPP (Hamersma and Kim, 2013; Gruber and Simon, 2008).
4.1. Denition of disability
The ofcial denition of disability for Medicaid eligibility is the
same denition the Social Security Administration (SSA) employs
for SSI qualication (Families USA, 2001). This denition also
applies for the higher income individuals qualifying under the
OBRA86 Medicaid expansions. Per SSA guidelines an adult is dis-
abled if she has a medically determinable physical or mental
impairment that prevents substantial gainful activity and is
expected to result in death or has lasted or can be expected to
last for a continuous period of not less than 12 months (Social
Security Administration, 2012). This is a broad denition that relies
heavily on a doctors subjective opinion as to the severity of the
condition rather than a decisive cutoff measurement. Household
surveys do not contain a measure of this exact denition and more
often contain information regarding different difculties or specic
conditions of an individual.
The closest measure of disability that is available in both the SIPP
and March CPS are the work disability variables in those surveys.10
These certainly have their limitations in that they are self-reported
in both surveys and individuals may not have a clear conception of
disability or impairment status when answering the question (US
Census Bureau, 2012). Despite these limitations, however, there is
some evidence to document that disability trends in work disability
status may be related to disability trends measured by impairment
status. This suggests that though the work disability measure is not
perfect it is correlated with other conceptualizations of disability
(Burkhauser et al., 2003).
Within the SIPP, 85% of individuals aged 2064 who report hav-
ing a work disability have a total personal income less than the
average per capita income in the United States and 65% of these
individuals have incomes less than half the average per capita
income. Since substantial gainful activity for disability is com-
monly measured in terms of an income cutoff, the work disabled
sample appears to be identifying a population that may have dif-
culties performing substantial gainful activity. Measurement error
for this disability denition also does not appear to be related to the
Medicaid legislative changes as individuals were not more likely to
report a work disability after the eligibility expansions went into
effect (results not shown).
4.2. Imputing eligibility
A person is eligible for Medicaid coverage if her personal
income minus income disregarded from eligibility determination
is less than the threshold in a persons state of residence. Eligibil-
ity is dened as a persons own eligibility.11 Income considered
10
The specic question from the SIPP survey is Does . . . have a physical, mental,
or other health condition that limits the kind or amount of work . . . can do at a job or
business? (United States Census Bureau, 1996). The specic question in the March
CPS is (Do you/Does anyone in this household) have a health problem or disability
which prevents (you/them) from working or which limits the kind or amount of
work (you/they) can do? (United States Census Bureau, 2012).
11
Gruber and Simon (2008) demonstrated the sensitivity of results to family
spillover effects where a child could benet from a sibling being Medicaid eligi-
ble. Since the sample I am considering includes working-age disabled adults rather
than children and one individuals Medicaid status does not inuence anothers, I
do not use family income to determine eligibility. I do, however, run an analysis
where I include spousal income in eligibility determination. In this analysis, a cou-
ple is eligible for Medicaid if at least one of them reports having a work disability
and the total income from both individuals is less than the Federal Poverty Level
during eligibility determination is referred to as countable income
and is constructed by subtracting disregards any income that
is not considered when determining Medicaid eligibility from
a persons total income. There are several types of income that
are disregarded from Medicaid eligibility determination. By fed-
eral law, every individual receives an automatic monthly disregard
of $20. Some states (such as California and Connecticut) choose
to disregard larger amounts. Also, income received through the
Supplemental Security Income program does not count toward
Medicaid eligibility.12 In addition to SSI payments, a portion of
earned income is disregarded.13 Subtracting the total amount of
disregards from an individuals total income produces countable
income for Medicaid eligibility.14
In reality, Medicaid eligibility is determined not only using a per-
sons income level but also her nancial assets. The SIPP, however,
only contains asset information once a year rather than a monthly
basis like income information. Due to this restriction, my main anal-
ysis does not consider assets when determining eligibility.15 This
is a restriction that most papers on crowd-out face and assets usu-
ally have not been used in the imputation of eligibility. The SIPP
does not uniquely identify every state and instead groups some
states together in some of its panels. As a result, I drop ve states
from the analysis (Maine, North Dakota, South Dakota, Vermont,
and Wyoming).
Given the research question, I limit the sample to respondents
within the SIPP who report a work-limiting disability. I eliminate
everyone under the age of 20 (who could potentially qualify for
Medicaid under eligibility rules for children which are more gen-
erous) and the elderly (those over the age of 65, most of who
automatically qualify for Medicare). Across all three panels of the
SIPP considered in this study, 41,554 individuals between 20 and 64
years of age reported having a work-limiting disability. In 1996, the
rst year of our sample, 819 individuals reported a work-limiting
disability and had incomes between 74 and 100% of the FPL and was
approximately 1.5% of the working age population in that year. Of
these 819 individuals, 37% resided in a state that would eventually
expand Medicaid to the disabled using the OBRA86 option.
Summary statistics for the ineligible and the eligible sample in
the SIPP are presented in Table 2. The observed differences between
the two samples are consistent with previous research. The eligible
sample contains a higher percentage of women, blacks, and those
individuals with lower levels of education which is not surprising as
these characteristics are associated with lower income levels and
disadvantaged populations. As expected, there is a dramatic dif-
ference in income between the ineligible and the eligible samples
for a family of size 2. The crowd-out measurements from the spousal analysis are
not substantially different from the individual eligibility analysis. Spousal analysis
Medicaid, Private, and Overlap take-up rates are 0.445 (0.187), 0.532 (0.152), and
0.031 (0.094) respectively. State clustered standard errors are in parentheses.
12
In the majority of states, a recipient of Supplemental Security Income is auto-
matically eligible for Medicaid coverage. In 2008, however, there were eleven states
that used the 209(b) option which allowed state Medicaid ofces to impose stricter
standards on Medicaid eligibility than federal SSI requirements.
13
$65 plus *(total income earned 65) per month of earned income is not
counted toward eligibility determination.
14
Given that the income information in the SIPP is self-reported, there is likely
measurement error in the imputed eligibility variable. Hamersma and Kim (2013)
proposed an estimation strategy in which we use the actual Medicaid income
threshold (as a percent of the FPL) rather than imputed eligibility. Following this
strategy, I nd the effect on Medicaid, Private, and Overlap insurance to be 0.133
(0.050), 0.150 (0.080), and 0.043 (0.019), respectively. State clustered standard
errors are in parentheses. The Hamersma and Kim strategy take-up rates sug-
gest similar point estimates as my primary analysis, but only crowd-out estimates
accounting for the overlap group are statistically signicant at the 5% level.
15
In an annual analysis, using asset information to determine eligibility does not
greatly inuence results.
 K.L. Wagner / Journal of Health Economics 40 (2015) 6982 73

Table 2 Table 3
Summary statistics for SIPP individuals reporting a work disability by eligibility Pre and post trends in state insurance coverage rates work disabled sample
status. 19922008 March CPS.

Variable Ineligible sample Eligible sample Medicaid only Private only Overlap Uninsurance
(1) (2) (3) (4)
Average age 50 45
% Female 0.42 0.59 Pre-adoption 0.0010 0.0008 0.0005 0.0008
% Black 0.14 0.21 trend (0.0018) (0.0016) (0.0006) (0.0013)
% Married 0.53 0.41
Post-adoption 0.0049*** 0.0034* 0.0001 0.0001
% No high school 0.17 0.33
trend (0.0016) (0.0020) (0.0005) (0.0012)
% High school 0.31 0.35
% College 0.51 0.33 Mean (Ins. rate) 0.27 0.40 0.04 0.14
Average real monthly personal income $2816.18 $503.70 Observations 867 867 867 867
% Medicaid 0.13 0.47 R-squared 0.6913 0.6806 0.3498 0.5885
% Medicare 0.32 0.16
% Private 0.67 0.34 Notes: Results are from the 19922008 March CPS. Regressions are OLS regressions
% Uninsured 0.12 0.25 of four types of state insurance coverage rates (Medicaid only, Private Only, Both
Average number of children 0.52 0.57 Medicaid and Private, and Uninsured) on a pre-adoption trend and a post-adoption
Average family size 2.47 2.68 trend. All regressions control for state and year xed effects. Standard errors are
% Employed 0.46 0.22 clustered at the state level.
*
% Receiving SSI 0.06 0.31 Represents a 10% signicance level.
**
Observations 72,678 98,231 Represents a 5% signicance level.
***
Number of individuals 23,399 27,448 Represents a 1% signicance level.
All regressions are weighted using the SHADAC constructed weight for CPS health
Source: Authors calculation. insurance coverage.
Summary statistics are weighted using the individual weight in the 1996, 2001, and
2004 SIPP panels. I can reject the null at a P-value of 0.01 that the sample means
are the same across samples for all variables. 9293 individuals were in both the 1996b) and construct a simulated instrument for eligibility. Since
ineligible and eligible samples over the time period resulting in a total of 41,554
my sample size is small and not computationally overwhelming,
individuals included throughout the entire analysis.
I use the entire sample of disabled respondents from each year
to simulate the instrument rather than select a smaller random
since this is a key determinant of eligibility. The eligible group also sample as has been done in previous work. Using this national pop-
has higher rates of Medicaid coverage and uninsurance and are less ulation, the simulated instrument is constructed as the percent of
likely to have private plans or be on Medicare. people that would be eligible for Medicaid coverage under each
Since the OBRA86 expansions raised Medicaid eligibility lev- states individual eligibility rules for each year. When constructing
els above SSI qualication levels, the majority of Medicaid take-up these shares, I exclude respondents from the state whose laws are
and crowd-out will be driven by non-SSI recipients. Disabled indi- being used to simulate eligibility (Ham and Shore-Sheppard, 2005).
viduals through SSI, however, may be inherently different from These percents are then matched by state and year (based on the eli-
non-SSI individuals perhaps due to various health conditions or gibility rules that constructed them) to respondents in the sample.
income levels. Non-SSI Medicaid recipients, however, constitute a By constructing the simulated instrument, we are effectively
large share of the disabled Medicaid population. Nearly 40% of dis- producing a measure of Medicaid generosity by state and year that
abled Medicaid enrollment was for non-SSI beneciaries in 2010 is not dependent upon the characteristics of the population within
(statehealthfacts.org). This is consistent with summary statistics a given state. Instead, the variation in eligibility across state and
in the SIPP which had 42% of work disabled individuals possess- time reects the legislative changes in eligibility rules. Thus, the
ing Medicaid coverage but not receiving income from SSI. Thus, instrument depends on changes in state legislation concerning eli-
though the majority of disabled Medicaid beneciaries originate gibility being exogenous. Essentially, we want it to be the case
from the SSI program, there is also a fairly large number of non-SSI that OBRA86 adopting states did not experience different trends
Medicaid beneciaries. Given the size of their population and since in coverage rates compared to non-adopting states prior to the
the majority of future Medicaid income-eligibility expansions will expansions. In Table 3, using data from the March CPS, I test the
apply to them, take-up and crowd-out rates for the non-SSI dis- difference between pre-trends and post-trends for adopting and
abled are relevant even if they are inherently different from SSI non-adopting states. I use data from the CPS rather than the SIPP
recipients. because it allows us to observe a longer period prior to the start
OLS estimates of equation (1) are potentially subject to an of expansions. The data is collapsed to construct the rate of cover-
omitted variables bias. Characteristics that determine eligibility age for the four forms of insurance by state and year. These rates
for Medicaid are likely linked with demand for health insurance. are then regressed on state and year xed effects and two vari-
Though many of these characteristics are accounted for in the ables that measure the pre- and post-adoption trends for states.
demographic controls vector, there may be omitted variables inu- The pre-adoption trend variable is a negative integer value that
encing insurance take-up. Reverse causality may also be a concern indicates the number of years until an expanding states adoption
within the model. Given that I am examining a sample of indi- or zero for non-adopting states and years after the expansion. The
viduals reporting a disability, it is likely that many of them face post-adoption trend variable is a positive integer value that indi-
difculties working. Since a large share of private health coverage cates the number of years after an expanding states adoption or
is provided through employers, a large portion of my sample may be zero for non-adopting states and years prior to the expansion.
less likely to have private coverage due to an inability to work which Table 3 presents the regression results for the pre-treatment
also results in lower levels of income. Thus, there is a potential for analysis. There is no statistically signicant difference in the
a negative spurious relationship between Medicaid eligibility and pre-adoption trends of coverage rates between adopting and non-
health coverage. Since Medicaid eligibility is also imputed using adopting states for any of the insurance types and the coefcients
reported income information from the SIPP, measurement error is are all near zero. This implies that states experienced little differ-
likely present in the model as well. ence in coverage rate trends leading up to the expansions. The
To account for the omitted variables bias, measurement error, post-adoption trends for Medicaid and private coverage, how-
and reverse causality concerns, I follow Currie and Gruber (1996a, ever, are statistically signicantly different between adopting and
74 K.L. Wagner / Journal of Health Economics 40 (2015) 6982

Table 4 Table 5
Changes in health insurance coverage, 19962007, by income group work disability Crowd-out from difference-in-difference analysis work disability sample from 1996,
sample from 1996, 2001, and 2004 SIPP panels. 2001, and 2004 SIPP panels.

Income group Eligibility Medicaid only Private only Overlap Income group No overlap Overlap
(1) (2) (3) (4) (1) (2)

<=74% FPL 0.010 0.035 0.025 0.024*** 74100% relative to 0.778*** 0.819***
(0.008) (0.024) (0.017) (0.007) <=74% FPL (0.273) (0.256)
74100% FPL 0.207** 0.124*** 0.095*** 0.043** 74100% relative to 0.161 0.163
(0.078) (0.044) (0.029) (0.016) 100150% FPL (0.305) (0.341)
100150% FPL 0.002 0.000 0.075*** 0.043*** 74100% relative to 0.433 0.446
(0.011) (0.020) (0.026) (0.012) 150% + FPL (0.340) (0.383)
150% + FPL 0.000 0.021* 0.049** 0.042*** 74100% relative to all 0.538** 0.603**
(0.000) (0.011) (0.021) (0.009) others (0.245) (0.254)
All groups (except 0.013 0.021 0.037*** 0.032***
Source: Authors calculation.
74100% FPL) (0.010) (0.016) (0.010) (0.004)
Notes: Standard errors are in parentheses and are constructed from Table 4 esti-
mates using the delta method. The table contains crowd-out measures using
Difference-in-difference estimates
difference-in-difference estimates in Table 4. The No Overlap column does not
74100% relative to 0.198** 0.090* 0.070** 0.019 account for the population reporting both private and Medicaid insurance cover-
<=74% FPL (0.075) (0.045) (0.027) (0.015) age. The Overlap column accounts for this population assuming that the change
74100% relative to 0.206** 0.124** 0.020 0.000 in overlap coverage represents a transition from private to public coverage. Income
100150% FPL (0.080) (0.047) (0.038) (0.019) groups represent monthly income of the individual minus Medicaid disregards as a
74100% relative to 0.207** 0.104** 0.045 0.001 percent of the federal poverty level.
*
150% + FPL (0.078) (0.048) (0.037) (0.020) Represents a 10% signicance level.
**
74100% relative to All 0.220*** 0.104** 0.056** 0.011 Represents a 5% signicance level.
***
Others (0.073) (0.045) (0.028) (0.016) Represents a 1% signicance level.

Source: Authors calculation

Notes: Standard errors are in parentheses and are clustered at the state level. Income
groups are based on countable income as a percent of the federal poverty level.
Sample sizes are 12,888 in <=74% FPL group, 2467 in 74100% FPL group, 3400 in
100150% FPL group, 5533 in 150% + FPL group. Overlap represents respondents who
report having both private and Medicaid coverage. Medicaid represents respondents
reporting Medicaid and no private coverage. Private represents respondents repor-
ting private and no Medicaid coverage. Results are weighted using the individual
weight from the SIPP.
*
Represents a 10% signicance level.
**
Represents a 5% signicance level.
***
Represents a 1% signicance level.
non-adopting states (see Columns 1 and 2 in Table 3). Adopting
states experienced an increase in their Medicaid coverage rates and
a decrease in their private insurance rates relative to non-adopting
states after the eligibility expansions went into effect. The post-
adoption coverage trends for Medicare and uninsurance are not
statistically signicantly different between the two types of states.
These results are indicative of crowd-out though the magnitudes
of the coefcients in this analysis do not suggest that crowd-out
was complete. Overall, the results contained in Table 3 indicate
that pre-adoption coverage rate trends were similar for adopting
and non-adopting states and support the validity of the instrument,
but by no means are a sufcient condition for no endogeneity.
5. Results
5.1. Tabular results
Following earlier literature, we rst construct difference-in-
difference measures of crowd-out using a tabular approach. The
top section of Table 4 contains the changes in health insurance cov-
erage from 1996 (the beginning of our sample) to 2007 (the end of
our sample) across four income groups using the work disability
sample of the SIPP. In order to demonstrate the sensitivity of our
crowd-out results to individuals who report possessing both pri-
vate and Medicaid insurance coverage at the same time, we dene
three forms of health coverage: Medicaid only, private only, and
overlapping coverage (Gruber and Simon, 2008).16 In addition to
changes in health coverage, Table 4 also reports changes in imputed
eligibility for each income group. Standard errors are in parenthe-
ses and are clustered at the state level.
Given that the OBRA86 expansions effectively raised the income
thresholds for Medicaid eligibility from 74 to 100% of the FPL,
we expect individuals with countable incomes within this range
to be the most affected by the expansion. Table 4 suggests that
this is the case. The lower section of Table 4 reports difference-
in-difference estimates of eligibility and health coverage from the
74100% FPL income group relative to three other income cate-
gories (less than 74% FPL, between 100 and 150% FPL, and greater
than 150% FPL). Table 4 also compares the 74100% FPL group to all
three of these other groups combined. Imputed Medicaid eligibility
for the 74100% FPL group increased by about 20% points relative to
every other income group from 1996 to 2007. The rate of eligibility
remained virtually unchanged for the other income groups across
the time period. Coverage by Medicaid displays a similar pattern
as eligibility. Relative to the other income groups, Medicaid cover-
age increased between 9.0 and 12.4% points for the 74100% FPL
income group and 10.4% points relative to all other income groups
combined. All income groups experienced a reduction in private
coverage over the time period, but the largest decrease was for
the 74100% FPL group. The change in overlap coverage is similar
across all income groups.
Table 5 presents crowd-out measures resulting from the tabular
results in Table 4. Following Gruber and Simon (2008), we con-
struct one measure of crowd-out that ignores the overlap group
and another measure of crowd-out which interprets the overlap
coverage as a transition from private to Medicaid coverage. Table 5
presents crowd-out estimates ranging from 16 to 80% that appear
to be insensitive to the treatment of the overlap group. Standard
errors are in parentheses and are calculated from the difference-in-
difference estimates using the delta method. Relative to all of the
other income groups combined, the rate of crowd-out was between
54 and 60% which is consistent with the crowd-out estimates in
the childrens Medicaid expansions (Gruber and Simon, 2008). The

coverage (such as military insurance or Medicare), but these are not accounted for
16
Coverage denitions for the tabulation results only take into account various in the coverage denitions for this analysis. These alternative forms are considered
forms of private and Medicaid coverage. Individuals may have alternative forms of in a later robustness analysis.
 K.L. Wagner / Journal of Health Economics 40 (2015) 6982 75

Table 6
OLS estimates of insurance coverage regressions, work disability sample from 1996, 2001, and 2004 SIPP panels.

Dep. var. Medicaid only Private only Overlap Uninsurance

(1) (2) (3) (4)

Eligible 0.290*** 0.142*** 0.025*** 0.075***

(0.011) (0.010) (0.003) (0.006)
Female 0.017*** 0.078*** 0.002 0.068***
(0.005) (0.006) (0.002) (0.003)
Black 0.096*** 0.091*** 0.003 0.022***
(0.008) (0.009) (0.003) (0.007)
Hispanic 0.036 0.062*** 0.013*** 0.043***
(0.028) (0.015) (0.004) (0.015)
No high school 0.194*** 0.243*** 0.003 0.006
(0.009) (0.007) (0.004) (0.008)
High school 0.080*** 0.110*** 0.007*** 0.020***
(0.005) (0.005) (0.002) (0.004)
Married 0.197*** 0.310*** 0.014*** 0.061***
(0.006) (0.009) (0.003) (0.004)

Mean of dep. variable 0.29 0.42 0.04 0.19

R-squared 0.2558 0.3258 0.0196 0.0698

Source: Authors calculation.

Notes: Results are from the 1996, 2001, and 2004 panels of the SIPP. Regressions are OLS regressions of four forms of health coverage (Medicaid only, Private Only, Both
Medicaid and Private, and Uninsured) on imputed Medicaid eligibility. All models included 170,909 observations and control for panel-wave, state and year xed effects, a
state-linear trend, age, age squared, a cubic in income, number of children in household, and family size. Standard errors are clustered at the state level.
*
Represents a 10% signicance level.
**
Represents a 5% signicance level.
***
Represents a 1% signicance level.
All regressions are weighted using the SIPP provided person-level weight.

95% condence intervals for these estimates allow us to rule out a
crowd-out rate below 29% and above 85%.
OLS estimates are consistent with the range of crowd-out lev-
els implied by the Diff-in-Diff tabulation estimates though slightly
smaller. Crowd-out measures from OLS estimates are between 49
and 53% (see results in Table 6). Though the tabulation and OLS esti-
mates provide us with a general idea of changes in coverage and
crowd-out, they are likely to be biased and we proceed accordingly
with an instrumental variables analysis.
5.2. Two-stage least-squares estimates
Table 7 presents 2SLS estimates that correspond to Equation 1.
The key covariate in these models is the variable, Eligible, the indi-
cator variable for whether an individual is imputed to be eligible
for Medicaid, is the dependent variable in these regressions. We
instrument for this with the simulated eligibility variable and the
rst-stage estimates are reported in column (1). All models include
state, year, and SIPP panel-wave xed effects, age, age squared,
female sex, black race, Hispanic origin, marital status, less than
high school education, high school diploma, a cubic in income, and
a state-year linear trend as additional controls.17 Standard errors
allow for arbitrary correlation in errors within a state. It is no sur-
prise that simulated eligibility and imputed eligibility are positively
correlated. Indeed, rst-stage estimates in column (1) suggest a
large positive effect on the likelihood of being Medicaid eligible
with more generous Medicaid programs (as measured by the sim-
ulated instrument). First-stage estimates indicate that for a 10%
point increase in simulated eligibility, there will be a 7.32% point
increase in actual eligibility. The coefcient on the instrument is
statistically signicant at the 1% level and the F-statistic for the
null of the coefcient on the instrument is 24.
The nal four columns of Table 7 presents 2SLS estimates from
the SIPP for various forms of insurance coverage using simulated
Medicaid eligibility as an instrument for Medicaid eligibility. We
consider the effects of Medicaid eligibility on the same forms of
coverage as in the tabulation analysis and also on whether an indi-
vidual is uninsured. 18 The effects on Medicaid only, private only,
and overlap coverage do not need to sum to the effect on the unin-
sured due to potential joint coverage under Medicare or military
insurance. We explore these possibilities in a robustness check later
in the paper.
The key variable of interest is Eligible, which measures the
effects of Medicaid eligibility on the four forms of coverage. Column
(2) of Table 7 presents the results for Medicaid only participation.
Results suggest a 41% take-up rate for Medicaid after the eligibil-
ity expansions. This means that for every 100 individuals made
eligible through the expansion, 41 chose to take up Medicaid cov-
erage. In conjunction with this increase in Medicaid coverage, there
is a roughly equivalent decrease in private coverage. Estimates
suggest that 45 individuals lose private coverage for every one
hundred who become eligible for Medicaid. Coverage for the over-
lap group increases by 0.15% points for every 10% point increase
in Medicaid eligibility but is statistically insignicant. Including
the effects from the overlap group implies a Medicaid take-up
rate of 42% and a private coverage reduction rate of 46.5%. The
rate of uninsurance increases by 1.11% points for every 10% point
increase in Medicaid eligibility though this result is not statistically
signicant.
Table 8 presents the crowd-out estimates implied by the take-
up rates in Table 7. For the purposes of this paper, we dene
crowd-out to be the change in private insurance coverage relative
to the change in Medicaid coverage.19 The results in the No Over-
lap column (1) ignores any effects from the overlap group while the

17
SIPP panel-wave xed effects were included in the regressions to account for
differences in interview structures. Excluding these xed effects does not greatly
affect the results.
18
We dene a respondent to be uninsured if she reports no coverage from Med-
icaid, private, military, or Medicare insurance.
19
Some crowd-out literature denes an alternative measure of crowd-out to be
the percent change in Medicaid that is not associated with a reduction in the
76 K.L. Wagner / Journal of Health Economics 40 (2015) 6982

Table 7
First-stage estimates for and 2SLS estimates of insurance coverage regressions, work disability sample from 1996, 2001, and 2004 SIPP panels.

First-stage 2SLS estimates

Dependent Eligible Medicaid only Private only Overlap Uninsurance

Variable (1) (2) (3) (4) (5)

Simulated eligibility 0.739***

(0.149)
Eligible 0.407*** 0.450*** 0.015 0.111
(0.146) (0.148) (0.061) (0.134)
Female 0.063***
0.009 0.098*** 0.001 0.070***
(0.005) (0.012) (0.011) (0.004) (0.008)
Black 0.031*** 0.093*** 0.081*** 0.003 0.023***
(0.006) (0.010) (0.010) (0.003) (0.008)
Hispanic 0.020 0.034 0.056*** 0.013*** 0.043***
(0.017) (0.028) (0.013) (0.005) (0.016)
No high school 0.082*** 0.185*** 0.218*** 0.003 0.003
(0.008) (0.015) (0.014) (0.005) (0.013)
High school 0.013** 0.078*** 0.106*** 0.007*** 0.020***
(0.006) (0.005) (0.006) (0.002) (0.005)
Married 0.010** 0.196*** 0.307*** 0.014*** 0.060***
(0.004) (0.006) (0.010) (0.003) (0.004)

Mean of dep. variable 0.575 0.28 0.42 0.04 0.19

1st stage F-test 24.70
(P-value on null)
R-squared 0.4755 0.2472 0.2757 0.0193 0.0688

Source: Authors Calculation.

Notes: Standard errors are in parentheses and are clustered at the state level.
*
Represents a 10% signicance level.
**
Represents a 5% signicance level.
***
Represents a 1% signicance level.
All models have 170,909 observations. All regressions are weighted using the SIPP provided person-level weight. All regressions control for panel-wave, state and year xed
effects, a state linear trend, age, age squared, a cubic in income, number of children in household, and family size.

Table 8
The extent of crowd-out from work disability population from 1996, 2001, 2004
panels of the SIPP.
No Overlap Overlap
(1) (2)
Crowd-out rate (%) 1.11 1.10
95% Condence interval [0.262, 3.899] [0.329, 2.938]
Notes: Numbers in brackets are 95% condence intervals and these values are con-
structed using a clustered bootstrap with 350 repetitions. The No Overlap column
estimates the extent of crowd-out without considering the population with overlap-
ping coverage. The Overlap column includes the overlapping coverage population in
its calculation assuming that the change in overlap coverage represents a transition
from private to public coverage.
Overlap column (2) accounts for the overlap group by assuming its
effect represents a transition from private insurance to Medicaid.
Crowd-out measures appear to be insensitive to the treatment of
the overlap group. The SIPP estimates suggest complete crowd-out
between 110 and 111%. Measures of crowd-out are statistically sig-
nicant at the 5% level for both the overlap and no overlap groups
using bootstrapped condence intervals.20 The condence inter-
vals allow us to reject the hypothesis that crowd-out was equal to
zero at a 5% level in the SIPP and also allow us to reject crowd-out
rates lower than 26% and higher than 390%. Given that the 95% con-
dence intervals suggest a broad range of crowd-out estimates, we
need to interpret the crowd-out point estimate with caution. The
intervals do suggest, however, that crowd-out for the disabled was
substantial. In a later robustness check, I derive a similarly large
point estimate of crowd-out using an alternative sample from the
March CPS. Since the estimates across the two samples are similar,
this is suggestive that crowd-out in the disabled population was
near 100%.
The crowd-out measure and take-up rates for the disabled
population in the 2SLS analysis are much larger than what has
been previously found for other populations such as children
and their parents. There are several reasons this might be the
case. First, the disabled are more likely to experience larger total
healthcare expenses than children due to their chronic medi-
cal conditions requiring continuing treatment. Given this, private
insurance plans for the disabled may have larger premiums, coin-
surance, or deductibles making Medicaid insurance plans (which
are free) a more attractive option.21
Second, disabled individuals may use the Medicaid expan-
sions as a method of pursuing Social Security Disability Insurance
(SSDI). Since most private health insurance is provided through
an employer, the two year waiting period for Medicare coverage
might discourage disabled individuals from securing SSDI coverage
due to the job lock phenomenon. With the possibility of Med-
icaid coverage now available to them during the waiting period,
individuals may have more incentive to leave their jobs (which

uninsurance rate. Since the analysis always nds an increase in the uninsurance
rate, the alternative measure always suggests a crowd-out rate of over 100%.
20
Following the methods of Gruber and Simon (2008) we construct condence
intervals for the crowd-out measures using a clustered bootstrap with 350 replica-
tions. For each of the 350 replications we construct crowd-out measures accounting
for and not accounting for the overlap group. We then take the 5th and 95th per-
centile of the constructed crowd-out measures to create the condence intervals.
21
It is puzzling that all disabled individuals in the sample would choose to remain
uninsured rather than accept Medicaid coverage. About sixteen percent of the dis-
abled sample is uninsured which is not incredibly different from the insurance rate
of the overall population during this time period. One possible explanation is that
individuals with disabilities are more strongly inuenced by social stigma than other
types of individuals. Another explanation could be that disabled individuals lack the
knowledge or ability to pursue the Medicaid application which is often a lengthy
and complicated process.
 K.L. Wagner / Journal of Health Economics 40 (2015) 6982
often provide income above SSDI qualication levels) and take-up
SSDI. If the majority of Medicaid take-up originates from individ-
uals pursuing SSDI, then after the two year waiting period, costs to
the Medicaid program will decrease as these individuals qualify for
coverage through Medicare. Costs to SSDI will increase, however,
as they have to provide Medicare for more participants. We leave
detailed analysis of any increases in SSDI participation to future
work.
5.3. Robustness
The basic results in Table 7 are robust to several alternative sam-
ples. We nd similar effects of Medicaid expansions on public and
private insurance coverage using the 1997 through 2008 March CPS
from the Integrated Public-Use Microdata Series (IPUMS). The CPS
is an annual, nationally representative survey that contains infor-
mation on income, family structure, and health insurance status
for the year prior to the interview. Similar to the SIPP analysis, the
CPS sample is restricted to contain individuals between the ages
of 20 and 64 who report having a work disability. We also drop all
observations from the ve states which are not uniquely identied
within the SIPP. Thus, the SIPP and March CPS samples should rep-
resent similar populations. We impute Medicaid eligibility as we
did in the SIPP analysis but adjust for annual Medicaid income lim-
its and disregards to account for the yearly observations of the CPS.
Simulated eligibility is constructed similarly except it is now at a
state and year level.
Panel A of Table 9 reports two-stage-least-squares health insur-
ance coverage estimates using the sample from the March CPS. The
models in Table 9, Panel A have the same controls on the right side
of the equation as in the SIPP analysis except they do not control for
panel-wave xed effects, which are specic to the survey structure
of the SIPP. The health coverage variables (Medicaid only, Private
only, Overlap, and Uninsurance) are dened the same as in the SIPP
analysis. 22 All estimates in Table 9, Panel A are weighted using the
State Health Access Data Assistance Centers (SHADAC) summary
health insurance weight in the March CPS.
Results in Table 9, Panel A for the CPS analysis suggest a Med-
icaid take-up rate of 25% accompanied by a private reduction rate
of 22%. These rates are lower in magnitude than what was found in
the SIPP analysis but may be partially explained by the longer recall
period of the CPS. The Medicaid take-up rate and private reduc-
tion rates reported in columns (2) and (3) of Table 9, Panel A only
account for changes in individuals reporting a single form of cov-
erage for an entire year. The yearly observations in the CPS do not
identify transitions in coverage as well as the monthly level data of
the SIPP. As a result, individuals who experience transitions from
one type of insurance to another are much more likely to appear in
the overlap group in the CPS than in the SIPP. The increased likeli-
hood of overlapping coverage potentially results in lower take-up
rates for those reporting a single form of coverage, but larger take-
up rates for those reporting two forms of coverage (i.e. private and
Medicaid) in the CPS data. Indeed, the CPS estimate of the effect
of Medicaid eligibility on overlap coverage is larger than the SIPP.
If we account for this effect by interpreting it as a transition from
22
Health coverage variables for the CPS analysis are constructed using the health
insurance summary variables produced by the State Health Access Data Assistance
Center (SHADAC) contained within the IPUMS version of the CPS. The CPS updated
their survey questionnaire in 1999 with a verication question to address the under-
reporting of health insurance. The verication question asked those who had not
reported health coverage earlier in the survey about their health coverage again and
allowed respondents a second chance to report forms of health coverage. SHADAC
then imputed health insurance variables in the CPS prior to 1999 to account for the
verication question.
77
private coverage to Medicaid, then the CPS estimates are closer to
the level of the SIPP estimates but are still slightly lower in mag-
nitude. This is consistent with Ham and Shore-Sheppard (2005)
which found that take-up rates were more similar across the CPS
and SIPP when SIPP responses were annualized to be more like CPS
data. When we account for the overlap group, the CPS estimates
suggest a crowd-out rate of 90% which is reasonably close to the
SIPP estimates of crowd-out. A 95% bootstrapped condence inter-
val, however, does not allow us to reject the null hypothesis for
CPS estimates. The effect on uninsurance in column (5) of Table 9
Panel A is larger and statistically signicant in the CPS suggesting
a 1.31% point increase for a 10% point increase in Medicaid eligi-
bility. There is no clear explanation as to why the uninsurance rate
appears to be worsening in both samples. This is a focus of future
work.
In order to qualify for Medicaid under the OBRA86 expansion, an
individual must not only have a low level of income they must also
be disabled. Though the denition of a self-reported work-limiting
disability in the SIPP is probably less stringent than what is used
by the government to determine disability status, it does identify a
group of people likely to qualify for disability services. Since work
disabilities in the SIPP are self-reported, we may be concerned that
the Medicaid expansions increased the incentive to report a disabil-
ity and resulted in a compositional shift of the sample across the
time period. I nd no evidence of an effect of eligibility on the like-
lihood of being work-disabled (results not shown) suggesting that
becoming eligible for Medicaid did not induce a person to report a
work disability.23
As a robustness test, we perform the same analysis as in Table 7
using non-work disabled individuals in the 1996, 2001, and 2004
panels of the SIPP. The results for this analysis are presented in
Table 9, Panel B. There is no signicant Medicaid take-up by non-
work disabled individuals who would qualify for Medicaid based
on their income alone (see column (2), Table 9, Panel B). In fact, the
sign on the coefcient suggests there was a slight reduction in Med-
icaid participation for working age non-work disabled individuals
though this result is not statistically signicant. Medicaid eligi-
bility also has small statistically insignicant effects on the other
forms of coverage (Private only, Overlap, and Uninsured) for the
non-work disabled. This supports the specication being properly
identied.
One concern with the panel nature of the SIPP is the possi-
bility of attrition bias. The response rates at the initial interview
of each SIPP panel are higher than later interviews. We may be
concerned that a select group of people choose not to respond or
enter the dataset in later interviews and this inuences the over-
all results. To account for this problem, we construct a baseline
sample by keeping only observations from the rst three waves
of each panel and respondents who were present for the origi-
nal interview (i.e. all entrants after the rst wave of the panel are
excluded). Table 9, Panel C displays the results from this restricted
sample. Overall, the point estimates for private only coverage and
Medicaid remain virtually unaffected though the estimates are
less precise given the decrease in sample size and the effect on
private insurance coverage loses statistical signicance. The attri-
tion sample also only contains observations from 3 years of data
from my 12 year study period which does not allow us to take
full advantage of the Medicaid eligibility expansions that vary by
23
To test this, I run a 2SLS regression of an indicator for a reported work-related
disability on an indicator for whether a respondent would be nancially eligible
for Medicaid using the same controls in my main analysis. The estimated effect of
Medicaid nancial eligibility on the likelihood of reporting disability status is 0.022
with a standard error of 0.065.
78 K.L. Wagner / Journal of Health Economics 40 (2015) 6982

Table 9
First-stage for and 2SLS estimates of insurance coverage regressions.

Robustness of results to alternative samples

First-stage 2SLS estimates

Dep. Var. Eligible Medicaid only Private only Overlap Uninsured

(1) (2) (3) (4) (5)

Panel A: work disability sample from the 19972008 March CPS

(80,727 observations)
Simulated eligibility 1.074
(0.092)***
Eligible 0.249* 0.218** 0.071* 0.131*
(0.150) (0.101) (0.043) (0.074)

Mean of dep. var. 0.55 0.29 0.39 0.04 0.14

1st stage F 135.257
(P-value on null)
R-squared 0.5310 0.2430 0.2663 0.0100 0.0340

Panel B: non-work disabled sample from 1996, 2001, and 2004 SIPP panels.
(1,290,995 observations)
Simulated eligibility 0.814
(.053)***
Eligible 0.037 0.010 0.004 0.025
(0.036) (0.047) (0.010) (0.065)

Mean Dep. Var. 0.342 0.03 0.78 0.01 0.18

1st stage F 235.62
(P-value on null)
R-squared 0.5272 0.0796 0.2415 0.0167 0.1844

Panel C: work disabled sample 1996, 2001, and 2004 SIPP panels rst three waves of panels only
(46,623 observations)
Simulated eligibility 0.977
(0.227)***
Eligible 0.422** 0.484 0.008 0.305
(0.215) (0.318) (0.107) (0.259)

Mean dep. var. 0.585 0.30 0.42 0.04 0.19

1st stage F 18.576
(P-value on null)
R-squared 0.4869 0.2615 0.2708 0.0154 0.0237

Source: Authors calculation.

Notes: Standard errors are in parentheses and are clustered at the state level.
*
Represents a 10% signicance level.
**
Represents a 5% signicance level.
***
Represents a 1% signicance level.
Estimates are from 2SLS regressions of four forms of health coverage (Medicaid only, Private Only, Both Medicaid and Private, and Uninsured) for different sample populations
using simulated eligibility as an instrument. All columns control for state and year xed effects, age, age squared, number of children in household, family size, female sex,
black race, Hispanic origin, marital status, cubic in personal income, less than high school education, and high school diploma.
Panel A: Results for this table are for the 19972008 March CPS surveys. Sample includes all individuals between the ages of 20 and 64 who report a work disability. The 5
states that are not uniquely identied in the SIPP are excluded from the CPS analysis. Results are weighted using the SHADAC constructed weight for CPS health insurance
coverage.
Panel B: Sample includes individuals between the ages of 20 and 64 in the 1996, 2001, and 2004 panels of the SIPP who do not report a work disability. All regressions are
weighted using the SIPP provided person-level weight.
Panel C: Sample includes individuals between the ages of 20 and 64 who report a work disability in the in all 3 of the rst three waves of each SIPP panel. All regressions are
weighted using the SIPP provided person-level weight.

year. Despite their lack of statistical signicance and poor preci-
sion, the point estimates of the effects on insurance coverage in
the attrition analysis remain considerably close to the estimates
of my main specication. This demonstrates that the results of
the study are not being driven by attriters or entrants in the SIPP
panels.
Another potential concern with the disabled population is its
increased likelihood of being covered through Medicare. In earlier
crowd-out studies for children, Medicare coverage was ignored
given childrens low tendency to be covered through this program.
Crowd-out literature has focused on the treatment of the overlap
population for private and Medicaid coverage. We might also be
concerned with the overlap population considering Medicare cov-
erage. That is, we might want to account for individuals who report
Medicaid and Medicare coverage or private and Medicare coverage
in addition to private and Medicaid coverage.24 The interpretation
for Medicare overlap groups will be different from the Medicaid-
private overlap group. Given that Medicaid is a fully comprehensive
plan, there is no need for Medicaid recipients to possess private cov-
erage while covered by Medicaid. This is the basis for the original
overlap assumption of a transition from private to public cover-
age. Medicare, though generous, is not completely comprehensive,
however, and individuals often hold either supplementary private
or Medicaid coverage to ll the gaps. 25
24
Very few individuals report all three forms of coverage in the same month in
the SIPP.
25
For Medicare eligible individuals, Medicare will always cover rst-order
expenses and Medicaid will cover any remaining charges.
 K.L. Wagner / Journal of Health Economics 40 (2015) 6982
Table 10
Effects of Medicaid eligibility on insurance coverage for work-disabled individuals in the 1996, 2001, and 2004 SIPP panels: adjusting insurance coverage measures to account
for Medicare (170,909 observations).
Dependent variable Private only Medicaid only Medicare only
Eligible 0.274** 0.239 0.083
(0.127) (0.160) (0.083)
Mean dep. var. 0.35 0.22 0.08
R-squared 0.2195 0.1983 0.0895
Source: Authors calculation.
Notes: Standard errors are in parentheses and are clustered at the state level.
*
Represents a 10% signicance level.
**
Represents a 5% signicance level.
***
Represents a 1% signicance level.
All regressions are weighted using the SIPP provided person-level weight. Estimates are from 2SLS regressions of various forms of health coverage on Medicaid income
eligibility using simulated eligibility as an instrument. Coverage forms are dened similar as in earlier tables, but now take into account Medicare coverage. Three represents
coverage by all three forms of insurance (Medicaid, Private, and Medicare). All columns control for panel-wave, state and year xed effects, a state-linear trend, age, age
squared, a cubic in income, number of children in household, family size, female sex, black race, Hispanic origin, marital status, less than high school education, and high
school diploma.
Table 10 presents the results accounting for these additional
Medicare overlap groups. Columns (1) and (2) of Table 10 present
the results for private only (now no Medicaid or Medicare) and
Medicaid only (now no private or Medicare) coverage. The results
in column (2) of Table 10 suggests the increase in Medicaid cov-
erage with no private or Medicare coverage is 2.39% points for
every 10% point increase in eligibility. This is roughly equivalent
to the decrease in private coverage with no Medicaid or Medicare
coverage. Considering only these two groups, crowd-out remains
large and consistent with the earlier specication (115%). Turning
to the Medicare overlap groups, there is a 1.78% point decrease in
Private-Medicare coverage for every 10% point increase in Medic-
aid eligibility. This is slightly larger in magnitude than the increase
in Medicaid-Medicare coverage which increases by 1.67% points
for a 10% point increase in Medicaid eligibility. If we consider only
these two coverage forms, the estimates imply a crowd-out rate of
107%. Bootstrapped condence intervals for crowd-out rates with
and without Medicare coverage are wide and do not allow us to
rule out zero crowd-out. Given the similarity in the magnitudes of
the opposite signed coefcients of Medicaid only and private only
and likewise Medicaid-Medicare and private-Medicare, however,
the results suggest that individuals are choosing to take-up Med-
icaid coverage over private coverage either as an individual policy
or as a supplementary plan for Medicare.26
6. Conclusion
Despite the large cost and important Medicaid eligibility
changes affecting the working age disabled population, little liter-
ature has examined the effects of Medicaid eligibility on insurance
coverage for this group. In this paper, I ll this void by investi-
gating a Medicaid eligibility expansion which allowed states the
option to enroll disabled individuals with incomes up to 100% of
the federal poverty level in Medicaid. Estimates of crowd-out range
from 50 to 100% (OLS and 2SLS specications, respectively). The
2SLS estimates suggest that all new Medicaid enrollees through the
expansions dropped a private plan in favor of public coverage. The
condence interval around the 2SLS point estimate is wide, how-
ever, and should be interpreted cautiously, I nd a similarly large
point estimate of crowd-out using an alternative sample derived
from the March CPS further supporting that crowd-out among the
26
If we interpret the overlap group between all three types of coverage as
a transition to Medicaid Medicare-supplementary plans from Private Medicare-
Supplementary plans, then crowd-out measures remain similar (130%).
79
Medicaid-Medicare Private-Medicaid Private-Medicare All three
0.167* 0.024 0.178 0.039
(0.080) (0.051) (0.121) (0.038)
0.07 0.03 0.07 0.02
0.0370 0.0100 0.0824 0.0013
disabled population was substantial. Using population and disabil-
ity rate information from 2008 (the year after the end of my study
period), I estimate that these expansions cost at most between 2.92
and 3.03 billion dollars.27 We rely on the average cost of a disabled
Medicaid recipient to construct these estimates, but it is likely that
the newly eligible are not as expensive at the margin as the aver-
age disabled Medicaid recipient. Thus, we should think of 3 billion
dollars as an upper bound of the cost of the expansions.
The analysis of these expansions is especially relevant given the
enactment of the Medicaid expansions under the Patient Protection
and Affordable Care Act in 2014. For states who accept the cur-
rent terms of the PPACA, Medicaid eligibility levels will increase
to include individuals under the age of 65 who have incomes
less than 138% of the federal poverty level. In 2008, the average
income eligibility level for the disabled was 87% of the federal
poverty level across all states. Given the expansions imposed by
the PPACA, we can expect more disabled individuals to become eli-
gible for Medicaid in 2014. Disabled individuals qualifying through
the overall Medicaid expansions of the PPACA are likely healthier
on the margin than individuals qualifying through SSI. Given this,
my estimates of total costs represent an upper bound, but are still
informative as to the potential costs of the Medicaid expansion.
It is important to note that though I nd the potential for s-
cal consequences from these expansions, the results say nothing
about any potential health benets gained through the increased
Medicaid take-up. Even though individuals are less reliant on pri-
vate health insurance, they may experience benets to health from
receiving public coverage. First, private health insurance plans may
not cover all services while Medicaid coverage provides for most
procedures suggesting that individuals could gain access to medi-
cal techniques previously unavailable to them and improve overall
physical health. Second, since Medicaid is a low cost (if not free)
policy with virtually no copays or deductibles, individuals may
benet from having additional resources available for consump-
tion. For many of these individuals, coverage through the Medicaid
program will greatly lessen nancial strain and can have an effect
on mental health or overall happiness (Finkelstein et al., 2012).
Another potential benet of disabled Medicaid coverage is that
27
In 2008, approximately 15,000,000 individuals between the ages of 20 and 64
reported having a working limiting disability. Eligibility increased by roughly 3 per-
centage points at the peak of the expansions implying that close to 450,000 disabled
people became eligible for Medicaid coverage. Using the estimated take-up rates
(between 41 and 42%) and an average cost per disabled beneciary of $15,840 (in
FY 2008) this suggests that the expansions cost 450,000*0.41*15,840 = $2.92 billion.
80 K.L. Wagner / Journal of Health Economics 40 (2015) 6982
sicker individuals will leave employer group plans and result in
smaller or slower growing insurance premiums for the privately
insured individuals who remain on the plan.
There are some potential health detriments to Medicaid cover-
age, however. Due to low reimbursement rates, not all physicians
will accept Medicaid coverage and beneciaries are restricted to a
limited group of health providers and may experience access prob-
lems. For adults with public health insurance, healthcare access is
slightly worse than patients with private insurance (Government
Accounting Ofce, 2012). There is also some evidence that the qual-
ity of care is worse for patients with public insurance compared
with private insurance, but this difference is small (Weissman et al.,
2013). Given these potential health benets and detriments it is
important to examine health outcomes in conjunction with the
scal consequences. We leave the analysis of health outcomes to
future work.
Acknowledgements
I would like to thank William Evans, Daniel Hungerman, James
Sullivan, Ethan Lieber, Thomas DeLeire, Mark Duggan, the co-
editor, two anonymous referees, and seminar participants at the
University of Notre Dame.
Appendix A. Pathways for disabled Medicaid eligibility
A.1. Categorical eligibility SSI recipients
The Supplemental Security Income (SSI) Program is run by the
Social Security Administration (SSA) and provides a cash supple-
ment to the aged and disabled who have few nancial resources.
Individuals are eligible to receive SSI if their countable income and
assets are less than the income and resource eligibility levels deter-
mined by the SSA.28 In 2012, the SSI income and resource eligibility
levels for an individual were 75% of the federal poverty level (FPL)
and $2000 respectively (Social Security Administration, 2012).
States are required to offer Medicaid to all SSI recipients unless
they use a 209(b) option that allows states to employ stricter eli-
gibility levels than SSI requirements. However, under this option,
the eligibility levels cannot be any more restrictive than Medicaid
eligibility levels used by the state in 1972 when SSI was created.
Eleven states used the 209(b) option in 2009.29 Each 209(b) state
has at least one aspect of Medicaid eligibility that is more restrictive
than federal SSI standards. The most obvious method of imposing
stricter standards is through lower income and resource eligibility
levels though some states choose stricter denitions of disability or
countable income/assets. Standards for SSI have changed little over
the time period considered in this paper. SSI/categorical eligibility
is the only mandatory Medicaid pathway for the aged and disabled
that all states must offer. All other pathways are at the option of
the state.
A.2. State supplemental payments (SSP)
States may choose to offer supplementary payments to SSI
recipients and also to those who do not qualify for SSI but have
28
Countable income is the applicants total income minus certain monetary disre-
gards. There is a federally mandated disregard of $20 dollars though some states are
more generous. In addition, $65 of earned income plus half of remaining earnings
are disregarded when constructing countable income. See the SSAs Under-
standing Supplemental Security Income website at http://www.ssa.gov/ssi/text-
understanding-ssi.htm for more information.
29
These states were Connecticut, Hawaii, Illinois, Indiana, Minnesota, Missouri,
New Hampshire, North Dakota, Ohio, Oklahoma, and Virginia.
lower levels of income. The state supplemental payment option
allows states to increase their cash assistance levels for the elderly
and disabled above the federal standard. States choose the size of
their supplemental payments effectively creating an income oor
for eligible state residents. States have complete control over SSP
eligibility levels and can vary them by living arrangement, type
of recipient (aged or disabled), and by regional differences (typi-
cally differences in the cost of living) within a state (Stone, 2002).
Throughout my analysis, I use SSP levels for an individual living
independently since the data only contain information for non-
institutionalized individuals. The income eligibility level for the
SSP pathway is roughly equal to the maximum allowance offered
through SSI plus the additional SSP payments. All states, however,
must use the SSI resource levels unless they are a 209(b) state.30
Individuals who receive SSP from the state but do not qualify
for SSI are referred to as SSP-only recipients and are not eligible
for Medicaid through the SSI pathway.31 States have the option
to extend Medicaid coverage to SSP-only individuals and most
states offering SSP pursue this option. In 2001, twenty-ve states
offered SSP-only Medicaid coverage to the qualifying individuals
living independently (Bruen et al., 2003). Most states offer a supple-
ment that raises an individuals income and subsequent Medicaid
eligibility level just above SSI levels. Some states, however, offer
large payments that can effectively raise the SSP-only Medicaid
income eligibility level above the federal poverty level (Bruen et al.,
2003).32 SSP policies have also varied little over the time period
considered in this paper.
A.3. Medically needy
A medically needy program allows individuals with high income
but large amounts of medical expenses to qualify for Medicaid cov-
erage. States with medically needy programs allow residents to
subtract their medical expenses from their countable income and
essentially spend-down to Medicaid eligibility. That is, for states
with medically needy programs there is effectively no income limit
on Medicaid applicants as long as an applicant incurs enough med-
ical costs to reach the medically needy income limit. 209(b) states
who do not have a medically needy program must allow individuals
to spend down to eligibility. Though states may choose to offer less
generous benets packages to medically needy individuals, medi-
cally needy packages are essentially the same as those offered for
income eligible recipients.
States can choose medically needy income and resource levels
unique from other pathways for eligibility. Most states choose to
set their medically needy levels well below poverty and sometimes
even below SSI standards. To qualify under the medically needy
pathway, applicants get a certain time period (16 months depend-
ing on the state) to reduce their income by their medical expenses.
If they reach the medically needy income limit after the deduction
of their medical expenses then they qualify for Medicaid coverage
for the rest of the period. Except for four states (Texas and South
Carolina eliminated their medically needy programs in 1996 and
1993 respectively. Oklahoma and Oregon both ended their medi-
cally needy programs in 2003), all states that had a medically needy
program in 1991 also had them in 2008. In 2008, 33 states had a
medically needy program.
30
In which case they use the resource level used by the 209(b) state.
31
SSP-only individuals are those who have incomes higher than SSI levels, but
lower than the income oor set by the state.
32
States with the maximum SSI/SSP for the independently living aged and disabled
above 100% of the FPL in 2001 were Connecticut and Alaska.
 K.L. Wagner / Journal of Health Economics 40 (2015) 6982
A.4. Medicaid buy-in for the working disabled
Though Medicaid does disregard certain amounts of earned
income when determining eligibility, many working disabled have
too much income to qualify for Medicaid coverage. This pathway
allows disabled individuals to work and earn up to 250% of the
federal poverty level. Unlike other Medicaid recipients, however,
workers qualifying under this pathway have to pay a premium and
buy-in to the Medicaid program. Premiums for the Medicaid buy-in
tend to be much smaller than those available on the private market.
A.5. Medicare cost sharing dual eligibles: QMBs, SLMBs, and
QIs
Medicaid always acts as the payer of last resort allowing all other
sources of payment to be exhausted before coverage. Disabled indi-
viduals who have been covered under Social Security Disability
Insurance for 24 months qualify for coverage by Medicare. Medi-
care, however, is not a full coverage health insurance plan though
its coverage is quite extensive. There are gaps within Medicare ben-
ets (such as long-term care expenses and prescription drug costs)
and it also requires premiums and copays that many low income
individuals may not be able to afford.
To prevent Medicare eligible individuals from dropping their
Medicare policies in favor of less expensive Medicaid coverage,
Medicaid began offering coverage of Medicare premiums and
copayments for individuals who qualify for both Medicare and
Medicaid (dual-eligibles). For dual-eligibles, Medicaid will only
cover what Medicare does not, allowing Medicare to cover many
of the rst order expenses.
There are three forms of dual-eligibles that qualify for differ-
ent provisions through Medicaid. A Qualied Medicare Beneciary
(QMB) was established under the Medicare Catastrophic Cover-
age Act of 1988. QMBs receive Medicaid coverage of Medicare
Part B premiums, deductibles, and copayments if they qualify for
Medicare Part A and have incomes less than 100% of the fed-
eral poverty level. Specied Low-Income Medicare Beneciaries
(SLMBs) were established in 1993 and qualify for Medicare Part
B premiums if they qualify for Medicare Part A and have incomes
above 100% but less than 110% of the federal poverty level. The
SLMB income limit was raised in 1995 to 120% of the FPL. The last
of the dual-eligible types is a Qualifying Individual-1 (QI-1) who
has incomes above 120% but below 135% of the federal poverty
level. Medicaid will pay the Part B premiums for QI-1s. All Medicare
cost sharing participants must have resources less than twice SSI
levels.
References
Aizer, A., Grogger, J., 2003. Parental Medicaid expansions and child Medicaid cover-
age. In: NBER Working Paper # 9907.
Baicker, K., Finkelstein, A., Song, J., Taubman, S., 2013. The impact of Medicaid on
labor force activity and program participation: evidence from the Oregon Health
Insurance Experiment. NBER Working Paper # 19547.
Baumrucker, E., 2010. Medicaid: The Federal Medical Assistance Percentage (FMAP).
Congressional Research Service, Washington, DC.
Blumberg, L.J., Dubay, L., Norton, S.A., 2000. Did the Medicaid expansions for children
displace private insurance? An analysis using the SIPP. J Health Econo. 19 (1),
3360.
Brown, J.R., Coe, N., Finkelstein, A., 2007. Medicaid crowd-out of private long-term
care insurance demand: evidence from the Health and Retirement Survey. Tax
Policy Econ. 21, 134.
Bruen, B.K., Wiener, J.M., Kim, J., Miazad, O., 1999. State usage of Medicaid cover-
age options for aged blind, and disabled people. Assessing the New Federalism
Discussion Papers.
Bruen, B.K., Wiener, J.M., Thomas, S., 2003. Medicaid eligibility policy for aged, blind,
and disabled beneciaries. The Urban Institute, Washington, DC.
Buettgens, M., 2011. Health Insurance Policy Simulation Model (HIPSM) Methodol-
ogy Documentation. The Urban Institute, Washington, DC.
81
Burkhauser, R.V., Houtenville, A.J., Wittenburg, D.C., 2003. A users guide to cur-
rent statistics on the employment of people with disabilities. In: Stapleton, D.C.,
Burkhauser, R.V. (Eds.), The Decline in the Employment of People with Disabil-
ities: A Policy Puzzle. Upjohn Institute for Employment Research, Kalamazoo,
MI, pp. 2386.
Card, D., Shore-Sheppard, L., 2004. Using discontinuous eligibility rules to identify
the effects of the federal Medicaid expansions on low income children. Rev.
Econ. Stat. 86 (3), 752766.
Coe, N.B., (MIT dissertation) 2005. Financing nursing home care: new evidence on
spend-down behavior.
Congressional Research Service, 1993. Medicaid Source Book: Background Data and
Analysis, Washington, DC.
Currie, J., Gruber, J., 1996a. Health insurance eligibility, utilization of medical care,
and child health. Q. J. Econ. 111, 431466.
Currie, J., Gruber, J., 1996b. Saving babies: the efcacy and cost of recent changes
in the Medicaid eligibility of pregnant women. J. Pol. Econ. 104 (6), 1263
1296.
Cutler, D., Gruber, J., 1996. Does public health insurance crowd-out private insur-
ance? Q. J. Econ. 111, 391430.
De Nardi, M., French, E., Jones, J.B., Gooptu, A., 2011. Medicaid and the elderly. NBER
Working Paper # 17689.
Dubay, L.C., Kenney, G., 1996. Revisiting the issues: the effects of Medicaid
expansions on insurance coverage of children. Future Child. 6 (1), 152
161.
Dubay, L.C., Kenney, G., 1997. Did Medicaid expansions for pregnant women crowd-
out private insurance? Health Aff. (Millwood) 16 (1), 185193.
Families USA, 2001. Could your state do more to expand Medicaid for seniors
and adults with disabilities? Expanding Medicaid: State options. Washington,
DC.
Finkelstein, A., Taubman, S., Wright, B., Bernstein, M., Gruber, J., Newhouse, J.P.,
Allen, H., Baicker, K., Oregon Health Study Group, 2012. The Oregon health
insurance experiment: evidence from the rst year. Q. J. Econ. 127 (3), 1057
1106.
Garthwaite, C., Gross, T., Notowidigdo, M.J., 2013. Public health insurance, labor
supply, and employment lock. NBER Working Paper No. 19220.
Gruber, J., Simon, K., 2008. Crowd-out 10 years later: have recent public insur-
ance expansions crowded out private health insurance? J. Health Econ. 27, 201
217.
Ham, J., Shore-Sheppard, L., 2005. The effect of Medicaid expansions for low-income
children on Medicaid participation and private insurance coverage: evidence
from the SIPP. J. Public Econ. 89, 5783.
Hamersma, S., Kim, M., 2013. Participation and crowd out: assessing the effects of
parental Medicaid expansions. J. Health Econ. 32, 160171.
Holahan, J., Buettgens, M., Carroll, C., Dorn, S., 2012. The cost and coverage impli-
cations of the ACA Medicaid expansion: National and state-by-state analysis
Executive Summary. Kaiser Commission on Medicaid and the Uninsured,
Retrieved from: http://kaiserfamilyfoundation.les.wordpress.com/2013/01/
8384 es.pdf (29.08.13).
Horvath, J., 1997. Medicaid Financial Eligibility for Aged, Blind, and Disabled:
Survey of State Use of Selected Options. National Academy for State Health
Policy.
Hudson, J., Selden, T., Banthin, J., 2005. Impact of SCHIP on insurance coverage of
children. Inquiry 42 (3), 232254.
Kaiser Family Foundation, 2012. Quick take: who benets from the ACA Med-
icaid expansion?, Retrieved from http://kff.org/health-reform/fact-sheet/who-
benets-from-the-aca-medicaid-expansion/ (22.08.2013).
Kaiser Family Foundation, 2014. Status of state action on the Medicaid expansion
decision, Retrieved from http://kff.org/health-reform/state-indicator/state-
activity-around-expanding-medicaid-under-the-affordable-care-act/2014
(5.06.14).
Kaiser Family Foundation: statehealthfacts.org. Medicaid & CHIP data.
Kaiser Commission on Medicaid and the Uninsured, 2010. Medicaid nancial eli-
gibility: primary pathways for the elderly and people with disabilities. Kaiser
Family Foundation.
Kassner, E., Shirley, L., 2000. Medicaid nancial eligibility for older people: state
variations in access to home and community-based waiver and nursing home
services. Public Policy Inst. AARP, Washington, DC.
LoSasso, A., Buchmueller, T., 2004. The effect of the state childrens health
insurance program on health insurance coverage. J. Health Econ. 23, 1059
1082.
Mississippi Division of Medicaid, 19912008. Annual Report.
Shore-Sheppard, L.D., 2008. Stemming the tide? The effect of expanding Medicaid
eligibility on health insurance. The B.E. Journal of Economic Analysis & Policy 8
(2), Article 6.
Social Security Administration, 2012. Understanding Supplemental Security
Income, Retrieved from http://www.ssa.gov/ssi/text-understanding-ssi.htm
(27.07.12).
Social Security Administration, 19912008. State assistance programs for SSI recip-
ients. SSA Publication No. 13-11975.
Stone, J.L., 2002. Medicaid: eligibility for the aged and disabled. Report for Congress.
Congressional Research Service.
Stone, J.L., 2011. Medicaid eligibility for persons age 65+and individuals with dis-
abilities: 2009 state proles. Congressional Research Service.
Thorpe, K., Florence, C., 1998. Health insurance coverage among children: the role
of expanded Medicaid coverage. Inquiry 35 (4), 369379.
82 K.L. Wagner / Journal of Health Economics 40 (2015) 6982

United States Census Bureau, 2012. Disability: CPS Annual Social and Economics
Supplement (CPS ASEC), Retrieved from http://www.census.gov/people/
disability/methodology/cps.html (01.21.14).
United States Census Bureau, 1996. Survey of Income and Program Participation
1996 Panel Wave 1 Core Microdata File.
Weissman, J.S., Vogeli, C., Levy, D.E., 2013. The quality of hospital care for Medicaid
and private pay patients. Med. Care 51 (5), 389395.
Yazici, E., Kaestner, R., 2000. Medicaid expansions and the crowding out of private
health insurance among children. Inquiry 37 (Spring 1), 2332.
 Journal of Health Economics 40 (2015) 8396

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

Aggregation and the estimated effects of economic conditions on

health
Jason M. Lindo
Texas A&M University, NBER, and IZA, Germany

a r t i c l e i n f o a b s t r a c t

Article history: This paper considers the relationship between economic conditions and health with a focus on dif-
Received 6 January 2014 ferent approaches to geographic aggregation. After reviewing the tradeoffs associated with more- and
Received in revised form 6 November 2014 less-disaggregated analyses, I update earlier state-level analyses of mortality and infant health and then
Accepted 26 November 2014
consider how the estimated effects vary when the analysis is conducted at differing levels of geographic
Available online 25 December 2014
aggregation. This analysis reveals that the results are sensitive to the level of geographic aggregation with
more-disaggregated analysesparticularly county-level analysesroutinely producing estimates that are
JEL classication:
smaller in magnitude. Further analyses suggest this is due to spillover effects of economic conditions on
I10
J20
health outcomes across counties.
E32 2014 Elsevier B.V. All rights reserved.

Keywords:
Health
Recessions
Mortality
Infant health
Aggregation

1. Introduction which health outcomes respond to changes in macroeconomic con-

Although Harvey Brenners pioneering research suggested that
health deteriorates during recessions (Brenner, 1973, 1975, 1979),
follow-up work has revealed that estimates based on aggre-
gate time-series data are quite fragile (Forbes and McGregor,
1984; McAvinchey, 1988; Joyce and Mocan, 1993; Laporte, 2004;
Gerdtham and Johannesson, 2005). As Ruhm (2000) points out, this
fragility is not surprising since any lengthy time-series is likely
to suffer from substantial omitted variables bias.1 Out of con-
cern for such biases, researchers have largely stopped considering
nationwide changes in favor of an area approach that considers
how the health of individuals living in an area changes over and
above changes occurring across all areas when its economic condi-
tions change over and above changes occurring across all areas. The
intent, however, has remained the same: to estimate the degree to
Correspondence to: Department of Economics, Texas A&M University, College
Station, TX 77843, United States.
E-mail address: jlindo@econmail.tamu.edu
1
For example, it is problematic for this approach that penicillin became increas-
ingly available as the United States began to recover from the Great Depression.
ditions. These studies have repeatedly concluded that recessions
are good for health in developed countries, though this interpreta-
tion relies on the assumption that health is similarly inuenced by
macroeconomic conditions (broadly dened) and more local area-
specic economic conditions.2 This study is motivated by the idea
that this assumption cannot be tested directly since credible esti-
mates of the effects of macroeconomic conditions (broadly dened)
would appear to be out of our reach, but that it can be tested indi-
rectly by investigating the extent to which more local and less local
economic conditions have different estimated effects on health.
More generally, this paper is concerned with the way in which geo-
graphic aggregation (or disaggregation) inuences the conclusions
we draw from our analyses and what analyses we can perform.3
2
See Ruhm (2000, 2003, 2005, 2007), Ruhm (2013), Dehejia and Lleras-Muney
(2004), Johansson (2004), Neumayer (2004), Tapia Granados (2005), Gerdtham and
Ruhm (2006), Lin (2009), Miller et al. (2009) and Stevens et al. (2011).
3
With few exceptions nearly all area studies using U.S. data dene areas as
states. To my knowledge, the only exceptions are Currie and Tekin (2011) who con-
sider foreclosures at the zip-code level in four states and Dehejia and Lleras-Muney
(2004) who consider unemployment rates and supplement their state-level analysis
with a county-level analysis of California.

http://dx.doi.org/10.1016/j.jhealeco.2014.11.009
0167-6296/ 2014 Elsevier B.V. All rights reserved.
84 J.M. Lindo / Journal of Health Economics 40 (2015) 8396
There are several issues to consider when choosing the level of
geographic aggregation and interpreting the results of the analysis.
From an economic perspective, we must recognize that there are
many mechanisms through which economic conditions can affect
health and, from a statistical perspective, we must recognize that
the level of geographic aggregation inuences the degree to which
the estimates capture these different mechanisms. For example, we
would expect the effects of individuals job losses to be captured
by changes in economic conditions in the local area where individ-
uals work. However, economic conditions both near and far may
affect an individuals health through impacts on re-employment,
migration decisions, perceptions about economic conditions, trafc
congestion, levels of pollution, the quality of medical care, gov-
ernment policies, and through effects on the members of ones
social network. In terms of identication, it is important to keep
in mind that the estimated effects of an areas economic con-
ditions are fully inclusive of spillover effects across subareas
within the area whereas the estimated effects of subarea eco-
nomic conditions are not. For example, estimated effects of state
economic conditions are fully inclusive of spillover effects across
counties within a state whereas more-disaggregated analyses are
not.
At the same time, more-disaggregated analyses can offer more
precise estimates because they use variation in economic condi-
tions idiosyncratic to the area in addition to variation driven by
broader changes. In particular, more-disaggregated analyses can
improve power by leveraging variation in economic conditions
that are masked in more-aggregate measures. For example, a con-
traction in one part of a state that is offset by an expansion in
another part of the state would contribute to county-level esti-
mates but not to state-level estimates. Similarly, a county-level
analysis would exploit variation in the severity of contractions (and
expansions) across different parts of the state whereas a state-
level analysis would not. Thus, more-disaggregated analyses may
be able to detect statistically signicant effects of economic condi-
tions where less-aggregated conditions cannot, even if they do not
capture some of the spillover effects captured in more-aggregated
analyses.4
Another important consideration is that economic indicators
are subject to measurement error, which is especially problem-
atic for xed-effects estimators (Griliches, 1977; Griliches and
Hausman, 1986; Hausman, 2001). The most common measure
of economic conditions used in this literature is the unemploy-
ment rate, produced by the Bureau of Labor Statistics (BLS).
However, as one considers smaller areas, one needs to be more
and more concerned about measurement error in unemploy-
ment rates since they are based in part on household surveys
(Bartik, 1996; Hoynes, 2000).5 For this reason, employment-to-
population ratios would seem preferable because they are based
solely on administrative data. Still, one may have concerns about
measurement-error bias that may be inuenced by the level of
aggregation.
4
On a related note, more-disaggregated analyses can have improved power
because they allow for a richer set of control variables in a manner that reduces
the amount of unexplained variation in outcomes. This will not necessarily be the
case, however, because smaller areas may have more variation in outcomes overall.
The richer set of control variables, of course, may also help to mitigate concerns
about omitted variable bias.
5 6
Angrist and Krueger (1999) provide intuition: errors tend to average out in
aggregate data. Another problematic aspect of unemployment rate data is that
the BLSs substate estimates prior to 1990 are no longer considered ofcial BLS
data because they have not been revised to be consistent with the BLSs current
estimation procedure.
It is also important to consider the fact that migration is
inuenced by economic conditions (Blanchard and Katz, 1992;
Saks and Wozniak, 2011), particularly among highly educated
(Bound and Holzer, 2000; Glaeser and Gyourko, 2005; Wozniak,
2010; Notowidigdo, 2011) and healthy individuals (Halliday, 2007).
This heterogeneity implies that the estimated improvements in
health associated with recessions will understate the true improve-
ments in health if standard demographic controls do not fully
capture these sorts of compositional changes. That said, it is
unknown whether education and other characteristics associated
with health are more- or less-strongly related to the economic-
conditions-migration relationship when one considers different
types of moves and different measures of economic conditions.
As such, it is unclear whether this sort of composition bias is
likely to be of greater or lesser concern for more-disaggregated
analyses.6
Given the complexity of these issues, I take as my starting
point that it is not at all clear what level of geographic aggre-
gation is preferred but that the tradeoffs deserve consideration
and that much can be learned by comparing the results of alter-
native approaches. As such, after describing the different ways
that I dene areas throughout the subsequent sections in Sec-
tion 2, I then replicate and update earlier state-level estimates
of the relationship between economic conditions and mortal-
ity (Ruhm, 2000; Stevens et al., 2011) and the relationship
between economic conditions at the time of conception and infant
health (Dehejia and Lleras-Muney, 2004) in Section 3. I then con-
sider how and why estimated effects vary when the analysis is
conducted at differing levels of geographic aggregation for mor-
tality and infant health in Sections 4 and 5, respectively, before
concluding.
My main ndings are as follows:
1. The estimated links between economic conditions and health
outcomes are sensitive to the level of geographic aggregation
with more-disaggregated analysesparticularly county-level
analysesroutinely producing estimates that are smaller in
magnitude.
2. Analyses that simultaneously consider the economic conditions
of a county and the economic conditions of surrounding areas
reveal signicant spillover effects on health outcomes. For exam-
ple, the economic conditions outside a county in the same state
have economically and statistically signicant effects on mortal-
ity. These estimates offer an explanation for why the estimated
effects of state (or economic area or region) economic conditions
are larger than the estimated effects of county economic con-
ditions when each is considered alonethe estimated effects of
state economic conditions are inclusive of spillover effects across
counties within a state, which are quite important.
3. Because they have more power, more-disaggregated analyses
have the potential to reveal statistically signicant links between
economic conditions and health outcomes even when the esti-
mated effects are smaller in magnitude. For example, while
state-level estimates using recent years of data suggest that the
link between mortality and economic conditions may no longer
exist (Ruhm, 2013), more-disaggregated analyses indicate that
the relationship remains highly signicant at conventional lev-
els.
The systematic outmigration that occurs when an areas unemployment rises
also highlights the importance of well measured population denominators in calcu-
lating mortality ratespopulation measures that do not account for the systematic
outmigration caused by economic downturns will lead to mechanical reductions in
mortality rates.
 J.M. Lindo / Journal of Health Economics 40 (2015) 8396
4. Two-way clustering (Cameron et al., 2011) on areas and years
tends to yield larger standard-error estimates than the typical
approach that clusters only on areasparticularly for county-
level analysessuggesting that it is important to allow shocks
to be correlated within areas over time and to be correlated
across all areas within year in estimating the effects of economic
conditions on health outcomes.
2. Dening areas
Throughout my analyses, I separately identify the effects of eco-
nomic conditions using data at ve different levels of geographic
aggregation, from counties to regions. Except where replicating
earlier research, all of my analyses are based on data that are avail-
able at the county level to ensure that all estimates are based on
the same underlying population. In particular, I do not use data
from: Alaska or Hawaii where county population data (described
in greater detail below) are unavailable prior to 1990; counties
without at least one person of each age (085) in 1990 so that
age-adjusted mortality rates can be calculated at the county level;
and Virginia where there have been substantial changes to county
denitions over time.7 Section 3 demonstrates the trivial impact
these sample restrictions have on the estimated effects of economic
conditions on health.
My most-aggregated analyses use region-level data based on
the eight regions dened by the Bureau of Economic Analysis (BEA),
which were developed in the 1950s with the intention of grouping
states with similar economic and social factors.8 My next most-
aggregated analyses use state-level data, followed by analyses that
use data at the BEA-Economic-Area level and analyses that use data
at the BEA-Component-Economic-Areas level.9 Last, as mentioned
above, I conduct analyses based on data at the county level. Natu-
rally, the sample restrictions described above reduce the number of
states, Economic Areas, Component Economic Areas, and counties
below the number that exist.
3. Methodology and updates to earlier research
Before presenting estimates of the link between economic con-
ditions and health outcomes using different denitions of areas in
the next section. In this section, I present the estimated effects from
prior state-level studies, adapt the estimation strategies in a way
that allows for more-disaggregated analyses, consider how these
7
In an additional effort to achieve consistency in the data over time, for all years I
combine the counties of La Paz and Yuma (Arizona) which split in 1983; Washabaugh
and Jackson (South Dakota) which merged in 1983; and Adams, Boulder, Jefferson,
and Weld (Colorado) from which Broomeld split off in 2001.
8
In contrast, the ve regions used in the United States census were developed
in the 1880s with similar intentions. The BEA region groupings are as follows.
New England: Connecticut, Maine, Massachusetts, New Hampshire, Rhode Island,
Vermont. Mideast: Delaware, District of Columbia, Maryland, New Jersey, New
York, Pennsylvania. Great Lakes: Illinois, Indiana, Michigan, Ohio, Wisconsin. Plains:
Iowa, Kansas, Minnesota, Missouri, Nebraska, North Dakota, South Dakota. South-
east: Alabama, Arkansas, Florida, Georgia, Kentucky, Louisiana, Mississippi, North
Carolina, South Carolina, Tennessee, Virginia, West Virginia. Southwest: Arizona,
New Mexico, Oklahoma, Texas. Rocky Mountain: Colorado, Idaho, Montana, Utah,
Wyoming. Far West: Alaska, California, Hawaii, Nevada, Oregon, Washington.
9
BEA Economic Areas dene the relevant regional markets surrounding
metropolitan or micropolitan statistical areas and are constructed such that all
counties are a part of some BEA Economic Area. Component Economic Areas are
much the same as Economic Areas except they can be based on fewer counties and
fewer residents. For example, a Component Economic Area would be considered too
small to form an Economic Area if it consisted of fewer than three counties and fewer
than 500,000 employed residents. Also note that 167 of 344 Component Economic
Areas are sufciently large that they also serve as independent Economic Areas; the
remaining 177 combine to create the remaining 12 Economic Areas.
85
changes alter the estimates, and then extend the analyses to include
additional years of data. I begin by focusing on mortality outcomes
and then turn to infant health outcomes.
3.1. Mortality
In his highly inuential study, Ruhm (2000) considers the effects
of macroeconomic conditions on health using mortality rates from
Vital Statistics of the United States publications and estimates of
the unemployment rate from the BLS. In order to focus on within-
area variation, Ruhms estimates are based on the regression
equation
Hat = Eat + Xat + t + a + a t + at , (1)
where Hat is a measure of health for individuals living in area a in
year t, Eat is a measure of economic conditions, Xat is a vector of
time-varying controls, t are year xed effects,  a are area xed
effects, and  a t are area-specic time trends. To operationalize this
identication strategy, Ruhm denes a as a state, Hat as the natural
log of the mortality rate (deaths per 100,000), Eat as the unem-
ployment rate, and Xat includes the fraction of the state population
that: is less than ve years old, is 564 years old, is greater than 64
years old, is black, is hispanic, is a high school dropout, has some
college, is a college graduate.10 In Column 1 of Table 1, I display
the estimate based on these data and denitions while clustering
the standard-error estimate on the state.11 The point estimate indi-
cates that a one-percentage-point increase in the unemployment
rate is associated with a 0.54-percent decrease in overall mortality.
As the starting point for my analysis, I begin with the approach
used in Stevens et al. (2011) who extend Ruhm (2000) in sev-
eral ways. First, they create mortality rates using death counts
from Vital Statistics micro-record multiple-cause-of-death les
and population data from SEER (described above). Second, they
pool monthly Current Population Survey data in order to construct
overall unemployment rates and unemployment rates for separate
demographic groups, which requires them to focus on data from
1978 to 2006. Third, they include a richer set of controls for the
fraction of a states population in various age brackets (less than 5,
517, 1830, greater than 64) using the aforementioned population
data, construct the fraction black with the aforementioned popu-
lation data, and use CPS data to measure the fraction Hispanic and
the fraction in each education group.12 Last, they replace the natu-
ral log of the mortality rate with the natural log of the age-adjusted
mortality rate to account for uneven changes in the age distribution
across states. The age-adjusted mortality rate is calculated by tak-
ing the weighted average of the mortality rate for each age, using
as weights the fraction of individuals in each age category in the US
in 1990. After making these changes, their estimate (shown in Col-
umn 2 of Table 1) indicates that a one-percentage-point increase in
the unemployment rate is associated with a 0.33-percent decrease
in overall mortality.
In columns 3 through 9 of Table 1, I progressively make changes
to this estimation strategy, ultimately arriving at an approach that
can also be used in more-disaggregated county-level analyses.
These changes are as follows:
10
These control variables are calculated by interpolating between decennial cen-
suses. Additionally, estimates are weighted by the population in each state.
11
As Ruhm (2000) did not present estimates that clustered standard errors, this
estimate is taken from Stevens et al. (2011) who replicate and extend Ruhms anal-
ysis.
12
The fraction in each educational group is based solely off of individuals who are
at least 25 years old.
86 J.M. Lindo / Journal of Health Economics 40 (2015) 8396
Table 1
State-level estimates of the effects of economic conditions on overall mortality (per 100,000 residents).
Prior estimates New estimates
Ruhm Stevens et al. Using BLS Using controls
(2000) (2011) unemployment available for
data counties
Years analyzed 19721991 19782006 19782006 19782006
(1) (2) (3) (4)
Unemployment 0.0054*** 0.0033*** 0.0039*** 0.0040***
rate (0.0010) (0.0010) (0.0010) (0.0010)
Emp-to-Pop
ratio
Observations 930 1479 1479 1479
Notes: See Section 3.1 for details on the data and measures used in each column. All regressions include year xed effects, state xed effects, and state-specic trends. In
addition, all regressions are weighted by the population counts in each cell and cluster standard-error estimates on the state.
* Signicant at 10%; ** signicant at 5%; *** signicant at 1%.
Use state-level unemployment rates published by the BLS for
convenience. This change causes the estimated semi-elasticity
to change from 0.0033 to 0.0039 (Column 3).
Omit controls for the fraction Hispanic and the fraction in each
education category, which cannot be reliably measured for small
counties. This change has a trivial impact on the estimated effect
(Column 4).
Add data from 19761977 and 20072010 to the analysis. Incor-
porating the earlier years of data causes the semi-elasticity
estimate to change from 0.0040 to 0.0049 (Column 5). Adding
20072010 data changes the estimate to 0.0044 (Column 6).
Restrict the sample to individuals living in counties with at least
one person of each age (085) in 1990 so that age-adjusted mor-
tality rates can be calculated at the county level. Also, omit data
from Alaska and Hawaii where county codes are unavailable in
the early years of the data and from Virginia where there have
been substantial changes to county denitions over time. So that
the unemployment rates correctly correspond to the counties
contributing to the state-level estimates, use county-level BLS
data to construct unemployment rates. These changes have a
negligible impact on the estimated effect (Column 7).
Use the death rate as the outcome variable rather than the log
of the death rate, which may be undened in small counties.
This changes the interpretation of the estimate, which indicates
that a one-percentage-point increase in the unemployment rate
is associated with 3.7 fewer deaths per 100,000 (Column 8). The
semi-elasticity implied by this estimate, calculated by dividing
coefcient estimate by the mean of the outcome, is nearly iden-
tical to the direct estimate provided in Column 7 (0.0044 versus
0.0045).
Use the employment-to-population ratio as the measure of eco-
nomic conditions. The estimated effect on mortality (Column 9)
is nearly identical in magnitude to the estimate based on unem-
ployment rates with the opposite sign.
3.2. Infant health
Like Ruhm (2000), Dehejia and Lleras-Muney (2004) conduct a
state-level analysisalso characterized by Eq. 1but instead focus
on various measures of infant health and the effects of economic
conditions at the time of conception. As one of their primary out-
come measures, they consider the fraction of children who have
low birth weight (less than 2500 grams), calculated at the race-by-
state-by-year level based on vital statistics birth records for white
and black mothers at least 18 years old. As their primary mea-
sure of economic conditions, they use state unemployment rates
in the year of conception, which they base on the timing of the
Adding earlier Adding later Using county Not taking log Alt. measure of
years years matched data of mortality economic
rate conditions
19762006 19762010 19762010 19762010 19762010
(5) (6) (7) (8) (9)
0.0049*** 0.0044*** 0.0045*** 3.6979***
(0.0010) (0.0013) (0.0013) (0.9378)
3.7142***
(0.8644)
1581 1785 1680 1680 1680
mothers last menstrual period. Column 1 of Table 2 displays their
estimates, which indicate that a one-percentage-point increase in
the unemployment rate reduces the fraction of newborns with low
birth weight by 0.018.13
In columns 2 through 9 of Table 2, I progressively make changes
to the estimation strategy, ultimately arriving at an approach that
can also be used in more-disaggregated county-level analyses. In
Column 2, I instead dene a childs year of conception as nine
months prior to birth and include in the analysis children born
to mothers for whom information on the last menstrual cycle is
missing. The subsequent columns progressively include children
classied as other race to the analysis and an indicator variable for
other race to the regression model (Column 3); include children
born to mothers of all ages (Column 4); collapse the state-by-year-
by-race level data to the state-by-year level while adding controls
for the fraction in each race category (Column 5); control for the
age distribution of mothers with variables corresponding to the
fraction of mothers who are less than 18 years old, 1822 years
old, 2328 years old, and 2934 years old (Column 6); update the
sample to include data from 20002010 (Column 7); and then con-
struct the sample based on data available at the county level in the
same manner described above (Column 8). These changes tend to
have only minor impacts on the estimated effects with one excep-
tion. When the data is extended from 19761999 to 19762006
the estimated effect of unemployment increases by approximately
80 percent, suggesting that effect of economic conditions on new-
born health has grown over time. Column 9 indicates that using
the employment-to-population ratio as the measure of economic
conditions produces similar results of the opposite sign.
4. Estimates of the using different denitions of area
In this section, I begin by exploring how the estimated effects of
economic conditions on mortality vary with different levels of geo-
graphic aggregation using the identication strategy described in
the previous section. Though the convention in this literature is to
cluster standard errors at the area level, it may instead be appro-
priate to cluster on broader areas where possible, e.g., clustering
on regions for state-level analyses, because shocks to health out-
comes (that are not captured by the variables in the model) may
be correlated across adjacent areas.14 That said, the standard-error
13
Estimates are weighted by the number of births in each cell and standard-error
estimates are clustered at the state level.
14
For example, a heat wave that strikes a particular region of the country would
likely lead to a positive correlation in the errors for the areas in that region, all else
equal.
 J.M. Lindo / Journal of Health Economics 40 (2015) 8396
Table 2
State-level estimates of the effect of economic conditions at the time of conception on fraction with low birth weight.
Dehejia Impute Including all Including all
and Lleras- conception races ages of
Muney year mothers
(2004)
Years analyzed 19761999 19761999 19761999 19761999
(1) (2) (3) (4)
Unemployment 0.0180*** 0.0171* 0.0167** 0.0157*
rate (0.0063) (0.0088) (0.0080) (0.0079)
Emp-to-Pop
ratio
Observations 2447 2448 3672 3672
Notes: See Section 3.2 for details on the data and measures used in each column. All regressions include year xed effects, state xed effects, and state-specic trends. In
addition, all regressions are weighted by the number of births in each cell and cluster standard-error estimates on the state.
* Signicant at 10%; ** signicant at 5%; *** signicant at 1%.
estimates based on the conventional approach tend to be the same
asor more conservative thanestimates that cluster on broader
areas, except when the analysis is conducted at the county level. For
county-level estimates, clustering the standard errors at the state
level yields the largest estimates (though quite similar to the esti-
mates that cluster on BEA economic areas), which indicates that it
is important to allow the errors to be correlated across counties
within states when conducting inference. As such, I take this
approach where applicable. Further recognizing that year-specic
shocks may be correlated across areas, I have also investigated the
additional importance of allowing the errors to be correlated across
areas within a given year by estimating two-way standard esti-
mates following Cameron et al. (2011). I demonstrate that this can
be quite importantby way of a comparison with standard-error
estimates that do not allow for two-way clusteringin the rst
table that follows before taking this approach in all subsequent
tables.
Table 3 focuses on how the estimated effect of economic
conditions on overall mortality varies with different levels of
geographic aggregation across columns 1 through 5. Using the
broadest denition of area, region, the estimate indicates that a
one-percentage-point increase in the employment-to-population
ratio is associated with an additional 3.44 deaths per 100,000. The
estimate is a bit larger when the analysis is instead conducted at
the state level (3.71 per 100,000) and then the estimates become
smaller and smaller as we progressively consider more narrowly
dened areas. That said, these changes are relatively modest when
we move from the state level to the BEA Economic Area to the
BEA Component Economic Area (3.71 to 3.03 to 2.82). In contrast,
the magnitude of the estimate falls substantially to 1.75 when the
analysis is conducted at the county level.
Table A1 in the Appendix takes the same approach to separately
estimate the effects on youth mortality (age 017), working-age
mortality (age 1864), and elderly mortality (age 65+). This exer-
cise is motivated by Miller et al. (2009) who explain that, even
though unemployment has signicant effects on younger and older
individuals, the additional deaths that tend to be observed during
recessions largely consist of the elderly. Again, the magnitudes of
the estimates are smaller in more-disaggregated analyses, except
for working-age mortality for which there is less robust evidence
of a link between economic conditions and mortality.
Table A1 also shows estimated effects on mortality due to
cardiovascular problems, motor-vehicle accidents, and suicides.15
15
This is a subset of the categories considered in Stevens et al. (2011). See their
appendix for cause-of-death codes.
87
Collapsing Adding Adding recent Using county Alt. measure of
race-level data mother age years matched data economic
controls conditions
19761999 19761999 19762010 19762010 19762010
(5) (6) (7) (8) (9)
0.0144* 0.0177** 0.0320*** 0.0325***
(0.0072) (0.0072) (0.0089) (0.0093)
0.0324***
(0.0084)
1224 1224 1785 1680 1680
Echoing earlier work, these estimates reveal that economic
downturns reduce deaths due to cardiovascular causes and
motor-vehicle accidents while increasing deaths due to suicides.
Moreover, these estimates exhibit a pattern consistent with those
for overall mortalityregional and state-level analyses yield esti-
mates largest in magnitude while county-level analyses yield
estimates smallest in magnitude, with economic-area-level and
component-economic-area-level analyses yielding estimates in-
between.
As a whole, the results shown in Tables 3 and A1 demonstrate
that the estimated effects of economic conditions on mortality
depend on the level of geographic aggregation. This raises the ques-
tion: do the results vary because of the level of aggregation of
the measure of economic conditions, because of the aggregation of
other variables, or something else? In addition, are the differences
in the estimates statistically signicant?
Table 4 sheds light on these questions by considering the effects
of the economic conditions of regions, states, economic areas, com-
ponent economic areas, and counties in separate regressions using
county-level data and the set of associated county-level control vari-
ables. While this alternative approach does bring the estimated
effects of economic conditions of regions, states, economic areas,
and component economic areas closer to the estimated effects
of county economic conditions, the latter remain considerably
smaller in magnitude.16 In particular, the estimated effect of a one-
percentage-point increase in county employment-to-population
ratios is 1.75 additional deaths per 100,000 while the estimated
effect of a one-percentage-point increase in state employment-
to-population ratios is 3.25 additional deaths per 100,000. This
difference is statistically signicant at the one-percent level. We
can similarly reject hypotheses that the effect of county eco-
nomic conditions is the same as the effects of economic conditions
of regions, economic areas, and component economic areas. We
generally cannot, however, reject hypotheses that the effects of
these broader measures are same.17 Table A2 in the Appendix
16
In ancillary analyses (not shown), I have taken the intermediate step between
Tables 3 and 4 by estimating models using county-level data to examine the effects
of the more-aggregated measures of economic conditions while controlling for other
factors in a similarly aggregated manner (e.g., estimating the effects of state eco-
nomic conditions controlling for state xed effects, state-specic trends, and state
demographics as opposed to estimating the effects of state economic conditions
controlling for county xed effects, county-specic trends, and county demograph-
ics). These estimates were nearly identical to those in Table 3. This indicates that
the reduced variability in the estimates in Table 4 is due to the richer set of controls
and not because of the level of aggregation of the outcome variable.
17
In order to estimate the standard error for the difference in parameter estimates
across the different models in a way that allows for two-way clustering on states
88 J.M. Lindo / Journal of Health Economics 40 (2015) 8396

Table 3
Estimated effects of economic conditions on mortality (per 100,000 residents) using different denitions of areas to aggregate data.

Level of aggregation Region State BEA econ area BEA comp. econ area County
(1) (2) (3) (4) (5)

Emp-to-Pop ratio 3.4422 3.7142 3.0278 2.8246 1.7495

(1.2055)** (0.8644)*** (0.6057)*** (0.5222)*** (0.2267)***
[1.3826]** [1.0089]*** [0.8357]*** [0.7343]*** [0.4984]***
Observations 280 1680 6160 11760 102865

Notes: Data aggregated to the area level, spanning 19762010, are based on employment data produced by the BEA, population data produced by SEER, and death counts
from restricted-use Vital Statistics micro-record multiple-cause-of-death les provided by the NCHS. All estimates control for year xed effects, area xed effects, area-
specic trends, and demographics, and are weighted by the population in each cell. Standard-error estimates shown in parentheses are clustered on the area dened by
the level of geographic aggregation, except for the county-level analyses where standard-error estimates are more conservative when clustered at the state level. Two-way
standard-error estimates that additionally allow errors to be correlated across areas within a given year are shown in brackets.
* Signicant at 10%; ** signicant at 5%; *** signicant at 1%.

Table 4
Estimated effects of economic conditions on mortality (per 100,000 residents) using county-level data and aggregate measures of economic conditions.

Level of aggregation of Emp-to-Pop ratio Region State BEA econ area BEA comp. econ area County
(1) (2) (3) (4) (5)

Emp-to-Pop ratio 3.2707*** 3.2508*** 3.0333*** 2.9374*** 1.7495***

(1.1607) (0.8484) (0.7381) (0.6893) (0.4984)

Notes: Data spanning 19762010 are at the county level and are based on employment data produced by the BEA, population data produced by SEER, and death counts from
restricted-use Vital Statistics micro-record multiple-cause-of-death les provided by the NCHS. All estimates control for year xed effects, county xed effects, county-
specic trends, and county demographics, and are weighted by the population in each cell. Standard-error estimates shown in parentheses allow for two-way clustering
(state and year).
* Signicant at 10%; ** signicant at 5%; *** signicant at 1%.

shows similar results for age-specic and cause-specic mortality increase the mortality rate (per 100,000) by 3.69 which is quite
rates. similar to the estimate when state employment-to-population
Table 5 takes another step towards understanding why the esti- ratios are considered on their own (3.71 and 3.25 in Tables 3 and 4,
mated effects of county economic conditions are smaller than the respectively).
estimated effects of broader measures by simultaneously consid- More broadly, these results suggest an explanation for why the
ering the effects of economic conditions in counties and in their estimated effects of state (or economic area or region) economic
surrounding areas. To begin, Column 1 shows the results from a
regression that estimates the effects of county and state economic
conditions, where statewide economic conditions are calculated for
Table 5
each county using information from other counties in the state. This Estimated effects of economic conditions on mortality (per 100,000 residents) using
estimate indicates that the health of individuals living in a county county-level data to simultaneously consider the effects of local and broader meas-
is signicantly related to economic conditions in the county itself ures of economic conditions.
and economic conditions in the rest of the state. In other words,
(1) (2) (3)
there appear to be spillover effects across counties in the effects of
County Emp-to-Pop 1.3054*** 1.2860*** 1.2552***
economic conditions on mortality.18 This estimate is notably con-
ratio (0.3104) (0.3018) (0.3125)
sistent with the estimates from regressions separately considering State Emp-to-Pop ratio 2.3897*** 2.1966***
the effects of state and county economic conditions. In particular, it (0.8452) (0.7921)
indicates that a one-percentage-point increase in the employment- Region Emp-to-Pop 0.5988
to-population ratio across all counties in a state is expected to ratio (0.6886)

In state in BEA econ 1.1505**

area Emp-to-Pop (0.4516)
ratio
and years, I follow Cameron et al. (2011) and obtain (1) an estimate of the variance
for the difference in the parameter estimates that allows for clustering on states, In state out of BEA econ 0.6498
(2) an estimate of the variance for the difference in the parameter estimates that area Emp-to-Pop (0.6779)
allows for clustering on years, and (3) an estimate of the variance for the difference ratio
in the parameter estimates that allows for clustering on state-years. I obtain these
estimates by (cluster) bootstrapping and then combine them in order to obtain In BEA econ area out of 1.2600**
an estimate of the standard error for the difference in parameter estimates that state Emp-to-Pop (0.5341)
allows for two-way clustering on states and years. In particular, this standard-error ratio
estimate is calculated as the square root of (1) plus (2) minus (3).
18
Though the point estimate suggests that the effect of county economic con- In region out of state 0.0428
ditions is smaller than the effect of economic conditions in the rest of the state, and BEA econ area (0.6677)
the difference is not statistically signicant. That said, the difference is statisti- Emp-to-Pop
cally signicant when focusing on youth mortality rates and motor-vehicle-accident
mortality rates, shown in Table A3 in the Appendix. This offers an interesting angle P-value from test of 0.224 0.272 0.470
for future research as motor vehicle accidents are the leading cause of death for equality
youths and one might suspect that this nding is related to the fact that over a
quarter of workers commute across counties for work (McKenzie, 2013) but very Notes: See Table 4.
few of these commuters are youths. * Signicant at 10%; ** signicant at 5%; *** signicant at 1%.
 J.M. Lindo / Journal of Health Economics 40 (2015) 8396
Table 6
Estimated effects of economic conditions on mortality (per 100,000 residents) using county-level data post-1990.
(1)
Region Emp-to-Pop ratio 2.3955
(1.6984)
State Emp-to-Pop ratio
County Emp-to-Pop ratio
In state in BEA econ area Emp-to-Pop ratio
In state out of BEA econ area Emp-to-Pop ratio
In BEA econ area out of state Emp-to-Pop ratio
In region out of state and BEA econ area Emp-to-Pop
P-value from test of equality
Notes: See Table 4.
* Signicant at 10%; ** signicant at 5%; *** signicant at 1%.
conditions are larger than the estimated effects of county economic
conditions when each is considered alonethe estimated effects of
state economic conditions are fully inclusive of spillover effects
across counties within a state, which this analysis has demon-
strated to be quite important. This is also evident for age- and
cause-specic mortality rates (Table A3).
Column 2 of Table 5 extends the analysis from Column 1 by also
considering the effects of the economic conditions of areas in differ-
ent states of the same region. While the estimated effect of regional
economic conditions has the same sign as the estimated effects of
county and state economic conditions, it is small in magnitude and
not statistically signicant. This sheds light on why the estimated
effects of region and state economic conditions tend to be similar
when each is considered alonewhile the effects of regional eco-
nomic conditions can fully capture spillover effects across states
within a region, these effects are not large enough to have a mean-
ingful inuence on the estimates. The estimates for suicides, shown
in Table A3 in the Appendix along with other mortality outcomes,
are something of an outlier but also serve to demonstrate this gen-
eral line of reasoning. In particular, the estimated effect of regional
economic conditions on suicides is statistically and economically
signicant, which is consistent with the rather large difference in
the estimated effects of regional and state economic conditions
when each is considered alone (Table A2).
Column 3 further exploits of the county-level data to consider
spillover effects by considering the effects of the economic con-
ditions of the following areas which are mutually exclusive by
construction: (1) the county itself, (2) counties in the same state
and economic area, (3) counties in the same state but different
economic areas, (4) counties in the same economic area but differ-
ent states, and (5) counties in the same region but different states
and economic areas. The results from this analysis indicate that
economic conditions outside of the state but within the same eco-
nomic area have independent effects on mortality while there is
weaker evidence that economic conditions within the state but
outside of the economic area have signicant effects on mortal-
ity. While this nding could be taken as evidence regarding the
mechanisms through which economic conditions affect health, as
it suggests that the effects of economic conditions on health are
unlikely to be strongly related to any state-level policy responses
to changing economic conditions, it is important to note that the
estimates are not precise enough for us to be able to reject the
hypothesis that the effects of economic conditions are the same for
all areas considered.
89
(2) (3) (4) (5) (6)
0.5006
(1.3607)
2.4651* 0.6381 0.4383
(1.3907) (1.0891) (1.1175)
1.5549*** 1.4470*** 1.4347*** 1.4689***
(0.4886) (0.3567) (0.3671) (0.3777)
0.9046
(0.6166)
0.6560
(0.7356)
0.0581
(0.4679)
0.1959
(1.2727)
0.416 0.350 0.012
Table 6 explores these issues using data from 1990 to 2010. This
analysis is motivated by Ruhm (2013), which consists of state-level
analyses that suggest that the link between economic conditions
and total mortality has weakened over timeso much so that esti-
mates using recent 15- and 20-year windows are not statistically
signicant, raising a question about whether there remains a link
at all.
In particular, columns 1 through 3 of Table 6 provide estimates
of the effects of the economic conditions of regions, states, and
counties, respectively, on county mortality rates as in Table 4. Con-
sistent with Ruhm (2013), the estimated effect of state economic
conditions (Column 2) is smaller when using data beginning in
1990 than when using data beginning in 1976 (2.46 versus 3.25).
Moreover, neither the estimated effect of region nor the estimated
effect of state economic conditions is statistically signicant at the
ve-percent level.19 While the estimated effect of county economic
conditions (Column 3) is smaller in magnitude than the estimated
effects of state economic conditions, it is sufciently precise to
leave little doubt that there remains a strong link between county
economic conditions and mortality.
Columns 4 through 6 of Table 6 simultaneously consider the
effects of economic conditions in a county and the economic
conditions in surrounding areas, as in Table 5. Column 4, which
simultaneously considers the effects of county and state economic
conditions, shows that county economic conditions have a sig-
nicant effect on overall mortality that is independent of the
effect of statewide economic conditions. While the correspond-
ing point estimate for statewide economic conditions is smaller
in magnitude and not statistically signicant, suggesting a more
limited role for spillover effects on overall mortality in recent years,
we cannot reject the hypotheses that the independent effects of
state and county economic conditions are the same. Moreover, for
the specic causes of death considered in Table A4 (cardiovas-
cular problems, motor vehicle, and suicide), the estimated effect
of state economic conditions is statistically signicant. In particu-
lar, these estimates indicate that state employment-to-population
ratios have positive and signicant (independent) effects on mor-
tality due to cardiovascular causes and motor vehicle accidents
while they have negative and signicant effects on mortality due
19
Also consistent with Ruhm (2013), the estimated effects on mortality due to car-
diovascular causes, motor-vehicle accidents, and suicides continue to be statistically
signicant, as shown in Table A4.
90 J.M. Lindo / Journal of Health Economics 40 (2015) 8396

Table 7
Estimated effects of economic conditions on at the time of conception on fraction of newborns with low birthweight.

Level of aggregation Region State BEA econ area BEA comp. econ area County
(1) (2) (3) (4) (5)

Panel A: All data aggregated to area level

Emp-to-Pop ratio 0.0353* 0.0324*** 0.0277*** 0.0240*** 0.0073*
(0.0193) (0.0102) (0.0069) (0.0064) (0.0042)

Panel B: County level data, only Emp-to-Pop ratio aggregated further

Emp-to-Pop ratio 0.0267* 0.0254** 0.0247*** 0.0222** 0.0073*
(0.0145) (0.0105) (0.0085) (0.0086) (0.0042)

Notes: Data spanning 19762010 are based on employment data produced by the BEA, population data produced by SEER, and birth information from restricted-use natality
les provided by the NCHS. The estimates in Panel A are based on data aggregated to the area level indicated by the column heading and control for year xed effects, area
xed effects, area-specic trends, and demographics at that level. The estimates in Panel B use county-level data and consider the effects of aggregate measures of economic
conditions (at the level indicated by the column heading) while controlling for year xed effects, county xed effects, county-specic trends, and county demographics.
All estimates and are weighted by the population in each cell. In Panel A the standard-error estimates shown in parentheses allow for two-way clustering (area and year),
except for the county-level analyses where standard-error estimates are more conservative when clustered at the state level. In Panel B the standard-error estimates shown
in parentheses also allow for two-way clustering (state and year).
* Signicant at 10%; ** signicant at 5%; *** signicant at 1%.

to suicides, which explains why the estimated effects on overall regions, states, economic areas, and component economic areas
mortality are somewhat muted. closer to the estimated effects of county economic conditions and
The estimate in Column 5, which considers spillover effects thus closer to one another, but the estimated effects of county
across areas in different states in the same region, is too impre- economic conditions remain signicantly smaller in magnitude.
cise to reject that these effects are zero but also too imprecise to Table 8 further explores why the estimated effects of county
reject that they are the same as the effects of state and county eco- economic conditions are smaller than the estimated effects of
nomic conditions. The one exception is for the estimates focused broader measures of economic conditions by simultaneously
on suicides (in Table A4), which indicates that regional economic considering the effects of the economic conditions in counties and
conditions are more strongly related to suicides than are county in their surrounding areas (as in Table 5). Like the results for mor-
or state economic conditions. This evidence is further supported tality, the results of this analysis indicate that health outcomes in
by Column 6, which considers the effects of economic conditions a county are signicantly related to economic conditions in sur-
of the county itself, counties in the same state and economic area, rounding areas in a manner that can explain why the estimated
counties in the same state but different economic areas, counties effects tend to be larger in analyses that focus on more-aggregate
in the same economic area but different states, and counties in the measures of economic conditions. In particular, the results in all
same region but different states and economic areas. of these panels indicate that infant health is signicantly related
As a whole, the estimates across Tables 3 through 6 (and the
associated tables in the Appendix) demonstrate that the estimated Table 8
effects of economic conditions on mortality are sensitive to the Estimated effects of economic conditions on at the time of conception on fraction
level of aggregation of the data, that this sensitivity results primar- of newborns with low birthweight using data at county level and simultaneously
considering the effects of local and broader measures of economic conditions.
ily from the use of aggregated measures of economic conditions,
and that this sensitivity results from spillover effects across areas (1) (2) (3)
in the effects of economic conditions on health outcomes. In the County Emp-to-Pop 0.0011 0.0014 0.0003
next section, I consider these issues in the context of the estimated ratio (0.0040) (0.0039) (0.0039)
effects of economic conditions on infant health. State Emp-to-Pop ratio 0.0273** 0.0303***
(0.0111) (0.0101)
Region Emp-to-Pop 0.0103
5. Estimates of the unemploymentinfant-health ratio (0.0137)
relationship using different denitions of area
In state in BEA econ 0.0247***
Panel A of Table 7 follows the identication strategy described area Emp-to-Pop ratio (0.0063)

in Section 3.2 to estimate the effects of economic conditions at the

In state out of BEA econ 0.0078
time of conception on the fraction of children born with low birth area Emp-to-Pop ratio (0.0094)
weight. In particular, the estimates are based on data aggregated
to the area level and models that control for area xed effects, In BEA econ area out of 0.0026
state Emp-to-Pop ratio (0.0082)
area-specic trends, year xed effects, and demographics, with dif-
ferent denitions of area across columns 1 through 5. Though the In region out of state 0.0095
estimates are always positive, indicating that downturns are asso- and BEA econ area (0.0144)
ciated with improvements in infant health, the magnitude of the Emp-to-Pop
estimated effects decline as areas are dened more narrowly in a
manner similar to the estimated effects on mortality. P-value from test of 0.038 0.035 0.004
Panel B of Table 7 begins to shed light on why the estimated equality
effects of economic conditions on infant health are sensitive to the Notes: Data spanning 19762010 are at the county level and are based on employ-
level of geographic aggregation by considering the effects of the ment data produced by the BEA, population data produced by SEER, and birth
economic conditions of regions, states, economic areas, component information from restricted-use natality les provided by the NCHS. All estimates
control for year xed effects, county xed effects, and county-specic trends, and
economic areas, and counties in separate regressions using county-
demographics of mothers. All estimates are weighted by the population in each cell.
level data and the set of associated county-level control variables Standard-error estimates shown in parentheses allow for two-way clustering (state
(as in Table 4). As with the analysis of mortality outcomes, this and year).
approach brings the estimated effects of economic conditions of * Signicant at 10%; ** signicant at 5%; *** signicant at 1%.
 J.M. Lindo / Journal of Health Economics 40 (2015) 8396 91

Table A1
Estimated effects of economic conditions on mortality (per 100,000 residents) using different denitions of areas to aggregate data.

Level of aggregation Region State BEA econ area BEA comp. econ area County
(1) (2) (3) (4) (5)

Outcome: Youth (017) mortality

Emp-to-Pop ratio 1.5358 1.4232 1.2067 1.0915 0.5296
(0.2398)*** (0.1969)*** (0.1129)*** (0.1067)*** (0.0610)***
[0.2618]*** [0.2255]*** [0.1456]*** [0.1428]*** [0.1035]***

Outcome: Working-age (1864) mortality

Emp-to-Pop ratio 0.8080 1.5571 0.9774 0.9348 0.9908
(1.2671) (0.9290) (0.5767)* (0.4794)* (0.1724)***
[1.3754] [1.0211] [0.7671] [0.6583] [0.4128]**

Outcome: Elderly (65+) mortality

Emp-to-Pop ratio 18.5204 17.0419 15.2447 14.4166 8.1578
(3.9578)*** (4.1183)*** (2.8501)*** (2.4747)*** (1.1032)***
[5.6205]*** [4.7881]*** [3.8160]*** [3.4055]*** [1.9851]***

Outcome: Mortality due to cardiovascular causes

Emp-to-Pop ratio 1.3715 1.5043 1.4673 1.3337 0.6987
(0.7584) (0.4965)*** (0.3163)*** (0.2700)*** (0.1494)***
[0.8789] [0.5591]*** [0.4539]*** [0.3822]*** [0.2242]***

Outcome: Mortality due to motor vehicle accidents

Emp-to-Pop ratio 0.4660 0.4148 0.3980 0.3575 0.1564
(0.1604)** (0.0880)*** (0.0691)*** (0.0573)*** (0.0319)***
[0.1698]*** [0.1004]*** [0.1080]*** [0.0940]*** [0.0444]***

Outcome: Mortality due to suicides

Emp-to-Pop ratio 0.2197 0.1675 0.1333 0.1019 0.0331
(0.0547)*** (0.0408)*** (0.0429)*** (0.0408)** (0.0212)
[0.0594]*** [0.0426]*** [0.0587]** [0.0509]** [0.0237]

Note: This table corresponds to Table 3 for additional outcomes.

to the economic conditions of other counties in the same state. In
contrast, these results also indicate that a countys own economic
conditions are not signicantly related to infant health.
In terms of the bigger picture question regarding how and why
economic conditions are related to infant health, we must consider
the degree to which economic conditions affect the number and
composition of newborns. Towards this end, Tables A5 and A6 in
the Appendix present analyses corresponding to Tables 7 and 8
but instead focus on birth rates (dened as the number of births
per 10,000 women between the ages of 15 and 44), the fraction of
newborns with a white mother, and the fraction of newborns with
a mother younger than 18 years old. The estimates for birth rates
in Table A5 are positive and usually signicant, indicating that fer-
tility is procyclical, while the estimates in Table A6 suggest that we
cannot rule out the possibility that fertility is affected equally by
local economic conditions and broader economic conditions.20 As
such, the estimated effects on infant health are difcult to inter-
pret because they may be inuenced by non-random selection.
Indeed, the estimated effects on the fraction of children born to
white mothers and to young mothers are sometimes statistically
signicant and are often too imprecise to rule out economically
signicant effects in one direction or the other.
6. Conclusion
Because time-series analyses of health outcomes are subject
to such a large suite of potential biases, approaches that make
use geographically disaggregated data are a natural alternative
to understanding the effects of macroeconomic conditions. This
paper shows that the level of (dis)aggregation has important conse-
quences for the estimated effects of economic conditions in panel
data models that exploit variation in the timing and severity of
contractions and expansions across areas using the now-common
area approach. Most signicantly, the estimated effects based
on county-level variation in economic conditions are signicantly
smaller in magnitude than those based more-aggregate variation.
Analyses using county-level data to simultaneously consider the
effects of the economic conditions in a county and the effects of the
economic conditions in surrounding areas indicate that this occurs
because there are signicant spillover effects of economic condi-
tions on health outcomes across counties. As such, estimates based
on more aggregate analyses may be more reliable in the sense that
they can fully capture such effects. That said, analyses using more-
disaggregated data can offer increased precision and the ability to
consider such spillover effects directly. Having demonstrated that
such spillover effects are empirically relevant, understanding how
and why they occur would appear to be a critical next step to be
taken in future work.
Acknowledgements
I am extremely grateful to Maria Padilla-Romo and Eric Stein-
mann for research assistance. I also thank the Editor, David Cutler,
and anonymous referees for their valuable comments, in addition
to Alan Barreca, Ben Hansen, Mark Hoekstra, Doug Miller, Adri-
ana Lleras-Muney, Marianne Page, Chris Ruhm, Jessamyn Schaller,
Ann Huff Stevens, Glen Waddell, Wes Wilson, and Xiaohan Zhang
and seminar participants at the University of Maryland, the Uni-
versity of Oregon, and the 2012 Meetings of the Society of Labor
Economists.

20
Schaller (2012) also nds that fertility is procyclical. Appendix A. Additional tables
92 J.M. Lindo / Journal of Health Economics 40 (2015) 8396

Table A2
Estimated effects of economic conditions on mortality (per 100,000 residents) using county-level data and aggregate measures of economic conditions.

Level of aggregation of Emp-to-Pop ratio Region State BEA econ area BEA comp. econ area County
(1) (2) (3) (4) (5)

Outcome: Youth (017) mortality

Emp-to-Pop ratio 1.1050*** 1.2549*** 1.1699*** 1.0575*** 0.5296***
(0.2837) (0.1689) (0.1584) (0.1574) (0.1035)

Outcome: Working-age (1864) mortality

Emp-to-Pop ratio 1.1550 1.1408 1.3134* 1.3698** 0.9908**
(1.1188) (0.8173) (0.7310) (0.6765) (0.4128)

Outcome: Elderly (65+) mortality

Emp-to-Pop ratio 16.8945*** 16.4891*** 14.0185*** 13.3693*** 8.1578***
(4.9889) (3.8895) (3.2345) (2.9448) (1.9851)

Outcome: Mortality due to cardiovascular causes

Emp-to-Pop ratio 1.2127* 1.2142** 1.2602*** 1.2252*** 0.6987***
(0.7130) (0.5040) (0.4678) (0.4207) (0.2242)

Outcome: Mortality due to motor vehicle accidents

Emp-to-Pop ratio 0.4345*** 0.4263*** 0.4037*** 0.3675*** 0.1564***
(0.1069) (0.0849) (0.0771) (0.0736) (0.0444)

Outcome: Mortality due to suicides

Emp-to-Pop ratio 0.1627*** 0.1022*** 0.0980** 0.0759* 0.0331
(0.0432) (0.0334) (0.0431) (0.0430) (0.0237)

Note: This table corresponds to Table 4 for additional outcomes.

Table A3
Estimated effects of economic conditions on mortality (per 100,000 residents) using county-level data to simultaneously consider the effects of local and broader measures
of economic conditions.

(1) (2) (3)

Outcome: Youth (017) mortality

County Emp-to-Pop ratio 0.3478*** 0.3449*** 0.3357***
(0.0909) (0.0938) (0.0835)
State Emp-to-Pop ratio 0.9947*** 0.9666***
(0.1698) (0.2044)
Region Emp-to-Pop ratio 0.0892
(0.2790)
In state in BEA econ area Emp-to-Pop ratio 0.5786***
(0.1211)
In state out of BEA econ area Emp-to-Pop ratio 0.2606
(0.1875)
In BEA econ area out of state Emp-to-Pop ratio 0.2075**
(0.0847)
In region out of state and BEA econ area Emp-to-Pop 0.0794
(0.2739)

Outcome: Working-age (1864) mortality

County Emp-to-Pop ratio 0.8377*** 0.8325*** 0.8047***
(0.2708) (0.2647) (0.2840)
State Emp-to-Pop ratio 0.8212 0.7703
(0.9085) (0.8578)
Region Emp-to-Pop ratio 0.1567
(0.7033)
In state in BEA econ area Emp-to-Pop ratio 0.5394
(0.5039)
In state out of BEA econ area Emp-to-Pop ratio 0.1337
(0.6878)
In BEA econ area out of state Emp-to-Pop ratio 1.1955**
(0.5718)
In region out of state and BEA econ area Emp-to-Pop 0.4151
(0.6509)

Outcome: Elderly (65+) mortality

County Emp-to-Pop ratio 6.0401*** 5.9017*** 5.9026***
(1.2783) (1.1664) (1.1140)
State Emp-to-Pop ratio 11.1979*** 9.6288***
(3.4992) (3.2606)
Region Emp-to-Pop ratio 4.7894*
(2.6656)
In state in BEA econ area Emp-to-Pop ratio 3.6381**
(1.8029)
In state out of BEA econ area Emp-to-Pop ratio 4.4421
(3.2410)
 J.M. Lindo / Journal of Health Economics 40 (2015) 8396 93

Table A3 (Continued)

(1) (2) (3)

In BEA econ area out of state Emp-to-Pop ratio 3.8752*

(2.1989)
In region out of state and BEA econ area Emp-to-Pop 3.2569
(2.8606)

Outcome: Mortality due to cardiovascular causes

County Emp-to-Pop ratio 0.5512*** 0.5356*** 0.5372***
(0.1755) (0.1565) (0.1522)
State Emp-to-Pop ratio 0.7940* 0.6393
(0.4397) (0.3963)
Region Emp-to-Pop ratio 0.4795
(0.5983)
In state in BEA econ area Emp-to-Pop ratio 0.3553
(0.2656)
In state out of BEA econ area Emp-to-Pop ratio 0.2309
(0.3861)
In BEA econ area out of state Emp-to-Pop ratio 0.3282
(0.2316)
In region out of state and BEA econ area Emp-to-Pop 0.1389
(0.6003)

Outcome: Mortality due to motor vehicle accidents

County Emp-to-Pop ratio 0.0896** 0.0882** 0.0809**
(0.0377) (0.0371) (0.0374)
State Emp-to-Pop ratio 0.3591*** 0.3445***
(0.0691) (0.0695)
Region Emp-to-Pop ratio 0.0453
(0.0775)
In state in BEA econ area Emp-to-Pop ratio 0.1330***
(0.0394)
In state out of BEA econ area Emp-to-Pop ratio 0.2306***
(0.0560)
In BEA econ area out of state Emp-to-Pop ratio 0.0384
(0.0364)
In region out of state and BEA econ area Emp-to-Pop 0.0488
(0.0795)

Outcome: Mortality due to suicides

County Emp-to-Pop ratio 0.0170 0.0138 0.0114
(0.0224) (0.0217) (0.0204)
State Emp-to-Pop ratio 0.0867*** 0.0546*
(0.0255) (0.0281)
Region Emp-to-Pop ratio 0.0994**
(0.0433)
In state in BEA econ area Emp-to-Pop ratio 0.0440
(0.0362)
In state out of BEA econ area Emp-to-Pop ratio 0.0322
(0.0415)
In BEA econ area out of state Emp-to-Pop ratio 0.0060
(0.0381)
In region out of state and BEA econ area Emp-to-Pop 0.0758*
(0.0441)

Note: This table corresponds to Table 5 for additional outcomes.

Table A4
Estimated effects of economic conditions on mortality (per 100,000 residents) using county-level data post-1990.

(1) (2) (3) (4) (5) (6)

Outcome: Youth (017) mortality

Region Emp-to-Pop ratio 0.4469 0.3735
(0.3679) (0.3555)
State Emp-to-Pop ratio 0.4625 0.0695 0.2167
(0.3205) (0.2984) (0.3430)
County Emp-to-Pop ratio 0.3837*** 0.3951*** 0.3865*** 0.3837***
(0.1291) (0.1248) (0.1230) (0.1178)
In state in BEA econ area Emp-to-Pop ratio 0.2489*
(0.1423)
In state out of BEA econ area Emp-to-Pop ratio 0.4351
(0.2799)
In BEA econ area out of state Emp-to-Pop ratio 0.0245
(0.1090)
In region out of state and BEA econ area Emp-to-Pop 0.3158
(0.3258)

Outcome: Working-age (1864) mortality

Region Emp-to-Pop ratio 2.0516 0.4238
(1.6222) (0.8304)
94 J.M. Lindo / Journal of Health Economics 40 (2015) 8396

Table A4 (Continued)

(1) (2) (3) (4) (5) (6)

State Emp-to-Pop ratio 1.9518 0.7182 0.5491

(1.3620) (1.0544) (0.9150)
County Emp-to-Pop ratio 1.0413** 0.9185*** 0.9077*** 0.9280***
(0.4434) (0.3107) (0.3050) (0.3148)
In state in BEA econ area Emp-to-Pop ratio 0.5303
(0.4743)
In state out of BEA econ area Emp-to-Pop ratio 0.0315
(0.5322)
In BEA econ area out of state Emp-to-Pop ratio 0.0241
(0.3402)
In region out of state and BEA econ area Emp-to-Pop 0.2472
(0.7998)

Outcome: Elderly (65+) mortality

Region Emp-to-Pop ratio 9.4872 4.3082
(6.2207) (5.4216)
State Emp-to-Pop ratio 8.8148 0.2639 1.4922
(5.6797) (5.0315) (5.2421)
County Emp-to-Pop ratio 7.4274*** 7.3827*** 7.2833*** 7.5996***
(2.0399) (1.3689) (1.4272) (1.5323)
In state in BEA econ area Emp-to-Pop ratio 0.7919
(2.5386)
In state out of BEA econ area Emp-to-Pop ratio 4.4903
(4.3451)
In BEA econ area out of state Emp-to-Pop ratio 1.4530
(2.5935)
In region out of state and BEA econ area Emp-to-Pop 2.9396
(5.1082)

Outcome: Mortality due to cardiovascular causes

Region Emp-to-Pop ratio 1.0599* 0.0908
(0.6038) (0.6370)
State Emp-to-Pop ratio 1.2295*** 0.9337** 0.9699**
(0.4255) (0.4228) (0.4811)
County Emp-to-Pop ratio 0.5108** 0.3529* 0.3551** 0.3902**
(0.2240) (0.1830) (0.1802) (0.1687)
In state in BEA econ area Emp-to-Pop ratio 0.4430**
(0.1802)
In state out of BEA econ area Emp-to-Pop ratio 0.2858
(0.4688)
In BEA econ area out of state Emp-to-Pop ratio 0.0674
(0.2781)
In region out of state and BEA econ area Emp-to-Pop 0.0114
(0.6071)

Outcome: Mortality due to motor vehicle accidents

Region Emp-to-Pop ratio 0.3237*** 0.0446
(0.0812) (0.1096)
State Emp-to-Pop ratio 0.3194*** 0.2365*** 0.2187**
(0.0794) (0.0857) (0.0994)
County Emp-to-Pop ratio 0.1068*** 0.0668* 0.0657* 0.0651*
(0.0341) (0.0355) (0.0364) (0.0351)
In state in BEA econ area Emp-to-Pop ratio 0.1169**
(0.0462)
In state out of BEA econ area Emp-to-Pop ratio 0.0947
(0.0890)
In BEA econ area out of state Emp-to-Pop ratio 0.0075
(0.0529)
In region out of state and BEA econ area Emp-to-Pop 0.0642
(0.1078)

Outcome: Mortality due to suicides

Region Emp-to-Pop ratio 0.1677*** 0.1043**
(0.0479) (0.0462)
State Emp-to-Pop ratio 0.1175*** 0.1076** 0.0659
(0.0381) (0.0422) (0.0510)
County Emp-to-Pop ratio 0.0292* 0.0110 0.0085 0.0093
(0.0175) (0.0193) (0.0177) (0.0179)
In state in BEA econ area Emp-to-Pop ratio 0.0076
(0.0305)
In state out of BEA econ area Emp-to-Pop ratio 0.0560**
(0.0240)
In BEA econ area out of state Emp-to-Pop ratio 0.0030
(0.0257)
In region out of state and BEA econ area Emp-to-Pop 0.1083**
(0.0475)

Note: This table corresponds to Table 6 for additional outcomes.

 J.M. Lindo / Journal of Health Economics 40 (2015) 8396 95

Table A5
Estimated effects of economic conditions on at the time of conception on birth rates and mothers characteristics.

Level of aggregation Region State BEA econ area BEA comp. econ area County
(1) (2) (3) (4) (5)

Panel A: All data aggregated to area level

Outcome: Birth rate
Emp-to-Pop ratio 0.4210 0.5286** 0.4679* 0.4877** 0.3797**
(0.3831) (0.2691) (0.2766) (0.2426) (0.1611)

Outcome: Fraction with white mother

Emp-to-Pop ratio 0.0654 0.0155 0.0311 0.0407 0.0432
(0.0880) (0.0692) (0.0434) (0.0365) (0.0337)

Outcome: Fraction with mother less than 18 years old

Emp-to-Pop ratio 0.0334 0.0065 0.0019 0.0028 0.0074
(0.0336) (0.0144) (0.0177) (0.0155) (0.0075)

Panel B: County level data, only Emp-to-Pop ratio aggregated further

Outcome: Birth rate
Emp-to-Pop ratio 0.4114 0.5278* 0.4820* 0.4987* 0.3797**
(0.3591) (0.3067) (0.2925) (0.2756) (0.1611)

Outcome: Fraction with white mother

Emp-to-Pop ratio 0.0095 0.0162 0.0255 0.0374 0.0432
(0.0714) (0.0627) (0.0559) (0.0499) (0.0337)

Outcome: Fraction with mother less than 18 years old

Emp-to-Pop ratio 0.0256 0.0041 0.0032 0.0037 0.0074
(0.0191) (0.0157) (0.0146) (0.0131) (0.0075)

Note: This table corresponds to Table 7 for additional outcomes, omitting demographic control variables.

Table A6
Estimated effects of economic conditions on at the time of conception on infant outcomes using data at county level and simultaneously considering the effects of local and
broader measures of economic conditions.

(1) (2) (3)

Outcome: Birth rate

County Emp-to-Pop ratio 0.3110*** 0.3027*** 0.2830***
(0.0894) (0.0911) (0.0983)
State Emp-to-Pop ratio 0.3029 0.1805
(0.2483) (0.2958)
Region Emp-to-Pop ratio 0.4066
(0.4645)
In state in BEA econ area Emp-to-Pop ratio 0.0404
(0.1879)
In state out of BEA econ area Emp-to-Pop ratio 0.2503
(0.2299)
In BEA econ area out of state Emp-to-Pop ratio 0.1833
(0.1491)
In region out of state and BEA econ area Emp-to-Pop 0.4014
(0.4182)

P-value from test of equality 0.970 0.891 0.221

Outcome: Fraction with white mother

County Emp-to-Pop ratio 0.0563** 0.0541** 0.0549**
(0.0242) (0.0243) (0.0239)
State Emp-to-Pop ratio 0.0578 0.0907
(0.0615) (0.0684)
Region Emp-to-Pop ratio 0.1092
(0.1093)
In state in BEA econ area Emp-to-Pop ratio 0.0656
(0.0461)
In state out of BEA econ area Emp-to-Pop ratio 0.0492
(0.0484)
In BEA econ area out of state Emp-to-Pop ratio 0.0598
(0.0460)
In region out of state and BEA econ area Emp-to-Pop 0.0953
(0.0998)

P-value from test of equality 0.097 0.093 0.014

96 J.M. Lindo / Journal of Health Economics 40 (2015) 8396
Table A6 (Continued)
(1)
Outcome: Fraction with mother less than 18 years old
County Emp-to-Pop ratio 0.0104*
(0.0061)
State Emp-to-Pop ratio 0.0130
(0.0151)
Region Emp-to-Pop ratio
In state in BEA econ area Emp-to-Pop ratio
In state out of BEA econ area Emp-to-Pop ratio
In BEA econ area out of state Emp-to-Pop ratio
In region out of state and BEA econ area Emp-to-Pop
P-value from test of equality 0.157
Note: This table corresponds to Table 8 for additional outcomes, omitting demographic control variables.
References
Angrist, J., Krueger, A., 1999. Empirical strategies in labor economics. In: Ashenfelter,
O., Card, D. (Eds.), Handbook of Labor Economics. Elsevier, North Holland.
Bartik, T.J., 1996. The distributional effects of local labor demand and industrial
mix: estimates using individual panel data. Journal of Urban Economics 40 (2),
150178.
Blanchard, O.J., Katz, L.F., 1992. Regional evolutions. Brookings Papers on Economic
Activity 1, 175.
Bound, J., Holzer, H.J., 2000. Demand shifts, population adjustments, and labor mar-
ket outcomes during the 1980. Journal of Labor Economics 18 (1), 2054.
Brenner, M., 1973. Mental Illness and the Economy. Harvard University Press,
Cambridge, MA.
Brenner, M., 1975. Trends in alcohol consumption and associated illnesses. American
Journal of Public Health 656 (12), 12791292.
Brenner, M., 1979. Mortality and the national economy. The Lancet 314 (8142),
568573.
Cameron, A.C., Gelbach, J.B., Miller, D.L., 2011. Robust inference with multiway clus-
tering. Journal of Business & Economic Statistics 29 (2), 238249.
Currie, J., Tekin, E., 2011. Is The Foreclosure Crisis Making Us Sick?, NBER Working
Paper No. 17310.
Dehejia, R., Lleras-Muney, A., 2004. Booms, busts, and babies health. Quarterly
Journal of Economics 119 (3), 10911130.
Forbes, J.F., McGregor, A., 1984. Unemployment and mortality in post-war Scotland.
Journal of Health Economics 3 (3), 239257.
Gerdtham, U.-G., Johannesson, M., 2005. Business cycles and mortality: results from
Swedish microdata. Social Science and Medicine 60 (1), 205218.
Gerdtham, U.-G., Ruhm, C., 2006. Deaths rise in good economic times: evidence from
the OECD. Economics and Human Biology 4 (3), 298316.
Glaeser, E.L., Gyourko, J., 2005. Urban decline and durable housing. Journal of Polit-
ical Economy 113 (2), 345375.
Griliches, Z., 1977. Estimating the returns to schooling: some econometric problems.
Econometrica 45 (1), 122.
Griliches, Z., Hausman, J., 1986. Errors in variables in panel data. Journal of Econo-
metrics 33 (1), 93118.
Halliday, T.J., 2007. Business cycles, migration and health. Social Science & Medicine
(64), 14201424.
Hausman, J., 2001. Mismeasured variables in econometric analysis: problems from
the right and problems from the left. Journal of Economic Perspectives 15 (4),
5767.
(2) (3)
0.0113** 0.0074
(0.0054) (0.0051)
0.0003
(0.0135)
0.0424**
(0.0192)
0.0047
(0.0128)
0.0069
(0.0162)
0.0054
(0.0094)
0.0419**
(0.0174)
0.007 0.075
Hoynes, H.W., 2000. Local labor markets and welfare spells: do demand conditions
matter? Review of Economics and Statistics 82 (3), 351368.
Johansson, E., 2004. A note on the impact of hours worked on mortality in OECD
countries. European Journal of Health Economics 5 (4), 335340.
Joyce, T.J., Mocan, H.N., 1993. Unemployment and infant health: time-series evi-
dence from the state of Tennessee. Journal of Human Resources 28 (1), 185203.
Laporte, A., 2004. Do economic cycles have a permanent effect on population health?
Revisiting the Brenner hypothesis. Health Economics 13 (8), 767779.
Lin, S.-J., 2009. Economic uctuations and health outcome: a panel analysis of Asia-
Pacic countries. Applied Economics 41 (4), 519530.
McAvinchey, I.D., 1988. A comparison of unemployment, income and mortality
interaction for ve european countries. Applied Economics 20 (4), 453471.
McKenzie, B., 2013. County-to-County Commuting Flows: 200610. US Census
Bureau.
Miller, D.L., Page, M.E., Stevens, A.H., Filipski, M., 2009. Why are recessions good for
your health? American Economic Review 99 (2), 122127.
Neumayer, E., 2004. Recessions lower (some) mortality rates. Social Science and
Medicine 58 (6), 10371047.
Notowidigdo, M.J., 2011. The Incidence of Local Labor Demand Shocks, NBER Work-
ing Paper No. 17167.
Ruhm, C.J., 2000. Are recessions good for your health? Quarterly Journal of Eco-
nomics 115 (2), 617650.
Ruhm, C.J., 2003. Good times make you sick. Journal of Health Economics 22 (4),
637658.
Ruhm, C.J., 2005. Healthy living in hard times. Journal of Health Economics 24 (2),
341363.
Ruhm, C.J., 2007. A healthy economy can break your heart. Demography 44 (4),
829848.
Ruhm, C.J., 2013. Recessions, Healthy No More?, NBER Working Paper No. 19287.
Saks, R.E., Wozniak, A., 2011. Labor reallocation over the business cycle: new evi-
dence from internal migration. Journal of Labor Economics 29 (4), 697739.
Schaller, J., 2012. Boom, Busts and Fertility: Testing the Becker Model Using Gender-
Specic Labor Demand, University of Arizona Working Paper.
Stevens, A.H., Miller, D.L., Page, M.E., Filipski, M., 2011. The Best of Times, the Worst
of Times: Understanding Pro-cyclical Mortality, NBER Working Paper No. 17657.
Tapia Granados, J., 2005. Recessions and mortality in Spain, 19801997. European
Journal of Population 21 (4), 393422.
Wozniak, A., 2010. Are college graduates more responsive to distant labor market
opportunities? Journal of Human Resources 45 (4), 944970.
 Journal of Health Economics 40 (2015) 97108

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

Physician responses to rising local unemployment rates: Healthcare

provision to Medicare and privately insured patients
Daifeng He a , Melissa McInerney b, , Jennifer Mellor a
a
Department of Economics and Thomas Jefferson Program in Public Policy at the College of William and Mary, Williamsburg, VA, United States
b
Department of Economics, Tufts University, Medford, MA, United States

a r t i c l e i n f o a b s t r a c t

Article history: Prior studies suggest that hospital care is countercyclical among Medicare beneciaries, and if anything,
Received 28 March 2013 procyclical among the non-elderly. In this paper, we provide the rst physician-level analysis of changes
Received in revised form in healthcare provision to Medicare and privately insured patients across the business cycle. Using Florida
19 December 2014
discharge data aggregated to the physician level, we nd that as county unemployment rates increase,
Accepted 29 December 2014
physicians treat fewer privately insured patients in both inpatient and outpatient settings. In contrast,
Available online 6 January 2015
physicians who are more exposed to income losses during recessions provide more care to Medicare
patients as the unemployment rate rises. Further analysis suggests that easing capacity constraints may
JEL classication:
I1
contribute to this rise in Medicare volume; however, even in areas that are not capacity constrained, care
H3 provided to Medicare patients remains countercyclical among physicians with a large share of privately
insured patients. This pattern is consistent with demand inducement in response to a negative income
Keywords: shock.
Provider-induced demand 2015 Elsevier B.V. All rights reserved.
Supply-side responses
Medicare
Health and recessions
Hospital care

1. Introduction Medicare patients as state unemployment rates rise (McInerney

Several studies from the recessions and health literature show
that as unemployment rates rise, hospital utilization by Medi-
care fee-for-service (FFS) beneciaries increases (Ruhm, 2007;
McInerney and Mellor, 2012a,b), and, if anything, hospital use
among the overall adult population falls (Ruhm, 2003). Standard
demand-side explanations do not account for increases in Medi-
care utilization since recessions are associated with decreases in
income and insurance coverage. Evidence of healthcare supplier
responses to the business cycle is scant, but one study nds that
physicians are more likely to report that their practice accepts new
We thank seminar participants at Cornell University Policy Analysis and Man-
agement, Bryn Mawr College, Northeastern University, Tufts University, Emory
University, and the University of Maryland. We thank Amy Filipek for excellent
research assistance and the Schroeder Center for Health Policy at the College of
William and Mary for nancial support.
Corresponding author. Tel.: +1 617 627 6653.
E-mail addresses: dhe@wm.edu (D. He), melissa.mcinerney@tufts.edu
(M. McInerney), jmmell@wm.edu (J. Mellor).
and Mellor, 2012a). One explanation for countercyclical Medicare
utilization is that recessions lead to revenue losses for physicians
because patients previously covered by private insurance use less
healthcare, and that, in response, physicians compensate by induc-
ing demand in the Medicare population. Another is that physicians
have more capacity to treat Medicare patients as demand by the
privately insured declines.
In this paper, we provide the rst physician-level examination
of the relationship between the business cycle and the provi-
sion of healthcare, and we investigate the role of both of these
supply-side explanations in explaining changes in inpatient hos-
pitalizations and outpatient care among Medicare beneciaries
during recessions. We use quarterly discharge data from hospi-
tals and ambulatory surgery centers (ASCs) in Florida to construct
separate physician-level measures of healthcare provided to Medi-
care patients and privately insured patients, where care provision
is measured by separate counts of inpatient and outpatient dis-
charges. We merge these physician-specic measures of care
provision to county-level quarterly unemployment rates and var-
ious controls. We nd that, on average, the amount of inpatient

http://dx.doi.org/10.1016/j.jhealeco.2014.12.008
0167-6296/ 2015 Elsevier B.V. All rights reserved.
98 D. He et al. / Journal of Health Economics 40 (2015) 97108
and outpatient care provided by physicians to privately insured
patients is procyclical. This pattern is most likely explained by the
fact that higher unemployment rates reduce both private health
insurance coverage (Cawley and Simon, 2005; Holahan, 2011) and
income, and thus dampen healthcare demand. In contrast, we nd
no evidence that care provided to Medicare patients is procyclical.
We then test for heterogeneous effects of local unemployment
rates by allowing the effects to vary by the physicians exposure
to changes in the unemployment rate. We dene exposure as
the share of the physicians patients with private insurance (i.e.,
patients likely to experience decreases in coverage or earnings)
in the year prior to our sample period. For both inpatient and
outpatient care, we nd statistically signicant increases in the
amount of care provided to Medicare patients among physicians
who are more exposed to the business cycle. These heterogeneous
responses are consistent with a form of physician-induced demand
in which physicians induce utilization among their continuously
insured patients to offset income losses, comparable to the type of
demand inducement examined in Gruber and Owings (1996).
We also test for the possibility that our results are explained
by easing capacity constraints. That is, higher unemployment rates
may free-up resources (e.g., time and facilities) and allow physi-
cians to meet pent-up demand by Medicare patients. To examine
this possibility, we test whether the patterns we observe are
limited to physicians who practice in areas that are capacity con-
strained. We dene capacity constraints in two ways, rst, using
the ratio of the population to the number of acute care beds in
the county, and second, using physician responses to the Commu-
nity Tracking Study (CTS) to identify areas where physicians report
an unwillingness to accept new Medicare patients. Although we
nd some evidence that easing capacity constraints may contribute
to the rise in Medicare volume, we also observe the same type of
heterogeneous effects of unemployment among physicians who
practice in areas not marked by capacity constraints (i.e., counties
with relatively low population-to-beds ratios or metropolitan areas
where large shares of physicians are willing to accept new Medi-
care patients). This result is consistent with physician induced
demand.
2. Relationship to the prior literature
While no prior study in the recessions and health literature
has examined physician-level data on healthcare provision, several
studies examine the relationship between the business cycle and
either individual-level utilization or aggregate state-level utiliza-
tion rates. Using data from the Behavioral Risk Factor Surveillance
System, Ruhm (2000) nds that state unemployment rates have
negative and in some cases signicant effects on routine medical
care use and receipt of preventive care, such as mammograms, pap
smears, and digital rectal exams. Lusardi et al. (2010) nd that more
than a quarter of 1865 year-old U.S. respondents to the TNS Global
Economic Crisis Survey reported a drop in routine medical care use
following the economic crisis of 200809. Using National Health
Interview Survey (NHIS) data, Ruhm (2003) nds a statistically sig-
nicant negative relationship between state unemployment rates
and the probability of having at least one hospital episode among
3055 year-old respondents.
The ndings above pertain to samples of, or heavily repre-
sented by, working-age adults; in contrast, several studies show
that, among the Medicare population, inpatient hospitalization
increases as local unemployment rates rise. Two studies do so using
state-level discharge rates for Medicare fee-for-service benecia-
ries from the Dartmouth Atlas of Health Care. Ruhm (2007) shows
that discharge rates for Medicare beneciaries undergoing certain
surgical treatments for heart disease increase with state-level
unemployment rates. McInerney and Mellor (2012b) nd that state
unemployment rates are positively and signicantly associated
with additional state-level measures of hospital utilization by
Medicare beneciaries (i.e., discharge rates for all medical and
surgical hospitalizations and end-of-life inpatient care). Evidence
using microdata from the Medicare Current Beneciary Survey
also conrms the countercyclical nature of inpatient care among
Medicare recipients (McInerney and Mellor, 2012a).
The present paper makes three contributions to the recessions
and health literature. First, we examine the relationship between
local economic conditions and physician-level measures of hospi-
tal care provision. To the best of our knowledge, no prior study in
this literature has examined healthcare provision at the level of
a provider. Second, we examine the amount of care provided to
Medicare patients and privately insured patients in separate mod-
els. This allows for a direct comparison of the effects of the local
unemployment rate on healthcare provision by payer. Prior stud-
ies look at the relationship between unemployment and healthcare
either in the Medicare population or in the adult population as
a whole; because these studies differ in the time period studied,
among other traits, the results are not directly comparable. Finally,
we explore some of the mechanisms behind the previously docu-
mented increase in inpatient care provided to Medicare patients as
unemployment rates rise. We use physician-level data to examine
two supply-side mechanisms in particular.
One mechanism is that economic downturns lead physicians to
induce demand among patients who retain health insurance. A the-
oretical basis for this mechanism is provided in Gruber and Owings
(1996), which presents a model where negative shocks to physician
incomes lead physicians to induce demand for services reimbursed
by third party payers at higher rates. Specically, their model posits
that negative shocks to obstetricians income brought on by declin-
ing fertility rates cause demand inducement of cesarean-section
deliveries. Their model is supported by empirical evidence linking
within-state decreases in fertility rates to within-state increases in
the likelihood of a cesarean-section delivery.
Rising unemployment rates may also represent a negative shock
to physician incomes. The job losses associated with rising unem-
ployment trigger reductions in private health insurance (Cawley
and Simon, 2005; Holahan, 2011), which in turn, decrease health-
care demand by workers who lose coverage. Even workers who
maintain coverage may be reluctant to take time off to obtain care
or may experience reductions in household income that dampen
demand. Together these responses are thought to contribute to the
slowing growth rate of private healthcare spending (Martin et al.,
2011; Hartman et al., 2010). As in the Gruber and Owings model,
physicians whose incomes are reduced by declining demand for
their services may induce demand, in this case by providing more
services to their remaining patients. Patients covered by Medi-
care are one such group. Although some Medicare beneciaries
could drop or lose supplemental coverage during economic down-
turns, Medicare eligibility is unaffected by the business cycle.1
Thus, evidence that higher unemployment rates reduce healthcare
provision to privately insured patients and increase healthcare pro-
vision to Medicare patients is consistent with physician induced
demand.
1
Demand for healthcare by the Medicare population could also decrease dur-
ing recessions, but we expect that the effect of economic downturns on healthcare
demand will be smaller among Medicare beneciaries than among the non-elderly
population, since seniors incomes are less closely tied to labor market earnings
than the incomes of non-elderly households, and since seniors primary insurance
coverage is not likely to be affected.
 D. He et al. / Journal of Health Economics 40 (2015) 97108
The second explanation we consider is that recessions ease
capacity constraints faced by physicians and allow them to treat
more Medicare patients. That is, as demand by the privately insured
falls during economic downturns, physicians have more time in
their schedules and space in their facilities to accommodate Medi-
care patients. If there is excess demand for healthcare by Medicare
patients, economic downturns may ease capacity constraints on
providers and thus enable them to increase the services provided
to the Medicare population.2 We elaborate on our empirical tests
for these alternate explanations in Section 4 below.
3. Data
We measure the amount of healthcare provided to patients by
individual physicians using inpatient and outpatient discharge data
from the Florida Agency for Health Care Administration (AHCA)
from 1998 through 2010. The data include the universe of dis-
charges from all reporting general short-term hospitals and ASCs
in the state. Florida is the fourth most populous state in the nation
and has more than 200 hospitals, so its size mitigates concerns
about the generalizability of our study. Changes in utilization in
Florida are also relevant to Medicare as a whole; Florida has the
second highest state Medicare spending and accounts for over 8%
of Medicare spending nationwide (CMS, 2007).
These data are well suited to our study because each Florida
discharge record includes a unique identier for the attending
physician.3 These identiers allow us to construct a panel of
physician-quarter-level measures of patient care provision over a
13-year period. National discharge data and discharge data from
other states do not allow researchers to link physician records
across years for as long a period of time. Changes in the algorithm
applied to physician identiers in the National Inpatient Sample
(NIS) and the Healthcare Cost and Utilization Project (HCUP) state
discharge databases prevent users from linking physicians before
and after 2003 (AHRQ, 2011a,b).4 Although the Florida AHCA also
changed the way physician identiers were recorded in 2005 for
outpatient discharges and 2006 for inpatient discharges, we were
able to address this issue in our study. Appendix A provides the
details.
The unit of observation in the raw data is a patient discharge;
in addition to the unique attending physician identication num-
ber, each discharge record also indicates the principal payer, the
facility identication code, the facility type and various diagno-
sis and procedure codes related to the patients treatment.5 We
use these data to construct, for each quarter, physician-specic
amounts of patient care provision measured as the number of
discharges. We construct separate measures of patient care pro-
vision depending on whether the discharge is covered by Medicare
or private insurance. We are especially interested in physician
2
Baker and Royalty (2000) provide some evidence of excess demand for health-
care by publicly insured patients; they nd that Medicaid eligibility expansions had
no signicant effect on the percent of poor patients seen by private physicians.
3
The data also include a unique identier for the operating physician. Over 80
percent of the physicians identied as operating physicians in both the inpatient
and outpatient setting also appear as attending physicians. As a robustness check,
we verify that the main qualitative conclusions are upheld when we perform the
same analysis for the two samples of operating physicians. Results are available
upon request.
4
We did not explore the availability of physician identiers in Medicare claims
because these data would not allow us to observe physician-level measures of care
provided to the privately insured.
5
Additional elds contain patient information (e.g., race/ethnicity, age, and sex)
and the charges for the discharge. Unfortunately our data do not include individual
patient identiers, so we are unable to determine whether or not a particular patient
receives more care over multiple visits as county unemployment rates rise.
99
responses under fee-for-service payment for two reasons. One is
that all of the prior studies of Medicare utilization in the reces-
sions and health literature focus on Medicare beneciaries with
traditional fee-for-service insurance. The second reason is that the
demand inducement explanation relies on the physicians ability
to raise additional income through additional volume, which is less
likely under the capitated or salary-based payment arrangements
in Health Maintenance Organizations (HMOs). For these reasons,
we include only those discharges where the primary payer was
either traditional FFS Medicare6 or a private plan that was either a
FFS or a PPO (preferred provider organization) plan, since PPO plans
typically reimburse on a discounted FFS basis. In all, we construct
four measures of care (i.e., care provided to Medicare and privately
insured patients in the inpatient and outpatient discharge records).
We restrict the physician samples to non-resident medical or
osteopathic physicians, and we exclude physicians whose practice
county does not remain the same across our entire sample period.
We dene a physicians practice county as the county where the
majority of the patients discharged by that physician reside, using
county of residence reported on individual discharge records.7
With this same-county restriction, the within-physician variation
in unemployment rates reects changing macroeconomic condi-
tions affecting a physicians existing patient base. We also exclude
physicians who never treat a Medicare patient in the sample period
because they are not affected by the types of supply-side mecha-
nisms we examine here.
Table 1 reports mean quarterly amounts of healthcare provided
by physicians to Medicare patients and the privately insured for
each of the two samples. The average physician discharges nearly
13 Medicare patients per quarter and nearly six patients with pri-
vate insurance per quarter from the hospital inpatient setting. As
with inpatient care, the amount of outpatient care provided to
Medicare patients also exceeds the amount provided to privately
insured patients, but the difference in the amount of outpatient
care provided to the two groups of patients is smaller.
We next construct quarterly unemployment rates for each
Florida county by averaging monthly county unemployment rate
data from the Bureau of Labor Statistics. Fig. 1 illustrates some
of the variation in quarterly county unemployment rates both
across the state and within a county over time. It depicts quar-
terly unemployment rates for the counties at the 10th, 25th, 50th,
75th, and 90th percentiles in the distribution of unemployment
rates in the rst quarter of 1998, the rst year of the estimation
samples. In 1998-Q1, the unemployment rate in Clay County was
3.2 (which is the 10th percentile of the unemployment rate distri-
bution); in comparison, the unemployment rate in Holmes County
was 6.8 (which is the 90th percentile of the distribution). Quarterly
changes in the unemployment rate have a mean absolute value of
0.6. Over our sample period, the mean county unemployment rate
was approximately 5.5, as shown in Table 1. We then merged the
physician-level measures of patient care provision to the county-
level unemployment rates using the physicians practice county
dened above.
6
Medicare HMO and Medicare PPO discharges cannot be counted separately
because the payer eld on the discharge record combines both types in the same
category.
7
Eighty-two percent of physicians practice in the same county from 1998 through
2010. For example, in the inpatient data, there are 41,364 attending physicians over-
all, and 7622 of them move to a different county at some point between 1998 and
2010. For the vast majority of the physicians who practice in the same county, 100%
of their patients reside in a single county. For example, across the entire period, 89%
of physicians have counts of inpatient care for patients from only one county and
84% of the physician-quarter observations have counts of inpatient care provided
to patients in only one county.
100 D. He et al. / Journal of Health Economics 40 (2015) 97108

Table 1
Means and standard deviations of dependent and key independent variables, by sample, 19982010.

Medicare patients Privately insured patients Unemployment rate Private share

Panel A: Inpatient care

12.82 5.78 5.47 0.33
(21.39) (9.35) (2.53) (0.21)
n = 649,460 n = 649,460 n = 649,460 n = 418,914
Panel B: Outpatient care
15.77 13.45 5.56 0.42
(41.74) (30.86) (2.59) (0.23)
n = 590,136 n = 590,136 n = 590,136 n = 299,396

Notes: Sample includes physicians who treat patients from the same Florida county for all periods they enter the sample, and who ever treat a Medicare FFS patient.
Discharges paid by Medicare MCOs or private HMOs are excluded from patient counts. Each discharge is counted separately in the patient counts. The variable private
share is constructed separately for each physician using discharge data for 1997, the year prior to the sample period used in the regression analysis. The private share is the
fraction of the total number of patients each physician treats in 1997 that hold private insurance (including private HMO). Physicians not observed in the discharge data in
1997 are not included in column 4 or in models that use the private share.

The dependent variable is the log of the discharge counts for

Fig. 1. Quarterly unemployment rates for selected Florida counties. Notes: The ve
counties depicted in the gure above represent the counties at the 10th (Clay
County), 25th (Collier County), 50th (Polk County), 75th (DeSoto County), and 90th
(Holmes County) percentiles in the distribution of quarterly unemployment rates
in the rst quarter of 1998. There are 68 counties in Florida; 60 counties are repre-
sented in the analysis presented in Tables 2 and 3.
We merge other county-year-level and hospital-year-level
variables to the physician-level observations. County-year-level
variables include estimates of the total population and the pro-
portions of the population age 65 and up, female, Hispanic, and
black.8 We merge these data to the physician-quarter datasets
again using the county in which the largest share of the physi-
cians patients resided in the period. We merge hospital-year traits,
including bed size, ownership status, teaching status, and statutory
rural hospital status,9 using the identier of the hospital in which
the largest share of the physicians patients were treated in the
period. Appendix Table A1 reports the summary statistics of the
main hospital-year-level and county-year-level variables used in
the analysis.
4. Empirical strategy
To examine how physicians provision of healthcare responds to
changes in the local unemployment rate, we estimate the following
model:
ln(patient care provisionit ) = 0 + 1 Unemct

+ Zct +Xht  + t + i + it (1)
8
With the exception of the proportion of the total population that is black, the
population data were obtained from the U.S. Census. Because Census reports pop-
ulation estimates by racial subgroup differently before 2000 and from 2000 on, we
obtained estimates of the black population from Florida Charts reports released by
the Florida Legislatures Ofce of Economic and Demographic Research.
9
We obtain these measures from January issues of Floridas Hospital Beds and Ser-
vices List for years 19992011 and we apply January data from each year to quarterly
observations from the prior year.
physician i in time period t where t ranges from 1998-Q1 to
2010-Q4. The main explanatory variable of interest is Unemct , the
quarterly unemployment rate in the county where the majority of
physician is patients reside. We control for annual county-specic
controls in vector Zct , hospital-specic controls in vector Xht , and
quarter-year xed effects (t ), and physician xed effects ( i ). The
parameter of interest is 1 , the coefcient of the local unemploy-
ment rate, which indicates how the average physician alters patient
care provision in response to rising unemployment. To examine
differential effects by payer, we estimate Eq. (1) separately for care
provided to Medicare patients and to privately insured patients.
To examine effects for different types of hospital care, we run this
pair of regressions separately for care provided in the inpatient and
outpatient settings.10,11
We next test whether physicians responses to changes in the
unemployment rate vary with their exposure to recessions. Fol-
lowing a standard approach in the demand inducement literature,
we measure each physicians exposure to changing nancial cir-
cumstances and interact that with the unemployment rate. In our
setting, exposure to the income losses from rising unemployment
would be greater for those physicians whose incomes are more
dependent on privately insured patients and smaller for those
physicians whose incomes depend more on patients covered by
public programs or those who pay out-of-pocket to begin with.
We thus dene the share of the physicians patients with private
insurance and estimate this model:
ln (patient care provision)it = 0 + 1 Unemct Sharei

+ 2 Unemct + Zct +Xht  + t + i + it (2)
where Sharei represents for each physician i the share of the total
number of patients for which private insurance was the payer, and
all other variables are the same as those used in Eq. (1).12 We con-
struct the share variable using discharge records from 1997, the
year prior to the sample period used in the regression analysis, and
10
A priori, it is theoretically ambiguous whether inpatient or outpatient care would
be more responsive to changes in unemployment rates. Patients may be more
responsive to inducement in the outpatient setting if the patients cost of treat-
ment is lower; alternatively a physician may be more prone to induce demand in
the inpatient setting since the return on inducement is greater.
11
Since we include ASC discharges in the counts in the outpatient setting, we do
not include controls for hospital characteristics in these models.
12
In constructing this measure, private insurance includes private HMOs in order
to allow for the possibility that patients with private HMO coverage may also lose
insurance and thus reduce healthcare demand, contributing to potential physician
income losses during economic downturns.
 D. He et al. / Journal of Health Economics 40 (2015) 97108
we construct separate shares for both physician samples.13 In the
inpatient setting, the mean share of patients with private insurance
is about one-third; the 10th percentile of the physician-quarter
distribution is 0.1 and the 90th percentile is 0.63. In the outpatient
setting, the mean share with private insurance is slightly higher,
above 0.4, and the 10th percentile is 0.11 and the 90th percentile
is 0.75. If highly exposed physicians respond to higher unemploy-
ment rates with larger increases in patient care provision, we would
expect estimates of 1 to be positive.
Finally, we consider whether our results t an alternative expla-
nation, in which declining demand by the privately insured releases
a binding capacity constraint originally faced by physicians, thus
allowing them to treat more Medicare patients. If this is the case,
we would expect that only capacity constrained doctors increase
healthcare provision to Medicare patients when the unemploy-
ment rate rises. We test for the impact of capacity constraints in
two ways. To measure capacity for inpatient care, we dene the
population-to-beds ratio in the county where the largest share
of the physicians patients reside using data from 1997. We then
construct several indicator variables based on whether the county
population-to-beds ratio is higher than a particular cutoff in the
distribution of the population-to-beds ratio (i.e., above the median
or the 75th, 90th, and 95th percentiles).14 We specify a three-way
interaction model that includes an interaction term between each
of these indicators of capacity constraint and our key explanatory
variable (the interaction of the unemployment rate with the pri-
vately insured share), as shown below:
ln(patient care provisionit ) = 0 + 1 Unemct Sharei
Capacity Constrainedc + 2 Unemct
Share + 3 Unemct
Capacity Constrainedi + 4 Unemct

+ Zct +Xht  + t + i + it (3)
where 1 is the coefcient on the three-way interaction term of the
county unemployment rate, the private share in the baseline year,
and an indicator for capacity constraints in the baseline year. This
specication also includes the two-way interaction between the
capacity constraint indicator and unemployment rate, in addition
to our key explanatory variable (the unemployment rate interacted
with the private share), and the unemployment rate alone.15
Because the population-to-beds ratio may partially reect the
health of the local population and does not describe capacity con-
straints in the outpatient setting, we also dene a second measure
using data from the Community Tracking Study (CTS)-Physician
Survey. This survey includes a question about how willing the
physicians practice is to accept new Medicare patients; the pos-
sible responses are accepting all, most, some, or no new patients.
Using data from the 1996/97 wave of the survey, we dene for each
Florida site the percent of practices that reported accepting no new
Medicare patients. We posit that the higher this percentage, the
more capacity constrained are physicians in the site. This measure
13
In the inpatient setting, almost two-thirds of the physician-quarter observations
in our data are associated with physicians who we observe in the 1997 discharge
data. In the outpatient setting, this share is just over 50%.
14
In 1997, the median population-to-beds ratio across Florida counties is 363
persons/bed; the 10th percentile is 259 persons/bed, the 75th percentile is 521 per-
sons/bed, the 90th percentile is 804 persons/bed, and the 95th percentile is 1002
persons/bed.
15
Measures of the private share, the capacity constraint indicator, and the inter-
action of those two variables are not included in the model since they do not vary
over time and because the model includes physician xed effects.
101
Table 2
Average effects of county unemployment rate on hospital care provision,
19982010.
Medicare patients Privately insured
patients
(1) (2)
Panel A: Inpatient care
All attending 0.003 0.003
physicians (0.002) (0.002)**
(n = 649,460) [0.009] [0.005]
Panel B: Outpatient care
All attending 0.002 0.005
physicians (0.002) (0.002)**
(n = 590,136) [0.009] [0.006]
Notes: Cells report the coefcient of the county-quarter unemployment rate in
regressions where the dependent variable is the log of the number of patients in
each category (i.e., Medicare patients or privately insured patients). Discharges paid
by Medicare MCOs or private HMOs are excluded from patient counts and each dis-
charge is counted separately in the patient counts. Standard errors in parentheses
are clustered by physician, and standard errors in brackets are clustered by county.
Inpatient regression models also include hospital characteristics for the hospital
where the physician practices the most (i.e., the number of acute beds and indi-
cators for whether the hospital is in a rural location, is a teaching hospital, and
is public or for-prot, where the omitted category is non-prot). Each regression
also includes controls for the following county characteristics: total county popu-
lation and proportions of county residents who are female, over age 65, Hispanic,
and black. Regression models also include year x quarter xed effects and physician
xed effects. Statistical signicance is indicated by * for signicance at the 10% level,
** for signicance at the 5% level, and *** for signicance at the 1% level.
is available for three Florida sites: Miami, Tampa, and West Palm
Beach. We use this continuous measure of capacity constraints in
the triple interaction specication in Eq. (3), and we estimate the
model using samples of physicians in these three Florida sites.
5. Results
Table 2 presents results from the estimation of Eq. (1). Each
cell contains the coefcient of the quarterly unemployment rate
in the physicians practice county.16 Column 1 reports results from
models of care provided to Medicare patients. We nd no evidence
that either inpatient or outpatient care provided to Medicare ben-
eciaries changes when the county unemployment rate rises. The
coefcient estimates are small and not statistically signicant. Col-
umn 2 reports results from models of care provided to privately
insured patients. Here, we nd that a one percentage point increase
in the unemployment rate is associated with statistically signicant
decreases in inpatient discharges (of 0.3 percent) and outpatient
discharges (of 0.5 percent). Thus, outpatient care provision to the
privately insured is more sensitive to rising unemployment rates
than inpatient care provision. The negative coefcients are con-
sistent with the negative effect of unemployment on incomes
and job-based coverage among the working-age population. These
ndings are also consistent with prior ndings that hospital utiliza-
tion falls among 3055 year-olds as state unemployment rates rise
(Ruhm, 2003), but because we focus exclusively on the privately
insured, results are not directly comparable.17
Table 3 reports results from estimating Eq. (2), which allows
the effect of unemployment to vary with the physicians share of
16
The full set of coefcient estimates for the inpatient care models is reported in
Appendix Table A2. The coefcient estimates for the control variables are similar in
the sample of physicians providing outpatient care; the full set of those estimates
is available upon request.
17
Ruhm (2003) found a small, negative relationship between rising state unem-
ployment rates and the likelihood of having at least one hospital episode in the past
year in a sample that pooled individuals with private insurance, public insurance,
and no insurance.
102 D. He et al. / Journal of Health Economics 40 (2015) 97108
Table 3
Effects of county unemployment rate on hospital care provision by physician share
of patients with private insurance in 1997, 19982010.
Medicare Privately insured
patients patients
(1) (2)
Panel A: Inpatient care
Attending physicians observed in 1997 (n = 418,914)
Unemployment * private share 0.060 0.013
(0.004)*** (0.004)***
[0.011]*** [0.009]
Unemployment rate 0.013 0.005
(0.002)*** (0.002)**
[0.010] [0.006]
Panel B: Outpatient care
Attending physicians observed in 1997 (n = 299,396)
Unemployment * private share 0.035 0.006
(0.004)*** (0.005)
[0.007]*** [0.007]
Unemployment rate 0.013 0.005
(0.004)*** (0.004)
[0.010] [0.006]
Notes: Standard errors in parentheses are clustered by physician, and standard
errors in brackets are clustered by county. Inpatient regression models also include
hospital characteristics for the hospital where the physician practices the most
(i.e., the number of acute beds and indicators for whether the hospital is in a
rural location, is a teaching hospital, and is public or for-prot, where the omit-
ted category is non-prot). Each regression also includes controls for the following
county characteristics: total county population and proportions of county residents
who are female, over age 65, Hispanic, and black. Regression models also include
year quarter xed effects and physician xed effects. Statistical signicance is indi-
cated by * for signicance at the 10% level, ** for signicance at the 5% level, and ***
for signicance at the 1% level.
privately insured patients in the baseline year.18 We report only
the coefcients of the unemployment rate and its interaction with
Share.19
Results from models of care provided to Medicare patients are
shown in column 1 of Table 3. The coefcients on the interaction
terms are positive and statistically signicant in both the inpatient
and outpatient settings. Thus, the amount of healthcare provided
to Medicare patients increases during recessions among physicians
who initially treat a higher share of privately insured patients. For
example, column 1, Panel A of Table 3 shows that a one percentage
point increase in the local unemployment rate corresponds to a
0.7 percent increase in the amount of inpatient care provided to
Medicare patients by physicians with the mean share of privately
insured patients (0.33).20 Overall, these results are consistent with
a model of physician-induced demand, in which those physicians
who are more exposed to economic downturns (and declines in
18
Note that the sample sizes in Table 3 are smaller than the sample sizes in Table 2.
This is because only those physicians who are observed in 1997 (the year we use to
dene the share of privately insured patients) are included in the estimation sample
for Table 3. Appendix Table A3 shows that the qualitative conclusions concerning
the average effect on the unemployment rate on hospital care provision (i.e., the
results from Table 2) are similar in the sample of physicians who are observed in
1997.
19
The full set of coefcient estimates for inpatient care is reported in Appendix
Table A2. The coefcient estimates for the control variables are similar in the out-
patient setting; the full sets of those estimates are available upon request.
20
Several factors may contribute to the unemployment rate effects being smaller
in the outpatient care model than in the inpatient care model. First, inpatient pro-
cedures are generally more expensive than outpatient procedures; thus, physicians
may have a greater incentive to induce demand for inpatient care. Second, Medi-
care coverage is more complete for inpatient care than for outpatient care; therefore,
demand for inpatient care may be more responsive to inducement. Third, inpatient
care often is considered more urgent than outpatient care; thus, again, demand for
inpatient care may be more responsive to inducement for this reason.
private insurance coverage) experience larger income effects and
a stronger incentive to induce demand among Medicare patients.
Two additional ndings from Table 3 are also worth noting.
First, the coefcient of the interaction term is positive and statis-
tically signicant in models of care provided to privately insured
patients. This positive coefcient suggests that for physicians with
relatively large shares of privately insured patients in the baseline
year, increases in unemployment are associated with increases in
care provision to the privately insured. The effect sizes are very
small relative to the Medicare volume effects. For example, the
expected effect of a one percentage point increase in the unemploy-
ment rate at the 75th (95th) percentile in private share is a 0.07 (0.5)
percent increase in inpatient care provided to privately insured
patients, but a 1.3 (3.4) percent increase in inpatient care provided
to Medicare patients. Within the context of demand inducement,
this implies that the positive income effect exceeds the negative
income shock from reduced demand among the privately insured;
we consider two alternate explanations for these positive effects at
the end of this section.
Second, the coefcient of the level term for unemployment is
always negative. The negative sign in the models of healthcare pro-
vided by the privately insured is as expected and consistent with
Table 2. The negative sign in models of Medicare patient care is
not observed in the average effects shown in Table 2, and suggests
that some physicians (those with relatively low shares of privately
insured patients) see reductions in the amount of care provided to
Medicare as unemployment rates rise. This effect may be explained
by the fact that Medicare beneciaries are not entirely insulated
from recession-timed reductions in earnings from work or sup-
plemental coverage and thus may also experience reductions in
healthcare demand as unemployment rates rise.
We conduct several checks on the robustness of the results
reported in Tables 2 and 3. First, we estimate models similar to
Eqs. (1) and (2) but where we also include county-specic linear
time trends. The results are similar to the results discussed above.
In results available upon request, we also show qualitatively sim-
ilar results when we relax the restriction that the physician treats
at least one Medicare patient.
We next consider the possibility that easing capacity constraints
accounts for the pattern of results in Table 3. Suppose that, ini-
tially, most physicians are working at or near capacity and cannot
accommodate all Medicare patients in their schedules. As the
local unemployment rate rises, the physicians with larger ini-
tial shares of privately insured patients would likely experience
larger releases in their capacity constraints as some previously pri-
vately insured patients reduce their healthcare consumption; these
physicians may be able to accommodate more Medicare patients
compared to physicians with a smaller initial share of privately
insured patients. Notably, however, this alternate explanation is
inconsistent with the positive coefcient of the interaction term for
the care provided to privately insured patients, because by deni-
tion easing capacity constraints arise from reductions in healthcare
provided to the privately insured.
To further examine whether easing capacity constraints could
plausibly explain the results shown in Table 3, we present esti-
mates of Eq. (3). By allowing the effect of the interaction between
the county unemployment rate and the private share to vary by
whether or not the area is capacity constrained, we can observe
if the effect observed in Table 3 is driven entirely by capacity
constrained areas. We construct a proxy for inpatient capacity con-
straints from the ratio of the county population to the total number
of acute beds in all short-term hospitals in the county in 1997. The
higher the population-to-beds ratio, the more likely a physician
practicing in the county is capacity-constrained. We dene capac-
ity constraints in four ways, based on whether a given countys
 D. He et al. / Journal of Health Economics 40 (2015) 97108 103

Table 4
Capacity constraints versus demand inducement. Sample is attending physicians providing inpatient care (n = 418,914).

Cap. constrained = 1 if 50th percentile 75th percentile 90th percentile 95th percentile
pop-to-beds ratio>

Medicare Privately insured Medicare Privately insured Medicare Privately insured Medicare Privately insured
patients patients patients patients patients patients patients patients
(1) (2) (3) (4) (5) (6) (7) (8)

Unemp. rate * private 0.019 0.014 0.010 0.010 0.018 0.032 2.577 0.175
share * cap. constrained = 1 (0.008)** (0.010) (0.010) (0.011) (0.021) (0.023) (0.468)*** (0.353)
[0.020] [0.015] [0.025] [0.015] [0.014] [0.011]*** [0.222]*** [0.269]

Unemp. rate * private share 0.073 0.022 0.058 0.014 0.060 0.014 0.060 0.013
(0.007)*** (0.008)*** (0.004)*** (0.005)*** (0.004)*** (0.004)*** (0.004)*** (0.004)***
[0.013]*** [0.010]** [0.013]*** [0.011] [0.011]*** [0.009] [0.011]*** [0.009]

Unemp. rate * cap. 0.009 0.012 0.009 0.004 0.020 0.000 0.520 0.022
constrained = 1 (0.004)** (0.004)*** (0.004)** (0.004) (0.010)* (0.009) (0.057)*** (0.038)
[0.011] [0.008] [0.010] [0.006] [0.007]*** [0.006] [0.024]*** [0.024]

Unemp. rate 0.019 0.014 0.011 0.004 0.012 0.005 0.013 0.005
(0.004)*** (0.004)*** (0.003)*** (0.003)* (0.003)*** (0.002)** (0.002)*** (0.002)**
[0.008]** [0.009] [0.011] [0.007] [0.010] [0.006] [0.010] [0.006]

Notes: See Table 3.

Table 5
Capacity constraints versus demand inducement: results from Eq. (2) using CTS data on % of practices accepting no new Medicare patients. Sample is Miami, Tampa and
West Palm Beach physicians.

Inpatient care (n = 83,805) Outpatient care (n = 123,990)

Medicare Privately insured Medicare Privately insured

patients patients patients patients
(1) (2) (3) (4)

Unemp. rate * private share * % 0.006 0.006 0.002 0.002

accepting no new Medicare patients (0.005) (0.006) (0.002) (0.002)
[0.005] [0.005] [0.001] [0.001]

Unemp. rate * private share 0.046 0.030 0.049 0.018

(0.028) (0.030) (0.019)** (0.021)
[0.025] [0.022] [0.019]* [0.013]

Unemp. rate * % accepting no new 0.004 0.006 0.001 0.002

Medicare patients (0.002)** (0.002)*** (0.001) (0.001)*
[0.002] [0.002]** [0.002] [0.001]

Unemp. rate 0.015 0.013 0.024 0.018

(0.016) (0.015) (0.011)** (0.012)
[0.018] [0.030] [0.019] [0.015]

Notes: See Table 3.

population-to-beds ratio is above the median, as well as the 75th,
90th, and 95th percentiles in the distribution.
Table 4 reports the results of estimating Eq. (3). We nd
some evidence that easing capacity constraints may contribute to
increased Medicare volume when we dene capacity constrained
counties as those two counties with population-to-beds ratios
above the 95th percentile. In this case, the coefcient on the triple
interaction term is large, positive, and statistically signicant, and
a one percentage point increase in the county unemployment
rate corresponds to a 33.6 percent increase in Medicare volume.
However, even in this setting, we continue to nd that unemploy-
ment increases Medicare volume among highly exposed physicians
working in counties that are not capacity constrained. This is similar
to what we document in Table 3: the coefcient on the interaction
between the county unemployment rate and the share of patients
with private insurance is 0.06, as in Table 3, and a one percent-
age point increase in the unemployment rate corresponds to a 0.7
percent increase in Medicare volume in areas that are not capacity
constrained. When we dene capacity constrained counties using
cutoffs at the 50th, 75th, or 90th percentiles, we nd fewer posi-
tive and statistically signicant coefcients on the triple interaction
terms, but in every case the coefcient estimates on the interaction
between unemployment and the share with private insurance are
similar to Table 3.21 Thus, we continue to nd evidence consistent
with physician demand inducement.
Of course, two concerns with the measure of capacity con-
straints presented above are that it does not apply to the outpatient
setting and that a county may have a low population-to-beds ratio
because the community is typically healthy, not because the area
is capacity constrained. Therefore, in Table 5, we present results
using a different measure of capacity constraints: a metropolitan
areas mean percentage of physicians who report their practice
is unwilling to accept new Medicare patients.22 In both models
of Medicare inpatient and outpatient discharges, the estimated
coefcient on the triple interaction term is not statistically signif-
icant but the coefcient estimate on the interaction between the
21
As shown in Appendix Table A5, when we include a continuous measure of
capacity constraints, we continue to nd positive and statistically signicant coef-
cient estimates for the interaction of the unemployment rate and the private share.
22
We have conrmed that our baseline results from our main specications (those
shown in Tables 2 and 3) remain similar when we use these three sites alone. See
Appendix Tables A3 and A4.
104 D. He et al. / Journal of Health Economics 40 (2015) 97108

Table 6
Effects of county unemployment rate by physician share of patients with private insurance, results by quarter.

Quarter 1 Quarter 2 Quarter 3 Quarter 4

Medicare Privately insured Medicare Privately insured Medicare Privately insured Medicare Privately insured
patients patients patients patients patients patients patients patients
(1) (2) (3) (4) (5) (6) (7) (8)

Panel A: Attending physician inpatient care

Unemp. rate * private share 0.063 0.017 0.067 0.017 0.053 0.011 0.062 0.005
(0.005)*** (0.006)*** (0.005)*** (0.006)*** (0.004)*** (0.005)** (0.005)*** (0.005)
***
[0.008] [0.012] [0.008]*** [0.011] [0.017]*** [0.007] [0.009]*** [0.009]
Unemp. rate 0.025 0.001 0.016 0.008 0.008 0.008 0.018 0.007
(0.004)*** (0.004) (0.005)*** (0.005)* (0.003)** (0.003)** (0.004)*** (0.004)*
[0.006]*** [0.010] [0.006]** [0.012] [0.009] [0.009] [0.008]** [0.009]

Panel B: Attending physician outpatient care

Unemp. rate * private share 0.038 0.008 0.041 0.001 0.032 0.011 0.034 0.006
(0.006)*** (0.006) (0.006)*** (0.006) (0.005)*** (0.006)* (0.005)*** (0.006)
[0.008]*** [0.010] [0.009]*** [0.010] [0.008]*** [0.009] [0.007]*** [0.007]
Unemp. rate 0.013 0.007 0.013 0.001 0.016 0.004 0.012 0.008
(0.006)** (0.006) (0.006)** (0.006) (0.005)*** (0.005) (0.005)** (0.005)
[0.012] [0.010] [0.013] [0.012] [0.012] [0.007] [0.010] [0.009]

Notes: See Table 3.

Table 7
Effect of unemployment rate on the proportion of discharges paid by Medicare.

Inpatient Medicare proportion Outpatient Medicare proportion

Discharges of persons Discharges of persons Discharges of persons Discharges of persons

age 65+ all ages age 65+ all ages

Coefcient on 0.002 0.003 0.002 0.007***

unemployment rate (0.002) (0.002) (0.003) (0.004)
Sample mean of dep. 0.94 0.58 0.96 0.72
variable
# of observations 3275 3275 2732 2732

Notes: Regression models also include mean hospital characteristics in the county (i.e., the mean number of acute beds at hospitals in the county and the proportion of
hospitals that are rural, teaching hospitals, public, and for-prot in the county). Each regression also includes controls for the following county characteristics: total county
population and proportions of county residents who are female, over age 65, Hispanic, and black. Regression models also include year quarter xed effects and county
xed effects. Statistical signicance is indicated by * for signicance at the 10% level, ** for signicance at the 5% level, and *** for signicance at the 1% level.

unemployment rate and share of patients with private insurance
is positive (though estimated with less precision in these smaller
samples). In the outpatient setting, the average physician in an area
where physicians are less willing to accept new Medicare patients
experiences a greater increase in Medicare volume in response to
an increase in the unemployment rate compared to physicians in an
area more accepting of new Medicare patients. While this suggests
that easing capacity constraints explains some of the increased
Medicare volume, we continue to nd evidence consistent with
demand inducement in the least capacity constrained areas.
We next examine two other alternate explanations for the
increased Medicare volume during recessions. In Florida, the popu-
lation of retired individuals increases during the winter months (at
least in certain counties). These relocated retirees are often called
snowbirds. If the unemployment rate is higher during winter
months, and the number of Medicare recipients in the state is also
higher during the winter months, then we may merely be pick-
ing up the effect of snowbirds relocating to Florida for part of the
year. To address this concern, we follow the same approach used
by Doyle (2011) and estimate Eq. (2) in separate samples quarter-
by-quarter. If snowbirds were driving our main results, we would
see different patterns in the results using observations from quar-
ters 1 and 4 as compared to using observations from quarters 2
and 3. Table 6 shows that our ndings are not driven by snowbirds.
Instead, the coefcient estimates for our key explanatory variable,
the unemployment rate interacted with the share of patients cov-
ered by private insurance at baseline, are very similar in magnitude
across all four quarters for both the Medicare patient models and
private patient models and always statistically signicant in the
Medicare patient models.
We also consider a second potential explanation for our ndings
in which higher unemployment rates lead Medicare beneciaries
to leave the workforce, in turn leading to increased discharges
where Medicare, instead of employer-sponsored private insur-
ance, becomes the primary payer. We directly test this explanation
using the Florida discharge records. We use inpatient discharges of
persons over age 65 where the primary payer was either Medi-
care fee-for-service or private insurance, and for each county
and quarter-year, we dene the proportion of those discharges
in which Medicare was the primary payer. We similarly dene
this Medicare proportion variable using the outpatient discharge
records. For robustness, we also obtain a variant of both meas-
ures without applying the 65 and older restriction; this allows
for the possibility that in recessions some privately insured shift
to Medicare even if they are below age 65. Thus, all together, we
dene four Medicare proportion variables (one for inpatient dis-
charges, one for outpatient discharges, and then with or without
the age 65 and older restriction). We then regress each of these four
Medicare proportion variables on the unemployment rate using Eq.
(4) below:
Medicare proportionct = 0 + 1 Unemct

+ Zct +t + c + ct (4)
 D. He et al. / Journal of Health Economics 40 (2015) 97108
where  c represents county xed effects and the county traits in
Z also include the average number of beds and the proportion of
hospitals by ownership and teaching status. The results from this
exercise are presented in Table 7. The unemployment rate coef-
cient estimates are very small, especially relative to the mean of
the dependent variables, and three of the four estimates are statisti-
cally insignicant; in the case of the one signicant estimate (in the
model of outpatient discharges among persons of all ages), the sign
is negative, meaning that the Medicare share is inversely related to
the unemployment rate. Thus, the shifting of elderly adults from
private insurance to Medicare does not seem to drive the main
results in our paper.
Finally, we return to the issue of the positive and signicant
effect of the unemployment rate on care provided to privately
insured patients among physicians who treat large shares of these
patients, and we consider two explanations for this other than
demand inducement. First, perhaps during economic downturns
patients of these doctors undergo more procedures in an ASC or
hospital outpatient department but fewer procedures in an ofce-
based setting; this, however, would affect only outpatient care.
Alternatively, if some physicians no longer treat any patients dur-
ing economic downturns, and their former patients turn to other
physicians who remain in practice (and who happen to treat large
shares of privately insured patients at baseline), this could also
explain our results in both inpatient and outpatient care mod-
els.
We cannot observe patient switching directly, but we look
for indirect support for this second scenario by estimating the
likelihood that a physician no longer treats patients. We use the
physicians included in the analytical samples in Table 3 and dene
an indicator variable equal to 1 if the physician does not treat pri-
vately insured patients in a given quarter, and 0 otherwise.23 We
regress that indicator on the county-level unemployment rate and
its interaction with the physicians baseline share of patients cov-
ered by private insurance. Support for this alternate explanation
would be seen in a positive signicant coefcient on the unem-
ployment rate (i.e., physicians with no or low private share patients
are more likely to stop treating patients during downturns) and
a negative signicant coefcient on the interaction term (i.e., the
effect of the unemployment rate on exit is smaller among high
private share physicians). Results from this exercise are reported
in Table 8 (column 2) and indirectly support the alternate explana-
tion for inpatient and outpatient care provided to privately insured
patients. We then do the same exercise for Medicare patients where
we dene an indicator variable equal to 1 if the physician does
not treat Medicare patients in a given quarter, and 0 otherwise.
Interestingly, we see a very different pattern for Medicare inpatient
care increases in the unemployment rate decrease the probabil-
ity a physician no longer treats Medicare patients, but less so for
physicians with high shares of private patients. These ndings lend
additional support for demand inducement as an explanation for
the Medicare inpatient results in Table 3, although the opposite
seems to be true for the Medicare outpatient results in the same
table.
23
We now add to the sample those quarters in which a physician has no dis-
charges, in the following way. We identify the rst quarter in which a physician
has any discharges, and we also identify the last quarter in which a physician has
any discharges. We add to the sample all quarters in between the physicians rst
observed practice quarter and one quarter after their last practice quarter (for physi-
cians who are observed to exit the sample during the period of our analysis). We
assign a value of 1 (exit) to these newly added quarters. Thus, the sample size
increases from 418,914 to 458,903 in the inpatient sample and from 299,396 to
328,100 in the outpatient sample.
105
Table 8
Models of the likelihood a physician treats no patients.
No Medicare No privately
patients insured patients
(1) (2)
Panel A: Inpatient care
Attending physicians observed in 1997 (n = 458,903)
Unemployment * private share 0.004 0.014
(0.002)** (0.002)***
[0.002]* [0.004]***
Unemployment rate 0.005 0.004
(0.001)*** (0.001)***
[0.002]*** [0.002]**
Panel B: Outpatient care
Attending physicians observed in 1997 (n = 328,100)
Unemployment * private share 0.004 0.003
(0.002)** (0.002)*
[0.002]* [0.002]
Unemployment rate 0.004 0.004
(0.001)*** (0.001)***
[0.002]* [0.002]**
Notes: Standard errors in parentheses are clustered by physician, and standard errors
in brackets are clustered by county. Each regression also includes controls for the
following county characteristics: total county population and proportions of county
residents who are female, over age 65, Hispanic, and black. Regression models also
include year quarter xed effects and physician xed effects. Statistical signi-
cance is indicated by * for signicance at the 10% level, ** for signicance at the 5%
level, and *** for signicance at the 1% level.
6. Conclusion
In this paper, we add to the recessions and health literature by
providing the rst physician-level examination of the relationship
between the business cycle and the provision of healthcare. Using
data on Florida physicians and examining changes in county-level
unemployment rates, we show that, on average, inpatient and out-
patient care provided to privately insured patients declines as local
unemployment rates rise. This pattern is consistent with reports
that private health insurance coverage falls during recessions, and
that some types of healthcare utilization by the overall popula-
tion decline as state unemployment rates rise. We nd no evidence
that the average physician provides less inpatient care to Medicare
fee-for-service patients as unemployment rates rise.
When we allow for heterogeneous responses across physicians,
we nd that rising unemployment rates increase Medicare inpa-
tient care provision by physicians who treat a greater share of
patients with private insurance. This nding is consistent with
models of physician induced demand, since those physicians with
large shares of privately insured patients are likely to experience
the large negative income shocks from economic downturns. We
also show that even though easing capacity constraints may explain
some of the increased Medicare volume during recessions, we con-
tinue to nd evidence consistent with physician induced demand.
These ndings add to the recessions and health literature by offer-
ing an explanation for prior ndings that hospital utilization by
Medicare beneciaries is positively associated with the unemploy-
ment rate (Ruhm, 2007; McInerney and Mellor, 2012a,b).
Our ndings of physician-level changes in the treatment of
Medicare patients also add to the prior literature on provider
incentives. These include empirical evidence of physician induced
demand in response to fee cuts for so-called overpriced or over-
valued Medicare procedures targeted by Congress in the Omnibus
Budget Reconciliation Act of 1987. Rice et al. (1999) shows that
fee reductions increased the volume of services paid by private
insurers and Yip (1998) nds that thoracic surgeons responded
to a large fee reduction by increasing the volumes of both Medi-
care and private pay procedures. In addition, He and Mellor (2012)
106 D. He et al. / Journal of Health Economics 40 (2015) 97108
nd that, consistent with provider demand inducement, private
FFS volume in hospital outpatient departments increased as Medi-
care payments for hospital outpatient services were cut under the
Outpatient Prospective Payment System (OPPS). Prior research by
Card et al. (2008) shows that eligibility for Medicare coincides
with increased hospitalization for non-urgent, high-cost proce-
dures, such as bypass surgery and hip and knee replacement. The
fact that these increases were not observed in hospitals that paid
physicians on a salaried basis led to the conclusion that discontinu-
ities in healthcare at Medicare onset are driven by an interaction
between supply-side incentives and shifts in insurance character-
istics (Card et al., 2008, p. 2243).
Our work also informs efforts to contain healthcare costs. Ris-
ing costs pose a major challenge to the goal of affordable health
care coverage, particularly for publicly funded programs like Medi-
care. Developing payment arrangements that counter incentives
for unnecessary care is one of the goals of the Affordable Care
Act (ACA) of 2010. The eventual savings from new arrangements
depends on many factors including how much excess utiliza-
tion occurs under fee-for-service. Some suggest that geographic
variation in healthcare spending and utilization is evidence that
wasteful spending arises from incentives created by fee-for-service
(e.g., Fisher et al., 2009). The evidence in this paper is consistent
with physicians responding to recession-induced income shocks
by increasing the amount of healthcare provided to Medicare FFS
patients. Whether this recession-induced increased spending is
wasteful depends on its effects on Medicare beneciaries health
status. Recent work shows that, in some settings, changes in health-
care provision that arise during recessions are health-improving.
Stevens et al. (2011) show that elderly mortality declines dur-
ing recessions likely because of quality improvements in nursing
homes. Thus an important avenue for future research would be to
examine whether the recession-induced increases in Medicare uti-
lization are timed with improvements in inpatient quality, patient
safety, and patient outcomes.
Appendix A. Addressing changes in FL discharge data
physician ID
As discussed in the text, the Florida AHCA changed the way that
the physician identications (ID) numbers were coded in 2005 (for
outpatient discharges) and 2006 (for inpatient discharges). We con-
tacted Florida AHCA staff to obtain additional information on the
nature of the recoding. We learned that prior to 2006 in the inpa-
tient data (and 2005 in the outpatient data), the AHCA padded
the physician IDs with leading zeros so that resulting IDs met a
Table A1
Means of hospital and county characteristics, by sample.
Acute beds
Panel A: Hospital Characteristics
Inpatient care
Attending physician sample (n = 649,460) 368.68
(214.82)
Total
population
Panel B: County Characteristics
Inpatient care
Attending physician sample (n = 649,460) 1,020,254
(741,624)
Outpatient care
Attending physician sample (n = 590,136) 1,013,190
(719,914)
Notes: See Table 1.
specic minimum eld length. Starting in 2006 for inpatient data
(and 2005 for outpatient data), the IDs were reported without lead-
ing zeroes (personal communication with AHCA staff, 2011). We
therefore removed leading zeros from the padded ID numbers and
created a physician ID that is uniformly reported in all the years
before and after the recoding.
We performed several tests to validate the ID transformation.
In the rst test, we examined the year-by-year attrition rates for
physician IDs in both the inpatient and outpatient data. For exam-
ple, in the inpatient data, 75.9% of the 2007 attending physician
IDs appear in 2008 and 76.3% of the 2004 attending IDs appear in
2005; both sets of years are years in which the IDs were reported the
same way. A very similar 76.0% of the transformed 2005 attending
physician IDs are found in 2006, the rst year when the new coding
practice started. This similarity in the attrition rates is observed for
the three other datasets as well (outpatient attending physicians,
inpatient operating physicians, and outpatient operating physi-
cians).
We also performed a second test for each physician in the inpa-
tient data. We identied the county and the hospital from which the
physician discharged the most patients (which we refer to below
as the most common county and the most common hospital,
respectively) and the most common DRG associated with the physi-
cians discharges. We then quantied the fraction of physicians
for whom the most common county was the same county in two
consecutive years, the fraction of physicians for whom the most
common hospital was the same hospital in two consecutive years,
and the fraction of physicians with the same most common DRG in
two consecutive years; we call each fraction an overlap rate. We
calculated these overlap rates across two consecutive years before
the recoding (i.e., in 2003 and 2004) and two consecutive years
after the recoding (i.e., in 2007 and 2008), all years when the IDs
were consistently coded. For the most common DRG associated
with a physicians discharges, we examined the overlap rate across
only 2003 and 2004 because CMS adopted the MS-DRG system in
2008, thus the DRG codes in 2007 and 2008 are not comparable.
The overlap rates in these years (all years with consistent coding)
then served as benchmarks against which we compared the over-
lap rates across 2005 and 2006 (the year immediately before the
change and the rst year with the change in effect) using the trans-
formed ID. Again, we found a very similar overlap rate across the
comparison periods for each of the three variables (i.e., the most
common hospital, the most common county, and the most com-
mon DRG) for both the attending and operating physicians. This
test further conrmed that our transformed IDs correctly identify
the same physician.
Rural status Public For-prot Teaching
0.02 0.13 0.34 0.09
Share female Share 65 Share Hispanic Share Black
0.51 0.17 0.20 0.16
0.51 0.18 0.19 0.16
 D. He et al. / Journal of Health Economics 40 (2015) 97108 107

Table A2
Full set of covariates for inpatient analysis shown in Tables 2 and 3.

Table 2 Table 3

Medicare patients Privately insured Medicare patients Privately insured

Unemployment rate * private share 0.060 0.013

(0.004)*** (0.004)***
[0.011]*** [0.009]
Unemployment rate 0.003 0.003 0.013 0.005
(0.002) (0.002)* (0.002)*** (0.002)**
[0.009] [0.005] [0.010] [0.006]
Acute beds (in 100s) 0.006 0.005 0.001 0.008
(0.003)** (0.003)** (0.003) (0.003)**
[0.005] [0.004] [0.006] [0.005]
Rural designation 0.020 0.021 0.053 0.044
(0.048) (0.048) (0.065) (0.065)
[0.073] [0.040] [0.098] [0.042]
Public 0.053 0.064 0.054 0.047
(0.016)*** (0.015)*** (0.019)*** (0.018)**
[0.026]** [0.029]** [0.026]** [0.020]**
For-prot 0.014 0.072 0.018 0.064
(0.010) (0.009)*** (0.012) (0.011)***
[0.020] [0.022]*** [0.023] [0.025]**
Teaching hospital 0.031 0.105 0.027 0.127
(0.012)*** (0.012)*** (0.016)* (0.016)***
[0.029] [0.035]*** [0.024] [0.038]***
Total population (in 1000s) 0.0003 0.0007 0.0002 0.0007
(0.0001)** (0.0001)*** (0.0001) (0.0001)***
[0.0004] [0.0004]** [0.0003] [0.0003]**
Proportion female 4.422 5.232 3.024 6.672
(2.604)* (1.948)*** (3.080) (2.336)***
[5.624] [5.143] [5.758] [5.605]
Proportion over age 65 1.177 1.213 1.108 1.512
(0.551)** (0.474)** (0.615)* (0.547)***
[1.020] [1.500] [1.014] [1.848]
Proportion Hispanic 2.497 2.638 2.298 2.562
(0.545)*** (0.449)*** (0.597)*** (0.505)***
[1.515] [1.337]* [1.273]* [1.179]**
Proportion Black 1.848 1.157 2.361 1.329
(0.450)*** (0.414)*** (0.502)*** (0.475)***
[1.323] [0.951] [1.253]* [0.972]
Sample size 649,460 418,914

Notes: See Tables 2 and 3.

Table A3 Table A4
Average effects of county unemployment rate, 19982010. Samples used in Effects of county unemployment rate on hospital care provision by physician share
Tables 3 and 5. of patients with private insurance in 1997, 19982010. Samples used in Table 5.

Medicare patients Privately insured Medicare Privately

patients patients insured
(1) (2) patients
(1) (2)
Sample used in Table 3
Inpatient attending 0.006 0.001 Inpatient attending physician sample (n = 83,805)
physician sample (0.002)*** (0.002) Unemp. rate * private share 0.083 0.011
(n = 418,914) [0.008] [0.005] (0.009)*** (0.010)
Outpatient attending 0.002 0.003 [0.013]*** [0.008]
physician sample (0.003) (0.003) Unemp. rate 0.009 0.052
(n = 299,396) [0.010] [0.006] (0.011) (0.010)***
[0.013] [0.028]
Sample used in Table 5
Inpatient attending 0.014 0.049 Outpatient attending physician sample (n = 123,990)
physician sample (0.011) (0.010)*** Unemp. rate * private share 0.026 0.000
(n = 83,805) [0.013] [0.026] (0.006)*** (0.007)
Outpatient attending 0.025 0.002 [0.010]** [0.006]
physician sample (0.003)*** (0.003) Unemp. rate 0.014 0.002
(n = 123,990) [0.010]* [0.006] (0.004)*** (0.004)
[0.008] [0.005]
Notes: See Table 3.
Notes: See Table 3.
108 D. He et al. / Journal of Health Economics 40 (2015) 97108

Table A5
Capacity constraints versus demand inducement: continuous measures of capacity constraints. Attending physician inpatient care sample (n = 418,914).

Capacity measure County percentile in distribution of pop/beds ratio (/100) County pop/beds ratio (/1000)

Medicare patients Privately insured Medicare patients Privately insured

patients patients
(1) (2) (3) (4)

Unemp. rate * private 0.026 0.017 0.034 0.016

share * capacity measure (0.015)* (0.018) (0.028) (0.024)
[0.025] [0.024] [0.031] [0.036]

Unemp. rate * private share 0.074 0.022 0.067 0.028

(0.009)*** (0.011)** (0.011)*** (0.013)*
[0.014]*** [0.014] [0.022]*** [0.018]

Unemp. rate * capacity measure 0.009 0.012 0.006 0.008

(0.007) (0.007)* (0.011) (0.010)
[0.015] [0.014] [0.020] [0.016]

Unemp. rate 0.018 0.012 0.010 0.009

(0.005)*** (0.005)** (0.005)* (0.005)*
[0.012] [0.010] [0.016] [0.010]

Notes: See Table 3.

References
AHRQ (Agency for Healthcare Research and Quality), 2011a. Introduction to the
HCUP Nationwide Inpatient Sample (NIS), 2009. Agency for Healthcare Research
and Quality, Rockville (MD).
AHRQ (Agency for Healthcare Research and Quality), 2011b. HCUP: NIS Description
of Data Elements. Agency for Healthcare Research and Quality, Rockville (MD).
Baker, L.C., Royalty, A.B., 2000. Medicaid policy, physician behavior, and health
care for the low-income population. Journal of Human Resources 35 (3),
480502.
Card, D., Dobkin, C., Maestas, N., 2008. The impact of nearly universal insurance cov-
erage on health care utilization: evidence from Medicare. American Economic
Review 98 (5), 22422258.
Cawley, J., Simon, K.I., 2005. Health insurance coverage and the macroeconomy.
Journal of Health Economics 24, 299315.
CMS (Centers for Medicare and Medicaid Services), 2007. Health Expenditure Data,
Health Expenditures by State of Residence (September 2007), Available at
http://www.cms.hhs.gov/NationalHealthExpendData/ (accessed 26.05.11).
Doyle, J.J., 2011. Returns to local-area health care spending: evidence from health
shocks to patients far from home. American Economic Journal: Applied Eco-
nomics 3 (3), 221243.
Fisher, E., Goodman, D., Skinner, J., Bronner, K., 2009. Health Care Spending, Quality,
and Outcomes: More Isnt Always Better. Topic Brief. The Dartmouth Insti-
tute for Health Policy and Clinical Practice: Dartmouth Atlas Project, Available at
http://www.dartmouthatlas.org/downloads/reports/Spending Brief 022709.pdf
Gruber, J., Owings, M., 1996. Physician nancial incentives and cesarean section
delivery. RAND Journal of Economics 27 (1), 99123.
Hartman, M., Martin, A., Nuccio, O., Catlin, A., the National Health Expenditure
Accounts Team, 2010. Health spending growth at a historic low in 2008. Health
Affairs (Millwood) 29 (1), 147155.
He, D., Mellor, J.M., 2012. Hospital responses to Medicares Outpatient Prospective
Payment System: evidence from Florida. Journal of Health Economics 31 (5),
730743.
Holahan, J., 2011. The 200709 recession and health insurance coverage. Health
Affairs (Millwood) 30 (1), 145152.
Lusardi, A., Schneider, D.J., Tufano, P., 2010. The Economic Crisis and Medical Care
Usage. NBER Working Paper 15843, Cambridge, MA.
Martin, A., Lassman, D., Whittle, L., Catlin, A., the National Health Expenditure
Accounts Team, 2011. Recession contributes to the slowest annual rate of
increase in ve decades. Health Affairs (Millwood) 30 (1), 1122.
McInerney, M., Mellor, J., 2012a. Recessions and seniors health, health behaviors,
and healthcare use: analysis of the Medicare Current Beneciary Survey. Journal
of Health Economics 31, 744751.
McInerney, M., Mellor, J., 2012b. State unemployment during recessions from
19912009 was linked to faster growth in Medicare spending. Health Affairs
(Millwood) 31 (11), 24642473.
Rice, T.H., Stearns, S.C., Pathman, D.E., DesHarnais, S., Brasure, M., Tai-Seale, M., 1999.
A tale of two bounties: the impact of competing fees on physician behavior.
Journal of Health Politics, Policy, and Law 24 (6), 13071330.
Ruhm, C., 2000. Are recessions good for your health? Quarterly Journal of Economics
115 (2), 617650.
Ruhm, C., 2003. Good times make you sick. Journal of Health Economics 22 (4),
637658.
Ruhm, C., 2007. A healthy economy can break your heart. Demography 44 (2),
829848.
Stevens, A.H., Miller, D.L., Page, M.E., Filipski, M., 2011. The Best of Times, the
Worst of Times: Understanding Pro-Cyclical Mortality. NBER Working Paper
No. 17657, Cambridge, MA.
Yip, W.C., 1998. Physician response to Medicare fee reductions: changes in the
volume of coronary artery bypass graft (CABG) surgeries in the Medicare and
private sectors. Journal of Health Economics 17, 675699.
 Journal of Health Economics 40 (2015) 109121

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

Who joins the network? Physicians resistance to take budgetary

co-responsibility
Maurus Rischatsch
Department of Economics, University of Zurich, Switzerland

a r t i c l e i n f o a b s t r a c t

Article history: Managed Care (MC) is expected to provide health care at a lower cost than conventional provision. There-
Received 15 September 2013 fore, Switzerland intends to promote MC by forcing health insurers to write MC contracts and introducing
Received in revised form 1 December 2014 budgetary co-responsibility for ambulatory care physicians. A discrete choice experiment conducted in
Accepted 6 December 2014
2011 including 872 physicians reveals a strong preference heterogeneity with respect to network partici-
Available online 15 December 2014
pation and alternative remuneration schemes. The number of physicians working in networks is unlikely
to rise on a voluntary basis, while general practitioners are more likely to join networks than special-
JEL classication:
ists with surgical activities. For physicians considering joining networks, cost savings are predicted to be
C5
D8
higher than the estimated willingness-to-accept payments.
I1 2014 Elsevier B.V. All rights reserved.

Keywords:
Physician reimbursement
Managed Care
Physician networks
Latent-class logit
Discrete choice experiment

1. Introduction
Over the last decade health care expenditure in most OECD
countries has grown at a faster rate than the gross domestic prod-
uct (GDP). Even if the national health care systems are organized
differently, most countries face similar challenges. Firstly, there is
a rising demand for medical services. On the one hand, an aging
society and other demographical changes lead to more people
in need of medical treatments and the number of patients with
chronic diseases or co-morbidities is on the rise. On the other
hand, additional demand originates from the coverage of uninsured
citizens. A prominent example for the latter is the United States
under the Affordable Care Act. Secondly, the number of innovative,
but expensive therapies has increased. As a consequence, health
care expenditure measured as a share of GDP increased to 17% in
the United States in 2012, the highest share worldwide accord-
ing to statistics from the OECD (2013). By comparison, Western
European countries typically spent between 10 and 12% on health
care. Only the Netherlands, France and Germany paid a higher
share than Switzerland, where health insurance premiums have
increased by 3.6% per year over the last decade according to statis-
tics from the Federal Ofce of Public Health.
In the aftermath of the nancial crisis and in the wake of the
economic downturns, nancing health care became an issue for
many governments. An important contribution to lower health
care costs is expected through a paradigm shift in physician reim-
bursement from fee for service to capitation. Under conventional
fee for service, medical services are remunerated according to an
administrated fee schedule, while under capitation a prospective
payment is paid to the health care provider, e.g. a payment per
enrollee per month. Transferring more cost responsibility to health
care providers is seen as one way to reduce health care expen-
diture due to the avoidance of unnecessary medical treatments.
For this reason, the so-called Alternative Quality Contract (AQC)
was designed in the United States and is currently used by the
Blue Cross Blue Shield of Massachusetts. Already more than two
decades ago, the United Kingdom introduced its GP fundholding
allowing larger general practices to work under a global budget
for hospital referrals with the possibility to retain a surplus as a
bonus.1

1
These two forms of risk sharing are discussed in more detail in the literature
E-mail address: maurus.rischatsch@econ.uzh.ch review.

http://dx.doi.org/10.1016/j.jhealeco.2014.12.002
0167-6296/ 2014 Elsevier B.V. All rights reserved.
110 M. Rischatsch / Journal of Health Economics 40 (2015) 109121

Switzerland intends to introduce similar cost incentives for target and not an actual payment to the organization. The medical
health care providers through the encouragement of Managed Care services are reimbursed through conventional fee for service, but
(MC). MC is understood as a model to introduce incentives for the global budget (called spending target in the remainder of this
health care providers and patients with the goal to increase treat- article) is used to introduce a bonus/malus system as discussed in
ment quality and to reduce health care expenditure. While MC is Section 3.
the dominant form of health insurance in the United States, it is
less established in Europe. Even if Switzerland was the rst Euro-
1.2. Health care reform
pean country to allow MC contracts in its social health insurance in
the 1990s, as discussed in Beck et al. (2009), the share of capitation
In 2012, Switzerland held a referendum with the intention
policies has remained low (5.5% in 2011). Although half the popu-
to increase the quality and the efciency of the health care sys-
lation signed a insurance policy accepting some kind of restriction
tem mainly through a better cooperation and coordination among
in return for a lower premium, the promotion of MC remains on the
health care providers. The government aimed at encouraging the
political agenda because several studies have shown that MC leads
nationwide development of MC networks. Among other changes,
to lower health care expenditure (e.g. Berchtold and Hess, 2006;
the legislative proposal encompassed that health insurers have to
Beck et al., 2009; Reich et al., 2012).
sign contracts with physician networks to govern their coopera-
tion, data exchange, quality assurance, and the remuneration. In
1.1. Institutional background
addition, the legal text contained that the health care providers
organized in physician networks are nancially responsible for
The Swiss health care system is nanced by a dual system com-
the medical provision of the network-insured individuals. In other
posed of a compulsory basic health insurance and an additional
words, the implementation of a budgetary co-responsibility for
voluntary insurance (e.g. free hospital and physician choice, private
ambulatory care physicians was intended. The referendum was
rooms in hospitals). The mandatory coverage includes most health
rejected by a strong majority of voters (76%). The main reason
care services and is written by about 70 competing private health
for rejecting the referendum was not the implementation of bud-
insurers. The health insurers are not allowed to make prots for the
getary co-responsibility. The physician community successfully
compulsory coverage but for the voluntary coverage. The premium
campaigned against the referendum with the argument that the
for the basic insurance depends on gender and area of residence,
reform would abolish free physician choice because the reform also
but not on health risk and income. Premium subsidies are granted
intended to impose a higher co-payment for patients that were not
to low-income citizens and are funded through general taxes. To
treated in a physician network (NZZ, 2012).
set incentives for individuals not to consume unnecessary medical
Even if the health care reform was rejected, budgetary co-
services, two cost-sharing incentives are installed. First, individu-
responsibility remains of central interest for both, private health
als have to choose one of six deductible levels that affect the nal
insurers as well as the government. Therefore, a better under-
premium level. Second, a co-payment of 10% limited to CHF 700
standing of physicians preferences for alternative reimbursement
(USD 840 in 2011) per year is imposed for annual costs exceeding
systems and their willingness to accept budgetary co-responsibility
the deductible. Insurers must accept all applicants for the manda-
is of great interest. Hence, the objective of this article is to measure
tory coverage but are allowed to use medical underwriting to reject
physicians preferences to work in networks and accept budgetary
applicants for supplementary coverage. To mitigate risk selection
co-responsibility. Willingness-to-accept (WTA) values are derived
for healthy risks in the basic insurance, a national risk adjustment
using a discrete choice experiment including 872 ambulatory care
scheme is in place, compensating insurers with a riskier portfo-
physicians surveyed in 2011.
lio than the Swiss population. Recently, a shift from conventional
The article is organized as follows. Section 2 gives a short lit-
insurance plans to MC policies took place. The main reason for
erature review of physician preference studies and the design
a shift in the demand toward MC policies are the lower premi-
of payment incentives for physicians. Section 3 discusses how
ums granted in return for the acceptance of certain restrictions,
budgetary co-responsibility can be designed and describes how
e.g. accepting a gatekeeper and giving up free physician choice.
physicians risk aversion can be expressed in WTA values. Section
As a consequence, health insurers have to contract with additional
4 presents the modeling approach, derives expected preference
physicians, physician networks, and Health Maintenance Organiza-
tendencies for or against reimbursement attributes used in the
tions (HMOs) to enroll their portfolio. Additional information about
experiment, and explains the design of the discrete choice exper-
the Swiss health insurance system can be found in Trottmann et al.
iment. In addition, physicians utility derived from alternative
(2012).
reimbursement designs and the applied econometric model used
In Switzerland, ambulatory care is predominantly provided by
to elicit the preference weights is discussed. Section 5 describes
independent private practice physicians. These are mostly paid
the survey data. The estimation results are interpreted in Section
through fee for service. Only a small number of ambulatory care
6. The concluding Section 7 compares the estimated willingness-
physicians works in MC-type arrangements, where alternative
to-accept values with potential cost saving through global budgets
remuneration models like capitation are used, but according to
and outlines the implications.
Berchtold and Peytremann-Bridevaux (2010) every second gen-
eral practitioner and more than 400 specialists have cooperated
with an established medical group in 2010. Contracting with these 2. Literature review
organizations is getting more important today with the increas-
ing demand for MC plans, mainly because these are cheaper than A broad international literature on incentive pay exists, which
conventional plans. A common contract form is what Reich et al. mainly revolves around the impacts of differently designed incen-
(2012) call a contract model with capitation. In this set-up, a health tive schemes to improve the performance of the employees (see
insurer contracts with either a group of (independent) physicians Ichniowski and Shaw (2003) and Gibbons (1996) for an overview).
organized as a network or an HMO. These organizations agree to With respect to physician payment, the focus is on installing incen-
provide health care for the enrollees and to accept a global bud- tives regarding the effort of ambulatory care or hospital physicians,
get, which is calculated as a per patient per year payment. As Reich mainly by replacing conventional fee for service by global bud-
et al. (2012) emphasize, the global budget is in practice a virtual cost gets. As discussed by Robinson (2001), physician remuneration
 M. Rischatsch / Journal of Health Economics 40 (2015) 109121
through fee for service rewards inappropriate services, fraudu-
lent up-coding of visits and procedures, and reciprocal referrals
among specialists. On the other hand, capitation may lead to the
refusal of expensive medical services and stimulate activities to
avoid patients with chronic diseases. This makes clear that the
introduction of budgetary co-responsibility for ambulatory care
physicians to tackle rising health care expenditure entails certain
drawbacks that might affect the treatment quality. For this reason,
the Blue Cross Blue Shield (BCBS) of Massachusetts introduced a
modied form of the conventional global budget models in 2009.
The so-called Alternative Quality Contract (AQC) has three major
components. First, a global budget with an annual spending growth
limit for the next ve years replaces the annually bargained bud-
get for the subsequent year. Second, incentive payments to improve
treatment quality are introduced, which was absent in the former
global budget models. This offers medical groups the opportunity to
make signicant prots if they meet the clinical performance target
set by the BCBS. Third, technical support for participating providers
is granted. For a more detailed discussion on AQC see Chernew
et al. (2011) who nd that all medical groups were rewarded with
a signicant quality bonus in the rst year of the AQC year.
In the United Kingdom, the National Health Service (NHS) intro-
duced the fundholding scheme in 1991 allowing larger general
practices to work voluntarily under a global budget for hospital
referrals. These practices received a bonus if costs were below the
budget. As a consequence, they reduced unnecessary use of sec-
ondary care, increased their effort to purchase secondary care at
lower rates, and decreased waiting times for hospital treatment.
Nevertheless, the fundholding was abolished in 1999 because it
was perceived as unfair and led to high transaction costs. Croxson
et al. (2001) nd empirical evidence that fundholders increased
their use of secondary care in the year before entering the scheme
to inate their budgets. Dusheiko et al. (2006) nd that the aboli-
tion of fundholding increased ex-fundholders admission rates for
chargeable elective admissions up to 5%.
Gaynor and Gertler (1995) investigate the determinants of the
internal organization of medical group practices. Focusing on the
inuence of the physicians risk aversion on the choice of com-
pensation method and physician panel size, they show that moral
hazard is present and that moving from fee for service to capita-
tion results in reduced physician effort, especially in large practices.
They conclude that physicians risk aversion inuences the choice
between risk diversication through larger networks and potential
lower productivity of the practice. Gaynor et al. (2004) study how
primary care physicians in MC networks respond to nancial incen-
tives to contain medical expenditure in the United States, where
MC organizations rely on a system of nancial and non-nancial
incentives that encourage physicians to control costs. They nd
that group-based incentives are more effective in reducing medical
expenditure for small physician panels because many group-based
incentives rely to some extent on the power of peer pressure to
abolish moral hazard. Budgetary co-responsibility to be effective
therefore requires an optimal choice of physician panel size.
The behavior of Swiss physician was analyzed in several
studies including Rischatsch and Zweifel (2012) who investigate
private-practice physicians willingness to apply some typical MC
treatment concepts. They nd that cost savings from the applica-
tion of these concepts have to be substantial to be able to pay the
compensations asked by the physicians for voluntary application.
Rischatsch et al. (2013) measure the impact of nancial incentives
on the prescribing behavior of Swiss physicians. The study uses
claims data to explain the disparity in the rate of generic substi-
tution between dispensing and non-dispensing physicians. They
show that the market share for generic drugs increases if physi-
cians prot from selling the drugs they prescribe. In Rischatsch
111
(2014), dispensing physicians drug margin optimization is stud-
ied and empirical evidence is found that nancial incentives from
sub-optimal pharmaceutical price regulation in Switzerland lead
to a cost-inefcient prescription of drugs if physicians are selling
drugs on their own account.
In general, these studies show that physicians are prone to nan-
cial incentives and that setting the right incentives can motivate
them to work more cost-conscious. While most studies focus on
measuring the success in reducing health care expenditure, the
objective of this article is to analyze the required compensation
payment to physicians for accepting a budgetary co-responsibility.
The next section explains how incentives can be designed in physi-
cian payment mechanisms.
3. Physician payment mechanisms
This section starts with the description of a general reim-
bursement mechanism that permits the incorporation of nancial
incentives to reduce health care expenditure. Depending on
the choice of its parameters, the payment system reduces to
conventional fee for service, the dominant form of physician reim-
bursement in Switzerland.
Following Ellis and McGuire (1986), physician utility is assumed
to depend on the networks prot () as well as patient health ben-
et (B), i.e. (, B). For simplicity, patient benet is ignored in the
theoretical part, focusing only on network prot with / > 0.
However, the effect of patient health benet considerations on net-
work participation is included in the empirical part. Assuming a
network panel of N physicians, the network prot can be written as
 = (1 s)R + [sR + b(T C)] 1(C T ) [m(C T )] 1(C > T )
+ pE C (1)
where R indicates the network revenues, s the share of retained
revenues by the health insurer (discussed below), T the spending
target or global budget, and C the costs of the medical services pro-
vided. The expression 1( ) is an indicator function, being one if the
expression in parenthesis is true and zero otherwise. The share of
cost savings (T C) that the insurer pays to the network is given
by b, while the share of costs (C T) to be borne by the network is
given by m. An additional source of prot is a payment per enrollee
(p), which is multiplied by the total number of enrollees (E).2
For simplicity, only one medical service is considered, so that
the networks total costs are given by C = i min {ci (q), L}, where
i indexes patients and ci (q) denotes patient-specic cost within a
year depending on the quantity of care (q) provided. It is assumed
that physicians can only inuence ci by choosing q, that the annual
number of patients is given exogenously, and that they do not reject
patients. To cap the risk caused by patients with catastrophic costs
(e.g. HIV patients), a so-called stop-loss limit (L) per patient per
year is introduced, meaning that if ci > L only L contributes to C, not
the actual amount ci . The health insurer pays the remainder. Net-
work revenues are given by R = (1 + )C with  being the loading
on costs to reward the network for the medical services provided.
The spending target (T) is an budget for the subsequent year. It
depends on the number of enrollees in the network and the pre-
dicted annual cost per enrollee (c ). The target level is determined
2
In this study, we investigate physicians maximizing the networks prot. The
individual prot for each member is then given by the network-specic internal
income allocation. Commonly, network members receive their own fee for service
claims and participate in a gain/loss. How specic gains or losses are distributed
among the members is not further discussed in this article.
112 M. Rischatsch / Journal of Health Economics 40 (2015) 109121
by T = w(c E), where w is a weighting factor taking into account
network-specic characteristics (e.g., specialization and skill mix),
forecasts of changes in demand, and expected price changes. The
factor  represents the factor for the intended cost reduction com-
pared to the non MC-insured collective.3
Gaynor et al. (2004) investigate an HMO payment mechanism
where a certain share of the revenues is retained by the insurer
until the end of the year (payment at risk) and only paid to the
HMO if costs are below the spending target. Otherwise, the insurer
keeps the retained fees. This type of incentive is represented in Eq.
(1) by the parameter s. In Switzerland, a bonus/malus system is
more common in practice that can be dened in absolute or rel-
ative terms. Because different types of physicians with different
cost structures are investigated, relative bonus/malus payments
are more appropriate than absolute payments. Thus, the param-
eter b denotes the share of the achieved cost savings that goes to
the network as a gain-sharing bonus and m denes the share of
the cost above a spending target that the network has to bear as a
loss-sharing malus. Hence, in addition to the income contribution
through , a network is able to increase prots or decrease losses
by providing the optimal quantity of care and optimize costs.
3.1. State-dependent network prots
One can dene the state-dependent network prot to be 1
if costs are below the spending target (state 1) and 2 if costs
exceed the spending target (state 2) with C1 < T < C2 . In this case,
the expected prot is
E[] =  [C1 + b(T C1 )] + (1 ) [C2 s(1 + )C2
m(C2 T )] ,
where  denotes the probability that state 1 occurs. If no budget
responsibility is introduced with s, b, m, and p being zero, the net-
works are reimbursed by conventional fee for service and Eq. (2)
reduces to
E[F ] = [C1 ] + (1 )[C2 ].
One can single out the uncertainty caused by budgetary co-
responsibility, which is reected by
E[B ] =  [b(T C1 )] + (1 ) [s(1 + )C2 m(C2 T )] ,
with E[] = E[F ] + E[B ]. Assuming risk averse physicians, a von
NeumannMorgenstern (VNM) utility function () permits a
comparison between certain and uncertain prots and the mea-
surement of the compensation required to bear the higher risk
through budgetary co-responsibility. The VNM utility function can
be written as
(E[] P) =  (1 ) + (1 )(2 ),
where P is the risk premium or willingness to pay to avoid the
uncertainty (compare Zweifel and Eisen, 2012, Chapter 2). The
risk premium is modeled in the reimbursement mechanism by
the inclusion of a payment per enrollee (pE). This should com-
pensate the physicians for the uncertainty introduced through the
budgetary co-responsibility. A more detailed discussion of the com-
pensation that serves as the price attribute in the discrete choice
experiment is given in Section 4.4
3
For simplicity, only one group of enrollees is modeled. In reality, many different
groups are applied depending e.g. on age, gender, and health care expenditure in
the previous year.
4 6
Because the bonus/malus system allows networks to make prots and losses,
prospect theory might be an alternative approach to model physicians preferences.
3.2. Conventional fee for service
Fee for service constitutes the reference reimbursement system
in the discrete choice experiment because it is the dominant reim-
bursement mechanism for independent private practice physicians
in Switzerland and does not include budgetary co-responsibility
(see Section 4.2). The prot function given by Eq. (1) reduces to
 = C.5 Thus, physicians may be incentivized to provide a higher
quantity of care than optimal leading to an increase in ci , which
constitutes the classical supplier-induced demand problem as dis-
cussed in McGuire (2000).
3.3. Alternative reimbursement designs
Budgetary co-responsibility can be modeled in different ways.
A rst example is described and analyzed by Gaynor et al. (2004)
and can be modeled using Eq. (1) with positive values for s, b, and
, so that the network prot reads
 = (1 s)R + [sR + b(T C)] 1(C T ) C, (6)
with R = (1 + )F and  = 0.25. The spending target is derived as
discussed previously. The stop-loss provision is limited at USD
15,000 per patient per year and revenues are calculated based on
Medicares fee schedule denoted by F. Bonuses are based on the
performance of panels of doctors which vary between three and
30 physicians. Realized cost savings are divided in equal shares
between the insurer and the network (b = 0.5). In the case that the
actual costs exceed the spending target, the networks loss sharing
amounts to the retained revenues (s = 0.2).
A second example replaces the retained revenues by loss-
sharing malus. The gain-sharing bonus is sometimes designed to
(2)
be more generous than the loss-sharing malus because insurers
are better able to manage risks than are networks. Services are paid
through fee for service (income contribution) where revenues are
R = (1 + )C. In addition, a payment per enrollee is paid. With this
specication, Eq. (1) reduces to
 = C + b(T C)1(C T ) m(C T )1(C > T ) + pE. (7)
(3)
This second example constitutes the most common form of
budgetary co-responsibility in Switzerland today. Therefore, the
discrete choice experiment is modeled using Eq. (7).
(4)
4. Modeling approach
4.1. Derivation of expected preference tendencies
Referring to Eq. (7), a health insurer can transfer more or less
budgetary responsibility to the network by choosing different val-
ues for b, m, and T. In the following, all parameters pertaining to the
(5) reimbursement design will be called reimbursement parameters.
This section derives the expected signs for the preference weights
under the assumption that / > 0. Because payments at risk
and changes in mark-ups are not of interest in the Swiss context,
these attributes are neglected.
It is assumed that physicians are not completely certain which
state, state 1 or state 2, they will end up in, hence 0 <  < 1. The
rst derivative of the expected prot with respect to the gain-
sharing bonus parameter is given by E[]/b = (T C1 ) > 0. Thus,
physicians are predicted to prefer higher bonuses so that 1 > 0
is expected.6 Increasing the loss-sharing malus parameter m is
5
In this case the reimbursement parameters s, b, and p are set to zero.
We assume that C1 < T < C2 so that an indifference arising from C1 = T = C2 is
excluded.
 M. Rischatsch / Journal of Health Economics 40 (2015) 109121
Table 1
Summary of attributes, expected signs, attribute levels, and units
Attributes Abbr. E[sign]
Bonus BON 1 > 0
Malus MAL 2 < 0
Stop-loss limit LIM 3 < 0
Panel size PHY 4 0
Target reduction TAR 5 <0
Payment PAY 6 > 0
undesired by physicians because E[]/m = (1 )(C2 T) < 0.
A higher malus parameter means that physicians have to bear a
higher share of the costs above the target if they end up in state 2,
and therefore 2 < 0 is predicted. The stop-loss limit enters Eq. (7)
through C = i min {ci , L} and therefore C/L > 0 if ci L. There-
fore, physicians are predicted to dislike an increase in the stop-loss
limit, hence 3 < 0. According to Gaynor et al. (2004), the num-
ber of physicians in a panel working under the same target is
an important decision variable for physicians joining a network.
Implementing network-level incentives introduces the problem
of moral hazard and the intensity of incentives might depreci-
ate as the panel size increases. Hence, the number of physicians
working under the same target is included as a reimbursement
parameter. However, Eq. (7) does not allow deriving predictions
for the network size (4 ) but considering risk diversication and
moral hazard one may expect a preference for medium size pan-
els. A higher spending target is preferred by the physicians because
E[]/T = b + (1 )m > 0 and therefore 5 > 0 as long as at least
b or m is positive. The spending target is modeled in the choice
experiment as target reductions and therefore, the coefcient 5
is expected to be negative. Finally, increasing the payment per
insured per month p is predicted to have a positive effect on physi-
cian utility (6 ), because E[]/p = E > 0. The predicted preference
tendencies or expected signs of preference weights are summa-
rized in Table 1.
4.2. Design of the choice experiment
Physicians preferences for or against budgetary co-
responsibility are measured using a discrete choice experiment
(DCE). In the following, the choice of attributes and their levels is
explained.
4.2.1. Attributes and their levels
The gain-sharing bonus representing the share of potential cost
savings (target minus actual cost) and the loss-sharing malus mea-
suring the share of excess cost (actual cost minus target) to be
borne by the network is modeled using the attributes BON and MAL
with six levels ranging from zero to 50%. Higher values than 50%
are unlikely to be implemented in Switzerland and therefore not
considered. In the Swiss context, the stop-loss limit (L) for general
practitioners is commonly about CHF 10,000 per patient per year
(1 CHF 1.20 USD in 2011). For specialized physicians costs are
more likely to reach this limit than for general practitioners. There-
fore, the stop-loss limit (LIM) is modeled to have three levels: CHF
10,000, 20,000 and 30,000 per patient per year. For an international
comparison, the HMO analyzed by Gaynor et al. (2004) limits the
stop-loss provision at USD 15,000 ( CHF 13,630) for primary care
physicians. The attribute PHY represents the size of the physician
panel. On the one hand, its inclusion in the experiment is important
because the intensity of incentives might decrease with panel size
and imposing network-level incentives may introduce moral haz-
ard. On the other hand, larger networks are better able to manage
nancial risks. This was already discussed in the literature review.
Referring again to Gaynor et al. (2004), the physician panels there
113
Description Attribute levels
Share of cost savings paid to network 0, 10, 20, 30, 40, 50 in %
Share of excess cost borne by network 0, 10, 20, 30, 40, 50 in %
Stop-loss limit per patient and year 10, 20, 30 in CHF 1000
Number of physicians with same target 10, 50, 100 physicians
Reduction compared to non-MC cost 0, 5, 10, 15, 20, 25 in %
Risk payment per insured per month 0, 0.5, 1, 1.5, 2, 2.5 in CHF
range from three to 30 physicians. The experiment uses three levels
to model network size: 10, 50 and 100 physicians. In Switzerland,
only a few of the largest physician networks have a panel size
of approximately 50 physicians. To model realistic scenarios in
the Swiss context and to measure the preferences for larger net-
works, an upper level of one hundred physicians was included in
the experiment. The levels where discussions with physicians and
health insurance representatives to ensure realistic scenarios. The
spending target (TAR) is included as percentage reductions of the
reference budget determined upon the cost of comparable non-MC
insured citizens. The reduction in the spending target ranges from
zero to 25% using intervals of 5%. The highest level is believed to
represent an upper bound for cost savings through MC based on
the ndings by Reich et al. (2012). Finally, monetary compensation
for the additional nancial risk is included as risk payments (PAY)
per insured per month ranging from zero to CHF 2.50 per insured
per month (see Table 1).
4.2.2. Experimental design
The total of six attributes and their levels combine to form
11,664 (=64 32 ) possible combinations of alternative reimburse-
ment designs. SASs statistical software package JMP was used to
generate a D-efcient design. The coefcients were assumed to
be zero and the design was specied to have 80 alternatives. The
resulting alternatives were randomly split into four blocks, each
with 20 alternatives. No xed comparator alternative was used and
the 20 alternatives randomly paired to constitute ten binary choices
between the hypothetical network alternatives A and B (compare
choice y1 in the sequential choice set-up discussed next). To reduce
the cognitive burden for respondents, only up to three attributes
were allowed to differ at the same time.7
4.2.3. Sequential choice set-up
In each of the 10 choice scenarios of each block, respon-
dents had to make two sequential decisions. First, two of the
generated alternatives were presented. The respondent was
asked to make a rst binary choice (y1 ) between Network A
and Network B differing only in the reimbursement attributes.
This choice constitutes a forced choice because no opt-out
alternative is given. Second, the respondents were asked to make a
conditional choice (y2 |y1 ) between the previously chosen network
(Network A or B) and the opt-out alternative representing to work
independently. The second choice of each scenario is an unforced
choice.
The sequential choice set-up with a forced followed by an
unforced choice was designed to mitigate strategic behavior by
choosing always the opt-out alternative. Under y1 , respondents
have always to trade off liked and disliked attributes so that no
alternative is available for strict protesting. Without this sequen-
tial set-up, the discrete choice experiment might have resulted in
7
A rst design was generate using NGENE. Unfortunately, the software was not
able to handle the mentioned restriction, which is the reason why JMP was used
instead.
114 M. Rischatsch / Journal of Health Economics 40 (2015) 109121

Table 2 whether network participation differs. To control for geographical

Example of the sequential choice scenario
differences, practice location is considered discriminating between
Network A Network B urban, suburban (SUBURB), and rural (RURAL) areas. In the past,
Share of cost savings paid to 50% 30% numerous studies (and referendums) have shown that preferences
network as a bonus differ between the language areas. Hence, an indicator variable for
Share of excess cost borne by 20% 20% German-speaking physicians (GERM) is used, where French- and
network as a malus Italian-speaking physicians constitute the reference group. Finally,
Stop-loss provision limit per CHF 10,000 CHF 10,000
the stated self-assessment of the probability of producing a cost
patient and year
Number of physicians with 10 physicians 50 physicians below the comparable practices (P(C < T), compare  in Eq. (2)) and a
same target covariate indicating if respondents think that the cost discussion in
Target reduction compared to 10% 10% Switzerland leads to a decrease in the treatment quality (HEALTH)
non-MC cost
is included. The latter variable is coded as minus one if the respon-
Risk payment per insured per CHF 1.50 CHF 1.00
month dent believes the statement is wrong, zero if indifferent, and one
if it is believed that the statement is true. The class-membership
Choice 1 (y1 ): I prefer to join... Network A  Network B 
equation used to explain segmentation of physicians into classes is
Choice 2 (y2 ): I prefer to... join preferred
remain independent  network above 
 z = 1 SP1 + 2 SP2 + 3 PSY + 4 MALE + 5 AGE + 6 MARR

only collecting information about network opposition measured + 7 SUBURB + 8 RURAL + 9 GERM + 10 P(C < T )
by the model constant regardless of the reimbursement attributes.
+ 11 HEALTH + 12 CONST (9)
The second choice is included to measure overall resistance. Table 2
shows an example of a choice scenario with its two sequential and explained in the next section.
choices y1 and y2 .
4.4. Applied nite mixture model
4.3. Specication of physician utility
The preference parameters of interest are estimated using a
Preferences are measured based on the theory of stochastic util- latent-class logit model. In addition, a random-coefcient logit
ity maximization (Luce, 1959; Manski, 1977; McFadden, 1981). The model is estimated as a reference model, but not discussed in
utility of alternative payment mechanisms is dened as a function detail here. The latent-class logit model assumes that the observed
of alternative-specic attribute levels. While a physicians util- choices come from two or more unobserved classes of physicians
ity from network participation is specied as being linear in the and where the proportions of class size are unknown a priori as
payment attributes, the utility from independent private practice discussed in Wedel and DeSarbo (2002). In contrast to the random-
(opt-out option) is given by an alternative-specic constant (com- coefcient model, the latent-class model incorporates preference
pare Marshall et al. (2009)). More precisely, a constant indicating heterogeneity among physicians by estimating class-specic coef-
the network option chosen under the rst choice is included in cients. The latent-class logit model can be extended to include
the second choice (NETW) to measure physician resistance against class-membership (also called concomitant) variables to explain
working in networks. An additional constant was included for the class membership by physician-specic characteristics as done by
rst binary choice, controlling for unmeasured preference in favor Kamakura et al. (1994), Leisch (2004), and Gruen and Leisch (2008).
of network B versus A (CONSTB) even though the pertaining coef- The choice probabilities are given by
cient is expected to be insignicant due to the experimental design.  T 
The basic model specication of the deterministic part of the ran- 
K  exp(k xint ) exp(k zn )
dom utility is Pin = J K , (10)
k=1 t=1 j=1
exp(k xjnt ) k=1
exp(k zn )

 x = 1 BON + 2 MAL + 3 LIM + 4 PHY + 5 TAR + 6 PAY
+ 7 CONSTB + 8 NETW.
Different functional forms are tested and discussed in Section
6.1. Depending on the econometric model, Eq. (8) is modied. The
stop-loss limits and the sizes of the physician panels are included
as categorical attributes. To be able to interpret the network con-
stant independently from the preferences for the reference levels
of these attributes, they are coded using effects coding. For a dis-
cussion on effects coding see Louviere et al. (2000) and Bech and
Gyrd-Hansen (2005).
Preference heterogeneity is modeled using a nite number of
physician classes with class-specic preference parameters. To
explain class membership of physicians, several socio-economic
variables are used. Physicians specialization is the rst segmenta-
tion variable and used to distinguish between general practitioners
who constitute the reference category and are compared to
specialists without (SP1) and with (SP2) surgical activities, as
well as psychiatrists (PSY). In addition, physician gender (MALE),
age (AGE), and marital status (MARR) are used to investigate
where the rst term is the logit probability of physician n choosing
alternative i at time t given to belong to class k (Pin|k ). The second
(8) term of Eq. (10) constitutes the class-membership function (Pnk ),
which explains the composition of classes using physician-specic
covariates and is given by Eq. (9). The model was estimated using
the flexmix package in R written by Gruen and Leisch (2008) and
parameter estimation is performed using the EM algorithm.
5. Data
The discrete choice experiment was included in online survey
and distributed using an e-mail newsletter. The anonymous survey
was addressed to approximately 11,000 Swiss physicians working
in ambulatory care. A pretest was conducted to check the respon-
dents understanding of the description of the attributes and their
levels and the questionnaire in general. Some comments disclosed
the sensitivity of the subject. Only some minor changes were nec-
essary (e.g. wording) before the main survey was nally elded
in August 2011. The survey was elded on a Wednesday evening
because most physicians have their administration day on Thurs-
day and the timing was expected to increase the response rate. The
 M. Rischatsch / Journal of Health Economics 40 (2015) 109121
survey was not personalized because of the sensitivity of the topic
and no reminders were sent out. The survey mode was chosen with
the knowledge of some limitations of web surveys, but avoiding to
mail a paper survey to thousands of physicians decreased response
time and cost and a higher response rate was expected. The rate of
return of 14% was slightly higher than expected based on previous
physician surveys. The sample size of physicians to be addressed
was in fact calculated using an expected response rate of 10%. This
rate was anticipated based on the previous experience with physi-
cian surveys in Switzerland made by the author as well as the Swiss
Medical Association.
All forced and unforced choices combined, a total of 15,748
choices from 872 physicians are observed and used in the esti-
mation of the latent-class model, some participating only in the
questions prior to the experiment. The share of choices in favor
of Network B versus A (forced choices) is 49.6%. This reects the
fact that no xed comparator alternative is used and alternatives
are randomly paired. The share of unforced choices in favor of the
previously chosen network alternative versus remaining indepen-
dent is 19.2%. This already reveals physicians strong preference
to remain independent. The share of physicians that made at least
once a choice in favor of a network and once in favor of the inde-
pendent practice option is 37.4% compared to 3.5% that always
preferred the network alternative and 59.1% that always preferred
the independent practice option. This shows a strong resistance
against physician networks in Switzerland. The distribution of
respondents between blocks not classes in the nal sample
is more or less uniform with 212 physicians (3891 choices), 235
(4246), 225 (3959), and 200 (3652), respectively.
General practitioners (including gynecologists and pediatri-
cians) account for 57% of the respondents, specialists without
surgical activities for 13%, specialists with surgical activities for
13%, and 17% are psychiatrists. The majority of physicians work
in independent single practice (49%). Ofcial statistics by Kraft
(2010) report a share of 63% working in single practice, indicating
an undersampling of single practice physicians. About one-third
work in shared practice on their own account (30%), while only 5%
work in shared practice on common account. Network participa-
tion in Switzerland is low and predominately for networks where
physicians continue to work on their own account (14%). Common-
account networks are the exception (2%). The average team size in
shared practice is 2.8 physicians. Males are oversampled, with a
share of 81% compared to 68% from the ofcial source cited above.
The average age in the sample is 54 years and equals the average age
reported by the Swiss Medical Association for all members working
in practices in 2010 listed in Kraft (2010). The respondents have 26
years of job experience on average. The majority (81%) is married.
Respondents are located in urban (47%), suburban (26%), and rural
(27%) areas and in the German-speaking (79%), French-speaking
(19%), and Italian-speaking (2%) part of Switzerland.
The parameter  in Eq. (2) denotes the probability of producing
an annual cost below the spending target. In the survey, respon-
dents were asked to assess their own probability of producing a
total cost (including pharmaceuticals and referrals to specialists
and hospitals) lower than comparable practices. Response options
ranged from zero to 100% in steps of 10%. On average, physicians
stated their probabilities to be at 52% (the median category is 50%).
Of those surveyed, 10% stated that they were certain to be over the
target, while 8% were sure to be below. These probabilities were
used to investigate whether physicians who assess their probabil-
ities to be high are more likely to join networks with budgetary
co-responsibility.
Physicians were asked if they believe that decreasing health
care expenditure is only possible with decreasing treatment qual-
ity. As to this, 79% think that decreasing expenditure is impossible
115
without degrading treatment quality, whereas only 11% think that
reducing costs is possible without affecting quality. The remainder
is indifferent. About 58% think that the cost discussion is a threat
to the treatment quality, while 25% think that this statement is
not true. Physicians see the highest potential for reducing health
care expenditure via the prescription of generic drugs. Neverthe-
less, they do not like to be forced to prescribe the generic drug
as found in Rischatsch and Zweifel (2012). A majority of 77% stated
that they consider cost when choosing treatments for their patients
while 9% do not consider cost when deciding about medical services
to provide.
In addition, respondents were asked if the they are currently
paid to some extent through capitation, a bonus or a malus. About
7% of the sampled physicians are paid to some extent through cap-
itation. Only a small share of Swiss ambulatory care physicians
have experience with alternative payment mechanisms. Informa-
tion campaigns addressed to physicians could lower the expected
resistance against cost sharing.
6. Results
The optimal number of latent classes is not known a priori. The
literature suggests to estimate latent-class models with different
numbers of classes and to keep the model with the lowest Bayesian
information criterion. Based on the Bayesian information criterion
shown in Table 3, the model with four classes ts the data bet-
ter than the same model with three and ve classes, respectively.
Modeling different functional relationships for the bonus and the
malus, the specication with continuous quadratic relations leads
to the best goodness of t compared to a linear continuous or cate-
gorical specication. Therefore, this model specication is retained
and discussed in the following. The estimated coefcients for the
model with three classes are shown in Table 5.
The choices are allocated to the class with the highest posterior
probability, while class membership is xed for each physician.
The estimation leads to segment sizes of 17%, 21%, 40%, and 21%
of observed choices, respectively. To circumvent being trapped in
a local maximum instead of reaching the global maximum of the
likelihood, the estimation process was repeated ve times using
different starting values.
6.1. Estimated preference weights
Table 4 shows the estimated class-specic coefcients together
with their standard errors (se) and p-values (pv). The upper
part presents the preference parameters for the reimbursement
attributes (), while the lower part lists the parameters pertaining
to the class-membership variables (). All statistically signi-
cant preference weights have the expected signs as predicted in
Section 4.1. This section starts with a discussion about the major
forces behind the physician segmentation. Subsequently, the dis-
cussion turns to the estimated preferences for the reimbursement
attributes.
6.1.1. Physician segmentation
Overall network aversion measured by the network constant is
revealed to be a major driver for physician segmentation. Using the
results of an initial estimation, the classes were ordered according
to the network aversion and the same model was re-estimated.
This facilitates the interpretation of the class-membership vari-
ables because the reference class is now the one that is most
likely to join a network if reimbursement attributes are neglected.
While the rst class has no statistically signicant preference for or
against networks in general and the second class weakly dislikes
116 M. Rischatsch / Journal of Health Economics 40 (2015) 109121

Table 3
Information criteria for model selection

Number of latent classes (k) Functional form of bonus/malus (k = 4)

a
k Bayesian Akaike ICL Relationship Bayesian Akaike ICLa

3 14,410.06 13,950.19 14,501.77 Cat. bonus/malus 14,503.53 13,675.77 14,772.27

4 14,399.97 13,756.16 14,682.50 Cont. bonus/malus (linear) 14,476.36 13,893.86 14,753.25
5 14,417.19 13,589.43 14,780.04 Cont. bonus/malus (quadratic) 14,399.97 13,756.16 14,682.50
a
ICL, integrated complete likelihood.

network participation, the third and the fourth class strongly dis-
like networks. Table 5 shows that if only three latent classes are
modeled, physicians from the third and fourth class are grouped
together, keeping the size of the rst two classes unchanged. Look-
ing at the reimbursement attributes elicits that the fourth class is
likely to be composed of protesters that are not considering the
reimbursement attributes when choosing between network alter-
natives. Drawing conclusions for this class should be limited to
the size of the class measuring the share of physicians protesting
against network participation. Therefore, the parameters for this
class are not discussed.
The class composition discussion is always in comparison to the
reference class representing physicians with the lowest network
aversion measured by the network constant. A strong impact on
physician segmentation is found to be the physicians specialty.
Predicting the class membership for each individual physician
permits to investigate class composition. The model estimation
results in classes of the sizes of 16.5% physicians (class 1), 20% (class
2), 42.5% (class 3), and 21% (class 4). Looking at the distribution
of physicians among classes based on their specialty shows that
37.3% of general practitioners are selected into the third class, while
24.8%, 21.6% and 16.3% are assigned to the second, rst, and fourth
class, respectively. 39.5% of specialists without surgical activity are
predicted to be in the third, 26.3% in the second, 22.8% in the fourth,
and 11.4% in the rst class, while specialists with surgical activ-
ity are split into the third (46.0%), fourth (37.2%), rst (10.6%) and
second (6.2%) class. Most psychiatrists are estimated to be in the
third class (59.7%) followed by the fourth class (22.8%). The remain-
ing two classes contain only 9.4% (class 2) and 8.1% (class 1) of all
sampled psychiatrists, respectively. Comparing the composition of
each class across specialties shows that the rst class is composed
of 41.8% general practitioners, 22.1% specialists without and 20.6%
with surgical activity, and 15.6% psychiatrists. In the same order of
speciality, predicted shares are 37.2%, 39.5%, 9.3%, and 14.1% for the
second class, 20.4%, 21.6%, 25.2%, and 32.7% for the third class, and
16.5%, 23.0%, 37.5%, and 23.0% for the fourth class.
The estimated class-membership coefcients shown in Table 4
support these ndings. General practitioners have the highest like-
lihood to be in the reference class. Physicians without surgical
activities are less likely to be in this class and have about the
same probability to be in the class with a weak or a strong aver-
sion. In contrast, physicians with surgical activities are more likely

Table 4
Estimation results of latent-class logit model with four classes

Latent-class logit

Class 1 Class 2 Class 3 Class 4

Attributes se pv se pv se pv se pv

BON 0.48 0.09 0.00 0.80 0.10 0.00 0.92 0.11 0.00 0.21 0.16 0.18
BON.SQ 0.06 0.02 0.00 0.11 0.02 0.00 0.14 0.02 0.00 0.04 0.03 0.18
MAL 0.06 0.09 0.51 0.72 0.10 0.00 0.50 0.10 0.00 0.37 0.14 0.01
MAL.SQ 0.04 0.02 0.03 0.07 0.02 0.00 0.03 0.02 0.17 0.01 0.02 0.55
LIM20.EC 0.10 0.07 0.18 0.17 0.07 0.02 0.25 0.08 0.00 0.27 0.11 0.01
LIM30.EC 0.26 0.07 0.00 0.41 0.08 0.00 0.76 0.09 0.00 0.27 0.08 0.00
PHY50.EC 0.31 0.06 0.00 0.28 0.06 0.00 0.43 0.07 0.00 0.01 0.08 0.86
PHY100.EC 0.14 0.06 0.03 0.01 0.07 0.92 0.42 0.08 0.00 0.79 0.13 0.00
TAR 0.03 0.07 0.63 0.72 0.07 0.00 0.42 0.06 0.00 0.06 0.09 0.46
PAY 0.42 0.05 0.00 0.43 0.05 0.00 0.17 0.06 0.00 0.01 0.07 0.85
CONSTB 0.20 0.07 0.00 0.12 0.07 0.07 0.31 0.07 0.00 0.27 0.09 0.00
NETW 0.04 0.19 0.81 0.60 0.23 0.01 5.20 0.32 0.00 5.82 0.74 0.00

Class membership Reference class se pv se pv se pv

SP1 0.78 0.42 0.06 0.71 0.39 0.07 0.87 0.43 0.04
SP2 0.68 0.58 0.24 0.85 0.40 0.03 1.38 0.42 0.00
SY 0.15 0.51 0.76 1.59 0.39 0.00 1.40 0.43 0.00
MALE 0.44 0.39 0.26 0.33 0.36 0.36 0.62 0.39 0.11
AGE 0.24 0.09 0.01 0.08 0.07 0.30 0.10 0.09 0.25
MARR 0.85 0.38 0.02 0.61 0.35 0.08 0.68 0.38 0.07
SUBURB 1.05 0.34 0.00 0.68 0.31 0.03 0.76 0.34 0.02
RURAL 0.43 0.33 0.19 0.34 0.28 0.24 0.16 0.33 0.64
GERM 0.56 0.37 0.13 0.20 0.29 0.49 0.53 0.31 0.09
P(C < T) 0.01 0.04 0.79 0.03 0.04 0.44 0.01 0.04 0.90
HEALTH 0.49 0.15 0.00 0.89 0.14 0.00 0.77 0.16 0.00
CONST Pnk 2.91 1.01 0.00 1.98 0.90 0.03 0.07 1.05 0.94

Class sizes: choices 2697 (17.1%) 3370 (21.4%) 6369 (40.4%) 3312 (21.0%)
Class sizes: physicians 144 (16.5%) 174 (20.0%) 371 (42.5%) 183 (21.0%)

Note: , are the class-specic coefcients, se are the standard errors, and pv are the p-values. The sufxes SQ and EC indicate the square-term and the effects-coded
variables. BON, BON.SQ, MAL, MAL.SQ, TAR, and P(C < T) are in 10%-points, PAY in CHF per insured per month, and AGE per 5 years.
 M. Rischatsch / Journal of Health Economics 40 (2015) 109121 117

Table 5
Estimation results of latent-class logit model with three classes

Latent-class logit

Class 1 Class 2 Class 3

Attributes se pv se pv se pv

BON 0.49 0.10 0.00 0.82 0.10 0.00 0.66 0.07 0.00
BON.SQ 0.06 0.02 0.00 0.11 0.02 0.00 0.10 0.01 0.00
MAL 0.07 0.09 0.44 0.71 0.10 0.00 0.38 0.07 0.00
MAL.SQ 0.03 0.02 0.04 0.07 0.02 0.00 0.02 0.01 0.20
LIM20.EC 0.09 0.07 0.18 0.16 0.07 0.02 0.17 0.05 0.00
LIM30.EC 0.26 0.07 0.00 0.42 0.08 0.00 0.47 0.04 0.00
PHY50.EC 0.31 0.06 0.00 0.30 0.06 0.00 0.21 0.04 0.00
PHY100.EC 0.13 0.06 0.04 0.03 0.06 0.61 0.05 0.04 0.21
TAR 0.02 0.07 0.80 0.71 0.06 0.00 0.27 0.04 0.00
PAY 0.42 0.05 0.00 0.43 0.05 0.00 0.14 0.04 0.00
CONSTB 0.19 0.07 0.00 0.14 0.07 0.04 0.06 0.04 0.12
NETW 0.01 0.19 0.95 0.73 0.23 0.00 5.06 0.28 0.00

Class membership Reference class se pv se pv

SP1 0.77 0.42 0.06 0.76 0.36 0.04

SP2 0.66 0.58 0.26 1.08 0.37 0.00
PSY 0.11 0.50 0.83 1.51 0.36 0.00
MALE 0.45 0.39 0.25 0.42 0.33 0.21
AGE 0.23 0.09 0.01 0.01 0.07 0.92
MARR 0.85 0.37 0.02 0.61 0.32 0.06
SUBURB 1.03 0.34 0.00 0.71 0.28 0.01
RURAL 0.45 0.33 0.17 0.27 0.26 0.30
GERM 0.46 0.36 0.20 0.35 0.26 0.17
P(C < T) 0.01 0.04 0.75 0.02 0.03 0.54
HEALTH 0.51 0.15 0.00 0.85 0.12 0.00
CONST Pnk 2.91 1.00 0.00 1.81 0.83 0.03

Class sizes: choices 2728 (17.3%) 3496 (22.2%) 9524 (60.5%)

Note: , are the class-specic coefcients, se are the standard errors, and pv are the p-values. The prexes SQ and EC indicate the square-term and the effects-coded
variables. BON, BON.SQ, MAL, MAL.SQ, TAR, and P(C < T) are in 10%-points, PAY in CHF per insured per month, and AGE per 5 years.

to be in the third class compared to the rst and second class
and most likely in the fourth class. A psychiatrist is most likely
segmented into the third class. While gender does not affect the
segmentation, older physicians are less likely to be in the second
compared to the rst class, but equally likely to be in one of the
other classes. Married physicians are most likely in the reference
class. A possible explanation is that networks allow physicians to
work part time and to avoid days on emergency call, which might
be important for physicians with family. While physicians located
in urban and rural areas disclose the same membership probabil-
ity for all four classes, physicians located in suburbs are less likely
to be in the reference class. German-speaking respondents are sig-
nicantly less likely to be in the fourth class compared to their
French- and Italian-speaking colleagues. Otherwise, language has
no explanatory power with respect to the segmentation. While
the self-assessed probability of producing cost below comparable
colleagues has no statistically signicant effect on segmentation,
physicians believe that the cost discussion negatively affects the
treatment quality is positively correlated with the likelihood of
being in a class with a stronger aversion.
In addition to the latent-class model, preference heterogene-
ity among physicians is also modeled using a random-coefcient
model. For a model description see Train (2003) and for a com-
parison between the latent class and the continuous mixture of
logit models see Hess et al. (2011). The estimated parameters
of the mixing distributions are shown in Table 6. The estimates
support the ndings of the latent-class model that general prac-
titioners and specialists without surgical activity have the lowest
network aversion measured by the constant followed by those with
surgical activity, and psychiatrists. While gender has no explana-
tory power with respect to class membership in the latent-class
model, the random-coefcient model estimates that males dislike
networks more than females. Older and married physicians are
more likely to join networks, while suburban and rural physicians
dislike networks more than urban physicians. German-speaking
physicians are not statistically different from their French- and
Italian-speaking colleagues. A higher self-assessed probability of
being able to produce cost below comparable practices decreases
the resistance to join a network. Finally, physicians who believe
that the cost discussion leads to a lower treatment quality also
dislike working in a network.
6.1.2. Preferences for reimbursement designs
Focussing on the rst three classes, all main- and square-term
effects for the bonus turn out to be statistically signicant with a
positive but decreasing impact on the choice probability. For the
malus, all main-effects coefcients take on the expected sign. For
the rst class, the main effect is insignicant while the quadratic
effect is signicant. The opposite is true for the third class. To inter-
pret the effects-coded attributes the estimates are transformed.8
Increasing the stop-loss limit from CHF 10,000 to 20,000 has no
statistical signicant effect on accepting an alternative reimburse-
ment scheme for the rst two classes. An explanation for this
nding is that even if a higher stop-loss limit increases the net-
works revenue at risk, it also permits to achieve higher cost savings
compared to non-MC insured patients. The group of physicians
with a strong dislike for network participation is less likely to
join a network if the stop-loss limit is increased to CHF 20,000.
8
To interpret
the effects-coded attributes (), the coefcients have to be trans-
formed as i = j + i , where i is the attribute level of interest and j is an index
j
for all attribute levels included. See Louviere et al. (2000) for a more detailed dis-
cussion.
118 M. Rischatsch / Journal of Health Economics 40 (2015) 109121

Table 6
Estimation results of random-coefcient logit model

Random-coefcient logit

Mixing distribution Mean Std.Dev. WTA

Attributes  se pv se pv Mean Median

BON Censored normal 0.15 0.02 0.00 0.30 0.03 0.00 0.67 0.49
MAL Censored normal 0.36 0.02 0.00 0.05 0.02 0.04 1.15 1.15
LIM20.ECa Normal 0.24 0.04 0.00 0.03 0.05 0.50 0.34 0.34
LIM30.ECa Normal 0.60 0.04 0.00 0.51 0.04 0.00 2.78 2.78
PHY50.ECa Normal 0.41 0.03 0.00 0.20 0.04 0.00 2.38 2.38
PHY100.ECa Normal 0.01 0.03 0.81 0.66 0.04 0.00 1.13 1.13
TAR Censored normal 0.38 0.13 0.00 1.64 0.16 0.00 0.00 1.53
CONSTB Normal 0.12 0.03 0.00 0.04 0.05 0.49 0.40 0.40
PAY xed 0.31 0.03 0.00
NETW Normal 4.23 0.36 0.00 3.71 0.13 0.00 13.48 13.48
SP1 Fixed 0.07 0.14 0.62 0.22
SP2 Fixed 2.12 0.19 0.00 6.74
PSY Fixed 2.71 0.20 0.00 8.62
MALE Fixed 0.29 0.13 0.03 0.91
AGE Fixed 0.06 0.03 0.03 0.20
MARR Fixed 1.00 0.13 0.00 3.18
SUBURB Fixed 0.88 0.12 0.00 2.82
RURAL Fixed 0.38 0.11 0.00 1.22
GERM Fixed 0.68 0.13 0.00 2.17
P(C < T) Fixed 0.07 0.02 0.00 0.21
HEALTH Fixed 1.19 0.06 0.00 3.80

BON, MAL, TAR and P(C < T) per 10%-points, PAY in CHF per insured per month, and AGE per 5 years. Displayed  coefcients for MAL and TAR are for the negative values.
a
WTA calculation based on the binary-coded coefcients.

An increase from CHF 10,000 to 30,000 has a negative impact for
all classes, the strongest for those physicians that dislike networks
most. Physicians prefer medium and large panel sizes but do not
distinguish between a panel size of 50 and 100. A reduction of
the spending target is perceived negatively by the second and the
third class but has no effect on the choice probability of the refer-
ence class. The reference class seems to believe in their ability to
achieve cost reductions, to meet the spending targets, and receive
a gain-sharing bonus. Finally, the price attribute in the experiment
is statistically signicant for all classes with the exception of the
protesting class four. Therefore, no willingness to accept is calcu-
lated for the fourth class in the next section. The rst two classes
place the same positive value on the payment per insured per
month, while the same payment to the third class has a signicantly
lower impact on network participation.
6.2. Willingness-to-accept values
The previously discussed preference weights can be used to pre-
dict how much it would cost an insurer (a network or an HMO) to
win physicians over to work under certain reimbursement designs
and take budgetary co-responsibility. To answer this question,
class-specic willingness-to-accept (WTA) values are derived and
discussed. The stop-loss limit and the size of the physician panel
enter the utility function as categorical variables using effects-
coded variables, while the reduction of the spending target is
modeled as a linear continuous variable. In these cases, the WTA
values are given by the ratio of the attribute-specic coefcient
and the one pertaining to the payment attribute. The estimates of
the effects-coded variables are rst transformed to dummy-coded
variables (see Bech and Gyrd-Hansen (2005) again). In contrast, the
bonus and the malus are modeled as quadratic continuous vari-
ables. Therefore, the derivation of the WTA values is different. The
willingness to accept is given by (1 xm + 2 (xm )2 )/3 , where 1 and
2 are the estimated coefcients pertaining to the main and the
square term of attribute m, respectively, and 3 is the preference
weight pertaining to the payment attribute. Figs. 1 and 2 illustrate
the derived WTA values that are discussed next. The standard errors
used for the calculation of condence intervals (see Table 7) are
estimated using the Delta method. All WTA values are measured
in CHF per insured per month with CHF 1 USD 1.20 in 2011.
Again, the payment coefcient for the fourth class is statistically
insignicant and therefore, compensation is not discussed for this
group.
6.2.1. Bonus and malus
While physicians from the rst class do not ask for a statisti-
cally signicant compensation to join a network without budgetary
responsibility, the physicians from the other classes do. The total
compensation required depends on the sum of an initial payment
for giving up independent practice and working in a network and
the compensations required for changes in the reimbursement
attributes. The initial payments derived using the network con-
stants amount to CHF 1.38 and CHF 31.11 per insured per month
for the second and the third class, respectively, while physicians in
the third class are likely physicians with surgical activities.
The positively valued bonus reduces the total compensation
while the introduction of a malus further increases the amount
required to compensate physicians for a voluntary participation.
Starting the discussion with the willingness to accept a malus, the
values amount to CHF 0.232.91 for the rst class, CHF 1.514.47
for the second class, and CHF 2.8511.20 for the third class. The
compensation asked by the rst class is relatively moderate in
comparison to the one asked by the third class (see Fig. 1 and
Table 7). Depending on the bonus level, compensation for accept-
ing an alternative remuneration reduces between CHF 1.022.40
(class 1), CHF 1.593.31 (class 2), and CHF 4.688.92 (class 3). The-
oretically, the lower valuation of a bonus of 50% compared to 40% is
counterintuitive and more a result of the functional form estimated.
Comparing symmetric bonus-malus combinations reveals that the
gain-sharing bonus overcompensates the loss-sharing malus for
combinations up to 40% in the case of the rst and up to 30% in
case of the third class. The required compensation is positive for
higher levels. The opposite is true for the second class where only a
combination of 10% is valued positively. For a comparison, the WTA
values from the random-coefcient model are shown in Table 6.
 M. Rischatsch / Journal of Health Economics 40 (2015) 109121 119

15 Bonus Malus

11.2
10 9.57

Willingness to accept (in CHF)

7.64

5.4
5
4.19 4.47
3.6
2.85 2.71 2.91
1.98
1.51 1.22
0.64
0 0 0.23
1.02
1.76 1.59
2.3 2.4 2.22
2.88 2.67
3.31 3.24

5 4.68

6.32 Class
7.65 1
8.47 2
8.92 3
10
50 40 30 20 10 0 10 20 30 40 50
Level (in %)

Fig. 1. Class-specic WTA values for different bonus and malus levels.

15 Limit of CHF 20k Limit of CHF 30k Panel of 50 phys. Panel of 100 phys. Target reduc. of 10pp
Willingness to accept (in CHF)

10
7.63

5
2.53
1.58 1.51 1.66
1.04
0.17 0.18
0
0.08
0.68
1.32 1.42
1.83

7.58 7.52 Class

10 1
2
3

LIM20 LIM30 PHY50 PHY100 TAR

Attributes

Fig. 2. Class-specic WTA values for stop-loss limits, panel size, and spending target reduction.

Table 7
Class-specic willingness-to-accept values

Class-specic WTA values

Class 1 Class 2 Class 3

Attributes WTA 95%-CI WTA 95%-CI WTA 95%-CI

BON10 1.02 1.45 0.59 1.59 2.07 1.10 4.68 7.67 1.69
BON20 1.76 2.45 1.06 2.67 3.47 1.87 7.65 12.54 2.77
BON30 2.22 3.03 1.40 3.24 4.19 2.30 8.92 14.62 3.22
BON40 2.40 3.23 1.57 3.31 4.26 2.36 8.47 13.95 2.99
BON50 2.30 3.14 1.45 2.88 3.75 2.00 6.32 10.66 1.98
MAL10 0.23 0.12 0.59 1.51 1.03 1.99 2.85 0.74 4.96
MAL20 0.64 0.06 1.21 2.71 1.91 3.52 5.40 1.56 9.25
MAL30 1.22 0.52 1.92 3.60 2.61 4.60 7.64 2.39 12.89
MAL40 1.98 1.19 2.77 4.19 3.10 5.28 9.57 3.18 15.97
MAL50 2.91 1.93 3.90 4.47 3.31 5.62 11.20 3.84 18.56
LIM20 0.17 0.37 0.72 0.18 0.29 0.65 1.58 0.27 3.44
LIM30 1.04 0.46 1.61 1.51 0.86 2.15 7.63 2.22 13.04
PHY50 1.83 2.61 1.05 1.32 1.97 0.66 7.58 13.13 2.03
PHY100 1.42 2.14 0.69 0.68 1.33 0.04 7.52 12.98 2.06
TAR 0.08 0.40 0.24 1.66 1.22 2.10 2.53 0.90 4.16
NETW 0.10 0.78 0.98 1.38 0.37 2.39 31.11 11.39 50.83

Note: WTA values are shown in CHF per insured per month, where 1 CHF 1.20 at 2011 exchange rates. The 95%-condence intervals (95%-CI) were estimated using the
delta method. The effects coding was considered when calculating the WTA values. WTA for the target reduction is in 10%-points.
120 M. Rischatsch / Journal of Health Economics 40 (2015) 109121
6.2.2. Stop-loss limit, physician-panel size and target reduction
Increasing the stop-loss limit from CHF 10,000 to 20,000 per
patient per year has no statistically signicant effect on network
participation as indicated by the condence intervals in Fig. 2.
As discussed previously, this imposes a higher risk but it also
permits networks to achieve higher cost savings that translate
into higher gain-sharing bonuses. However, increasing the limit
from CHF 10,000 to 30,000 requires a compensation payment to
the tune of CHF 1.04 (class 1), CHF 1.51 (class 2), and CHF 7.63
(class 3). The third class that is more likely composed of physi-
cians with surgical activity shows a willingness to accept that
is ve and seven times higher than the one of the other two
classes.
Increasing the panel size of physician working under the same
target from 10 to 50 or 100 physicians, respectively, decreases the
required compensation for all three classes. While physicians from
the rst and second class lower their demand by CHF 1.83 and CHF
1.32 for a panel of 50 physicians, those from the third class are
willing to give up CHF 7.58. The compensation reduces by CHF 1.42
(class 1), CHF 0.68 (class 2), and CHF 7.52 (class 3) if the panel size
increases from 10 to 100 physicians.
Acceptance of a lower spending target requires a compensation
payment for the second and the third class but not for the rst. The
compensation amounts to CHF 1.66 per 10%-points (pp) reduction
for the second class and CHF 2.53 for the third class.
7. Discussion and conclusion
In Switzerland, increasing health care expenditures have fueled
political debates on possible regulatory changes to promote Man-
aged Care (MC). Several studies have shown that gatekeeping
patients to health care providers and remunerating physicians
through capitation instead of conventional fee for service con-
tributes to lower health care expenditure and a higher treatment
quality. While MC constitutes the dominant form of health insur-
ance in the United States, it was less dominant in Switzerland
until recently. This has changed between 2008 and 2013 dur-
ing which period the market share for gatekeeping policies with
capitation has more than doubled. Even though MC is gaining
popularity in Switzerland, two major changes in regulation are dis-
cussed to further promote its market share. First, some politicians
want to force health insurers to write MC contracts. Second, the
introduction of budgetary co-responsibility for ambulatory care
physicians is believed to encourage physicians to better control
costs. Co-responsibility embodies that nancial responsibility is
shared between the health insurer and the physicians arranged in
physician networks. A grouping of physicians to networks is neces-
sary for risk diversication because single events are less inuential
with an increasing number of patients enrolled in the network. If
working in physician networks under a global budget is voluntarily,
a pivotal question is if the required payments to compensate physi-
cians for the additional nancial risk are lower than the potential
cost savings. In this article, the required willingness-to-accept val-
ues are derived and compared with potential cost savings from the
health insurers perspective.
The discrete choice experiment, completed by 872 physicians
in 2011, reveals a strong preference heterogeneity among Swiss
ambulatory care physicians. On the one hand, a rst group of
physicians mainly general practitioners is indifferent between
working in a network or in independent practice as long as no bud-
getary responsibility is imposed, but preferences change depending
on the design of the reimbursement mechanism. On the other hand,
a major share of physicians mainly specialists with surgical activ-
ities are unwilling to work in a physician network regardless
of how the reimbursement scheme is designed. Between these
extremes, two additional physician classes are identied: one class
that asks for a relatively low compensation to join a network with-
out taking budgetary responsibility and one class that has to be
strongly compensated. Using the estimated willingness-to-accept
values and assuming a reimbursement design with a symmetric
bonus and malus, a stop-loss limited of CHF 10,000 per insured per
year, a panel of 50 physicians, and a spending target that is 20%
below the conventionally insured collective, compensation pay-
ments between CHF 3.304.97 (low-compensation class) and CHF
26.3433.47 (high-compensation class) are required, depending on
the magnitude of the bonus and the malus. Again, one class of physi-
cians accepts such a design without asking for a compensation,
while a group of protesters is reluctant to change its willingness
to join a network in return for the risk payments modeled in the
choice experiment.
Comparing the estimated willingness-to-accept values with the
potential cost savings permits to assess if the introduction of bud-
getary co-responsibility contributes positively or negatively to the
health insurers prot as long as independent practice without
nancial responsibility remains an alternative. Reich et al. (2012)
estimate cost containments for capitated health insurance poli-
cies to be 21%. According to ofcial statistics provided by the BAG
(2011) average annual claims cost per MC insured adult (above 26
years) amounted to CHF 1,922 in 2011, excluding co-payments and
deductibles paid by the insured. Assuming that this MC collective
has 20% lower costs than the conventionally insured collective due
to efciency gains adjusted for the composition of the portfolios
about CHF 40 per insured per month can be saved by the insurer
with a target reduction of 20%. Comparing these savings with the
estimated willingness-to-accept values reveals that health insur-
ers are likely to achieve net savings by contracting with physician
networks.
Assuming 600 enrollees per physician a realistic value for
Switzerland a physician from the low-compensation class and
very likely a general practitioner would receive CHF 3.42 per
insured per month (or CHF 2052 per month) for the previously dis-
cussed reimbursement design in combination with a bonus and a
malus of 20%. According to Reichert (2010), the average monthly
income for a Swiss general practitioner is CHF 16,100. Therefore,
the payment would lead to a substantial increase in income by 13%.
This shows that it would be protable for both, general practi-
tioners and health insurers, to contract on such a reimbursement
design as long as physicians prove to be able to reduce treat-
ment cost to meet the spending target. Again, considering the
estimated efciency gains by Reich et al. (2012) this seems
possible. Nevertheless, the large share of physicians with a rela-
tively high compensation demand (42.5%) and the strong aversion
group (21%) show that a major acceptance of budgetary co-
responsibility is unlikely to be achieved on voluntary basis.
Acknowledgements
The author would like to express his thanks to Dr. med. I. Cassis
(FMH) for making the data collection possible. Support by Esther
Kraft is also gratefully acknowledged. Special thanks go to Prof. Dr.
P. Zweifel and my former colleagues M. Trottmann, H. Telser, Y.
Schneider, and Ph. K. Widmer for their very helpful comments. The
ndings, interpretations and conclusions expressed in this article
are entirely those of the author.
References
BAG, 2011. Statistik der obligatorischen Krankenversicherung 2011. In: Statistik der
Krankenversicherung., pp. 70.
 M. Rischatsch / Journal of Health Economics 40 (2015) 109121
Bech, M., Gyrd-Hansen, D., 2005. Effects coding in discrete choice experiments.
Health Economics 14 (10), 10791083.
Beck, K., Kser, U., Trottmann, M., Rotz, S.v., 2009. Efzienzsteigerung dank Managed
Care? (Efciency increase through Managed Care?). Data Master 5, 1521.
Berchtold, P., Hess, K., 2006. Evidenz fr Managed Care. In: Technical Report. Swiss
Health Observatory., pp. 16.
Berchtold, P., Peytremann-Bridevaux, I., 2010. Integrated care organizations in
Switzerland. International Journal of Integrated Care, 11.
Chernew, M.E., Mechanic, R.E., Landon, B.E., Gelb Safran, D., 2011. Private-payer
innovation in Massachusetts: the Alternative Quality Contract. Health Affairs
30 (1), 5161.
Croxson, B., Propper, C., Perkins, A., 2001. Do doctors respond to nancial incentives?
UK family doctors and the GP fundholder scheme. Journal of Public Economics
79, 375398.
Dusheiko, M., Gravelle, H., Jacobs, R., Smith, P., 2006. The effect of nancial incentives
on gatekeeping doctors: evidence from a natural experiment. Journal of Health
Economics 25, 449478.
Ellis, R.P., McGuire, T.G., 1986. Provider behavior under prospective reimbursement.
Journal of Health Economics 5, 129151.
Gaynor, M., Gertler, P., 1995. Moral hazard and risk spreading in partnerships. RAND
Journal of Economics 26, 591613.
Gaynor, M., Rebitzer, J.B., Taylor, L.J., 2004. Physician incentives in health mainte-
nance organizations. Journal of Political Economy 112 (4), 915931.
Gibbons, R., 1996. 1996 Incentives and careers in organizations. In: NBER Working
Paper Series 5705.
Gruen, B., Leisch, F., 2008. Flexmix version 2: Finite mixtures with concomitant
variables and varying and constant parameters. Journal of Statistical Software
28 (4), 135.
Hess, S., Ben-Akiva, M., Gopinath, D., Walker, J., 2011. Advantages of latent class over
continuous mixture of logit models. In: Working paper.
Ichniowski, C., Shaw, K., 2003. Beyond incentive pay: insiders estimates of the value
of complementary human resource management practices. Journal of Economic
Perspectives 17, 155180.
Kamakura, W., Wedel, M., Agrawal, J., 1994. Concomitant variable latent class mod-
els for conjoint analysis. International Journal of Research in Marketing 11 (5),
451464.
Kraft, E., 2010. 30273 rztinnen und rzte fr die Schweiz. Schweizer rztezeitung
92 (12), 440444.
Leisch, F., 2004. Flexmix: a general framework for nite mixture models. Journal of
Statistical Software 11 (8), 118.
Louviere, J.J., Hensher, D.A., Swait, J.D., 2000. Stated Choice Methods. Analysis and
Applications. University Press, Cambridge.
Luce, D., 1959. Individual Choice Behavior. Wiley and Sons, New York.
121
Manski, C.F., 1977. The structure of random utility models. Theory Decision 52,
229254.
Marshall, D.A., Johnson, F.R., Kulin, N.A., zdemir, S., Walsh, J.M., Marshall, J.K., Beb-
ber, S.V., Phillips, K.A., 2009. How do physician assessment of patient preferences
for colorectal cancer screening tests differ from actual preferences? A com-
parison in Canada and the United States using a stated-choice survey. Health
Economics 18 (12), 14201439.
McFadden, D., 1981. Econometric models of probabilistic choice. In: Manski, C.,
McFadden, D. (Eds.), Structural Analysis of Discrete Data with Applications. The
MIT Press, Cambridge, MA, pp. 198272.
McGuire, T.G., 2000. Physician agency. In: Culyer, A., Newhouse, J. (Eds.), In: Hand-
book of Health Economics, vol. 1. Amsterdam, North Holland, pp. 461536
(chapter 9).
NZZ, 2012. Wuchtiges Nein zur Managed-Care-Vorlage. In: Neue Zrcher Zeitung
17. Juni.
OECD, 2013. Health at a Glance: Europe 2013. OECD Publishing.
Reich, O., Rapold, R., Flatscher-Thoeni, M., 2012. An empirical investigation of the
efciency effects of integrated care models in Switzerland. International Journal
of Integrated Care 12, 112.
Reichert, M., 2010. Einkommensverhltnisse der freien rzteschaft der Schweiz in
den Jahren 2006 [neu] und 2005 [Re-Evaluation]. Schweizer rztezeitung 91,
479487.
Rischatsch, M., 2014. Lead me not into temptation: drug price regulation and dis-
pensing physicians in Switzerland. European Journal of Health Economics 15,
697708.
Rischatsch, M., Trottmann, M., Zweifel, P., 2013. Generic substitution, nancial inter-
ests, and imperfect agency. International Journal of Health Care Finance and
Economics 13 (2), 115138.
Rischatsch, M., Zweifel, P., 2012. What do physicians dislike about Managed Care?
Evidence from a choice experiment. European Journal of Health Economics 14
(4), 601613.
Robinson, J.C., 2001. Theory and practice in the design of physician payment incen-
tives. The Milbank Quarterly 79, 149177.
Train, K.E., 2003. Discrete Choice Methods with Simulation. University Press, Cam-
bridge.
Trottmann, M., Zweifel, P., Beck, K., 2012. Supply-side and demand-side cost shar-
ing in deregulated social health insurance: which is more effective? Journal of
Health Economics, 231242.
Wedel, M., DeSarbo, W.S., 2002. Mixture regression models. In: Hagenaars, J.A.,
McCutcheon, A.L. (Eds.), Applied Latent Class Analysis. Cambridge University
Press, pp. 366382.
Zweifel, P., Eisen, R., 2012. Insurance Economics. Heidelberg, Springer.
 Journal of Health Economics 40 (2015) 122131

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

Long-term care insurance: Does experience matter?

Norma B. Coe a,b,c, , Meghan M. Skira d , Courtney Harold Van Houtven e,f,b
a
Center for Retirement Research at Boston College, United States
b
Netspar, The Netherlands
c
NBER, United States
d
University of Georgia, United States
e
Durham Veterans Affairs Medical Center, United States
f
Duke University Medical Center, United States

a r t i c l e i n f o a b s t r a c t

Article history: We examine whether long-term care (LTC) experience helps explain the low demand for long-term care
Received 25 October 2012 insurance (LTCI). We test if expectations about future informal care receipt, expectations about inheritance
Received in revised form receipt, and LTCI purchase decisions vary between individuals whose parents or in-laws have used LTC
31 December 2014
versus those who have not. We nd parental use of a nursing home decreases expectations that ones
Accepted 3 January 2015
children will provide informal care, consistent with the demonstration effect. Nursing home use by in-
Available online 12 January 2015
laws does not have the same impact, suggesting that individuals are responding to information gained
about their own aging trajectory. Nursing home use by either a parent or in-law increases LTCI purchase
JEL classication:
G22
probability by 0.8 percentage points, with no signicant difference in response between parents and
D81 in-laws use. The estimated increase in purchase probability from experience with LTC is about half the
J14 previously estimated increase from tax policy-induced price decreases.
2015 Elsevier B.V. All rights reserved.
Keywords:
Long-term care
Insurance
Informal care
Expectations
Behavioral economics

1. Introduction
Long-term care (LTC) is one of the largest nancial risks facing
the elderly today, yet very few people 13 percent of current 65
We would like to thank participants and discussants at the 2009 iHEA con-
ference, the 2010 AcademyHealth Health Economics Workshop, the 2010 and 2012
ASHEcon conferences, the Aging Workshop at the NBER Summer Institute, Financial
Capability Research Workshop, International Long-Term Care Network, and Boston
College for helpful comments. We are especially thankful to Richard Frank and David
Cutler for their insights. We would also like to thank the John A. Hartford Foundation
Pilot Studies Grant through Duke University (Coe and Van Houtven), the Veterans
Affairs Health Services Research and Development Merit Review Program [MRP 05-
311] (Van Houtven), and the National Institutes of Health [NIH 1R01NR13583] (Coe
and Van Houtven) for funding. Any remaining errors are our own. The opinions and
conclusions expressed are solely those of the authors and do not represent the opin-
ions of Boston College, University of Georgia, University of Washington, NBER, Duke
University, or the Department of Veterans Affairs.
Corresponding author. Present address: University of Washington School of Pub-
lic Health, 1959 NE Pacic Street, Seattle, WA 98195-7660, United States.
Tel.: +1 206 616 8530.
E-mail address: nbcoe@uw.edu (N.B. Coe).
year olds insure against the risk. This lack of insurance cover-
age has spurred much work that tries to explain this phenomenon
(Finkelstein et al., 2005; Finkelstein and McGarry, 2006; Brown and
Finkelstein, 2007, 2008; Brown et al., 2007). It has also spurred pub-
lic policy into action, with 24 states and the District of Columbia
offering tax breaks for private long-term care insurance (LTCI) pur-
chases as of 2008 (Goda, 2011). These policy changes, however,
have not led to a widespread increase in coverage (Wiener et al.,
2000; Meiners, 2001; Courtemanche and He, 2009; Goda, 2011).
The Patient Protection and Affordable Care Act also included a now-
defunct program, the Community Living and Supported Services
(CLASS) Act, which would have created a new government-run LTCI
pool.
In other domains of catastrophic insurance, studies have shown
that after individuals experience the insurable risk, they are
more likely to purchase insurance against that risk in the future
(Kunreuther et al., 1978; Kunreuther, 1996; McCall et al., 1998;
Browne and Hoyt, 2000; Gallagher, 2014). This response could
be explained by behavioral heuristics, such as an increase in the
salience of the risk, or by an increase in information about the
insurable risk.

http://dx.doi.org/10.1016/j.jhealeco.2015.01.001
0167-6296/ 2015 Elsevier B.V. All rights reserved.
 N.B. Coe et al. / Journal of Health Economics 40 (2015) 122131
In the LTCI context, lack of experience with LTC could help
explain the low demand for insurance. Experience could help over-
come behavioral biases since people tend to dislike and delay
discussions about becoming frail and needing assistance (Sperber
et al., 2014). In addition, experience could ll several potential
information gaps in the LTCI context: the likely need for LTC; the
expense entailed in getting such care; the gaps in ones exist-
ing insurance coverage for LTC; or, ones familys willingness to
provide informal care. Any of these sources of incomplete informa-
tion would lead to lower than expected LTCI purchases. Previous
research has provided evidence of incomplete information in at
least two of these domains. Bacon et al. (1989), Pauly (1990), and
AARP (2006) report that most elderly believe that LTC expenses
are covered by Medicare. Further, AARP (2006) found considerable
underestimation of the costs of nursing home care among individ-
uals age 45 and older. While Brown and Finkelstein (2009) highlight
limited consumer understanding of both the insurance product and
the expenditure risk as one of the top ve reasons for low LTCI
demand, there has been no empirical testing, to our knowledge, of
the extent to which lack of information depresses LTCI purchase in
the United States.1
To evaluate the role experience may play in LTC and LTCI deci-
sions, we exploit variation between individuals who have someone
close to them a parent or in-law who needs and uses LTC and
those who do not.2 Individuals who have parents or in-laws who
use LTC may be more aware of the risks and costs. This paper
examines how parents and in-laws use of LTC impacts an adult
childs LTCI purchase decisions. Although we cannot completely
disentangle whether the effects are driven by a change in informa-
tion or behavioral economics explanations, we exploit variation in
the person who needed care (parent or in-law), the type of care
used (nursing home or informal care), and the person who pro-
vided informal care (the adult child him or herself or a sibling)
to discern which explanations may be at play. We also analyze
the mechanisms underlying any resulting change in LTCI holding
by examining how expectations about the future change after a
parent or in-law uses LTC. In particular, we analyze how a parent
or in-laws use of LTC impacts an adult childs expectations about
informal care from children and inheritance receipt.
The rest of the paper is as follows: Section 2 describes how
experience with LTC via parents or in-laws use can affect expec-
tations and LTCI purchase decisions. Section 3 discusses the data
and the sample in detail. Section 4 presents our empirical strategy.
Section 5 presents the results, in which we examine the impact
of parents and in-laws LTC use on the adult childs expectations
about informal care and inheritance receipt, and on the purchase
of LTCI, controlling for these expectations. Section 6 concludes that
experience does matter in the decision to purchase LTCI. While the
experience effect is slightly smaller than that of other policies tried,
such as decreasing the Medicaid asset limits (Brown et al., 2007)
or the recent tax credits and partnership programs (Goda, 2011),
it does not have the same associated costs to the federal or state
budget.
2. How experience can affect expectations
In the specic domain of risk perception and insurance, previ-
ous work has shown that individuals who experience the insurable
1
Zhou-Richter et al. (2010) assess the role of information among adult children
on the demand for LTCI for their parents in Germany.
2
While LTC can encompass many types of care arrangements, in the empirical
framework, due to data limitations, we limit the term to only mean two types
informal care from children and nursing home care.
123
risk are more likely to purchase insurance against that risk in the
future. Kunreuther et al. (1978) nd that individuals who experi-
enced property damage in a prior natural disaster are more likely
to purchase ood insurance. Similarly, ood insurance purchases
increase in areas that experienced a recent ood (Browne and Hoyt,
2000; Gallagher, 2014). Palm et al. (1990) found an increase in
the perceived need for earthquake insurance after the Loma Pri-
eta earthquake, and Kunreuther (1996) nds an increase in the
demand for insurance following earthquakes more generally. Coe
et al. (2014) nd an increase in the stated demand for disability
insurance among individuals who are informed about disability
risks.
In the LTC literature, some studies have explored how knowl-
edge about LTC or individual health risks impact LTCI demand
behavior. For example, McCall et al. (1998) nd that individuals
who have a family member or friend that needs LTC are more
likely to purchase LTCI, and Zhou-Richter et al. (2010) nd that
German adult children are more likely to want to buy LTCI for their
parents after being informed about the population average risks
and costs of LTC. Brown et al. (2012) nd that individuals beliefs
about the likelihood of not being able to live independently due
to health problems were strongly correlated with LTCI coverage in
the RAND American Life Panel. Meier (1999) formulates a theoret-
ical model that shows individuals might wait to purchase LTCI to
obtain more information about their own probability of becoming
disabled as well as the costs of disability and associated care. Our
paper contributes to this literature by examining how actual expe-
rience with LTC via a parent or in-laws use of nursing home or
informal care impacts LTCI purchase behavior, exploiting variation
in the relationship to the care recipient, type of care received, and
who provided care.
Disentangling why experience with LTC matters in forming
expectations and LTCI purchase decisions is often difcult with
secondary data, but important in order to make policy recom-
mendations. We explore three potential explanations below: (1)
rational updating; (2) overcoming inertia and increasing salience;
and, (3) overreaction to risks.
2.1. Rational updating
The rational expectations model assumes that individuals form
expectations using all available information and make optimal
decisions that are based on (and consistent with) these expecta-
tions. When new information is learned, expectations about the
future are updated, typically assumed in a Bayesian manner. The
rational expectations model is consistent with the self-regulation
model (Leventhal et al., 1992). The self-regulation model posits that
individual understanding of disease is based on the individuals
rationalizations about its causes, consequences, timeline, trajec-
tory, prognosis, and their ability to control, treat, and prevent the
illness. Individual understanding is informed by beliefs, experi-
ences, family, friends, health care practitioners, and the media.
Importantly, both positive and negative experiences or information
can alter ones beliefs, coping strategies, and behavior.
The rational expectations model posits that LTC use by par-
ents and parent-in-laws (hereafter, the G1 generation) could impact
ones knowledge of LTC risks and costs. Under this framework, we
hypothesize that LTC use by the G1 generation will have an informa-
tion spillover effect and change the insurance status of their adult
children. If individuals are already fully informed about the risks
they face and their associated costs, then LTC use by parents or in-
laws will have no effect on their own expectations about future use
or their nancing plan. However, if individuals learn during their
parents or in-laws LTC spell, expectations about future care use,
lifetime income, and preferences for different types of care delivery
124 N.B. Coe et al. / Journal of Health Economics 40 (2015) 122131
may evolve, and the decision to purchase LTCI to prepare nancially
for these future needs may change.
We try to disentangle the type of information that is learned
through the G1 generations use of LTC by exploiting across-family
variation in who provides care, who receives care, and the type of
care received. For example, helping or watching anyone navigate
the LTC system will make one aware of the complexities they may
face during their own aging process. Thus, information about the
aging process and the costs associated with care are revealed if any
individual in the G1 generation uses any type of LTC. Additional
information can be learned as outlined below.
Relationship to care recipient: Genetic information is revealed
through a health decline of a parent, but not an in-law. Because
of the common genetic link, a parents health decline may differ-
entially inuence an adult childs expected need for future LTC
compared to an in-laws health decline.
Type of care received: Whether a parent or in-law receives formal
or informal care could impact an adult childs expectations or
preferences about his or her own future care needs. The type of
care received may provide information about the costs of care,
the willingness of family to provide informal care, and the ability
to meet LTC needs in an informal setting.
Provision of care: As a result of providing informal care to a mem-
ber of the G1 generation, an individual may experience and learn
about caregiver burden, which has many denitions in the geron-
tology literature (Zarit et al., 1980) or conversely, caregiver gain,
such as role satisfaction (Tarlow et al., 2004), improved relation-
ships with a parent (Roth et al., 2005), and the knowledge that
one has fullled a familial duty (Van Houtven and Norton, 2004).
Providing care to a parent or in-law may also provide a signal that
an adult child expects his or her own children to do the same (i.e.
the demonstration effect) (Stark, 1995; Cox and Stark, 2005). We
hypothesize that an individual who provides informal care will
learn more about this aspect of LTC than non-care providing adult
children.
One complicating factor is that parents or in-laws use of LTC
might impact ones lifetime income expectations through chang-
ing the size or the probability of receiving a bequest; thus, we also
examine how expectations about inheritance receipt vary with LTC
use in the G1 generation. For example, a G1 member receiving infor-
mal care might change expectations about who gets the bequest if
the exchange motive is operative (Bernheim et al., 1985; Norton
and Van Houtven, 2006), while formal care use may impact the
expected size of the bequest, particularly if such care is paid out-
of-pocket.
Further, we examine whether there are direct or indirect effects
of the G1 generations use of LTC on how adult children plan to pay
for their future care needs by analyzing their decision to purchase
private LTCI. The hypothesized changes in the perceived need for
insurance described above are consistent with Bayesian updating
within a rational expectations framework (Viscusi, 1991). In other
words, individuals gain new information from the G1 generations
use of LTC and update their own expectations, which could result
in a change in behavior in the form of purchasing LTCI.
2.2. Overcoming inertia and increasing salience
Changes in expectations and LTCI purchase decisions among
individuals with more experience with the insurable risk may
also be consistent with models proposed in behavioral economics,
based on the psychology literature. Two likely concepts at play in
the LTC and LTCI context are inertia and salience.
Inertia: When faced with the difcult calculation of determining
whether and how much LTCI is needed, individuals may simply
follow the status quo from a previous decision or avoid making a
decision altogether. In this case, the status quo is non-purchase,
and could lead to lower levels of insurance rates than predicted
to be optimal from a rational expectations standpoint. Having
someone in the G1 generation use LTC may be a precipitating
event to end the inertia.
Salience and availability bias: Salience refers to the phenomenon
that when ones attention is differentially directed to one portion
of the environment rather than to others, the information con-
tained in that portion will receive disproportionate weighting
in subsequent judgments (Taylor and Thompson, 1982). Having
someone in the G1 generation use LTC may increase the salience
of LTC and LTCI decisions for the adult child. The behavioral
economics literature has noted that salient risks (even low-
probability ones) are often given excessive weight and result in
excessive reactions by individuals. Closely related, availability
bias means individuals base the probability of an event on the
ease with which instances of the event can be brought to mind
(Tversky and Kahneman, 1974). Thus, changes in expectations
and LTCI purchase in response to LTC use by the G1 generation
may reect adult children putting excessive weight on LTC risks
because they are salient or can be brought to mind easily, rather
than a rational updating of the information set.
2.3. Changing risk perception
Kahneman and Tversky (1979) observe that people have
limited ability to comprehend and evaluate extreme probabil-
ities. Indeed, the literature has found that individuals tend to
overestimate low-probability events. Coe and Webb (2009) suggest
that people tend to overestimate their probability of receiving an
inheritance; however, previous work such as Lindrooth et al. (2000)
and Taylor et al. (2005) suggests that average nursing home utiliza-
tion predictions from individuals in the early years of the AHEAD
data are fairly accurate compared with their actual nursing home
usage 5 to 7 years later. Another phenomenon concerning prob-
ability perception is the tendency to overreact to fearsome risks
and overestimate their frequency (Sunstein and Zeckhauser, 2011).
Thus, changes in expectations and LTCI purchase following LTC
use by the G1 generation may capture adult children excessively
reacting to LTC risks.
It is difcult to separately identify the mechanisms underlying
behavior changes following experience with LTC using secondary
data sources that do not specically ask why individuals changed
their expectations or decision to purchase LTCI (or without a full
parameterization of the decision-making model). However, by
exploiting variation in the relationship to the care recipient, the
type of care received, and who provided care, we might take some
steps to disentangle the different explanations. For example, within
a married household with shared decision-making, it is difcult to
use the salience argument to explain large differential responses
to in-laws using LTC and parents using LTC. We generally discuss
our ndings within the context of the rational expectations and
Bayesian learning framework, but we acknowledge that for the
most part, we cannot completely rule out or disentangle the role
of the behavioral tendencies and non-learning responses outlined
above.
3. Data
We use data from the 1998 through 2006 waves of the Health
and Retirement Study (HRS). In 1992, the HRS sampled the
 N.B. Coe et al. / Journal of Health Economics 40 (2015) 122131
non-institutionalized 5161 year old US population, and respon-
dents have been followed every two years. We limit the years of
data due to question differences across years. We start with the
1998 wave due to the inconsistent denition of LTCI in the survey
and other key variables prior (Finkelstein and McGarry, 2006),
and end in 2006 due to the lack of questions about inheritance
expectations in later years.
3.1. Sample
We create a respondent-level dataset consisting of respondents
from the original HRS cohort and their spouses. We limit our anal-
ysis to individuals over the age of 50. Given we are interested in
expectations about informal care from children, an individual has
to have at least one living child to be included in the sample. Since
we are interested in LTCI purchase, an individual must not have LTCI
in the prior survey wave and must be able to pass the LTCI under-
writing screening as of the prior wave (i.e. no IADLs or ADLs and
no recent strokes).3,4 Table A1 in the Appendix details our sam-
ple restrictions. Our nal estimation samples consists of 26,979
person-wave observations, representing 8,349 unique individuals.
3.2. Variables
First, we analyze how G1 LTC use impacts expectations about
informal care and inheritance receipt. Regarding future informal
care expectations, the HRS asks, Suppose in the future, you needed
help with basic personal care activities like eating or dressing. Do
you have relatives or friends (besides your [husband/wife/partner])
who would be willing and able to help you over a long period of
time? If the respondent says yes, he or she is asked, What is the
relationship to you of that person or persons? Respondents are
allowed to indicate multiple options from among a variety of people
listed. If respondents include child/child-in-law in the follow-up,
we code them as expecting children to provide informal care in the
future.
Having a parent use LTC could impact expectations about future
inheritance receipt (either the amount or the likelihood of receiving
one), and consequently impact ones expectations about lifetime
wealth. Such changes could subsequently impact the nancial plan-
ning that he or she may need to undertake to handle LTC expenses
in the future, including the choice between self-insuring and hav-
ing private LTCI. The HRS asks, What are the chances you will
receive an inheritance during the next 10 years? Individuals can
respond with a probability between 0 and 100, which we rescale
to be between 0 and 1.5
Second, we test the inuence of the G1 generations use of
LTC on LTCI purchase, controlling for the adult childs expecta-
tions. Regarding LTCI, the HRS asks, Not including government
programs, do you now have any long term care insurance which
specically covers nursing home care for a year or more or any part
of personal or medical care in your home? Individuals can respond
yes or no. We dene purchase to mean an individual answered
yes to the LTCI question in the current survey wave but answered
no in the prior wave. As mentioned earlier, we examine purchase
only among individuals healthy enough to pass the underwriting
3
IADLs are instrumental activities of daily living which include food preparation,
housekeeping, managing nances, taking medication, among other activities. ADLs
are activities of daily living, which including bathing, dressing, eating, toileting, and
transferring (walking).
4
We focus on LTCI purchase and not LTCI coverage to ensure the potential infor-
mation gained from G1 LTC use occurs before the LTCI purchase.
5
In 2006, the question wording changed to What are the chances that you (or
your [husband/wife/partner]) will receive an inheritance during the next 10 years?
125
screening as of the prior survey wave. The primary explanatory
measures of interest are whether parents and in-laws received LTC
and which type of care. We can identify informal care and nurs-
ing home care use of the G1 generation. The HRS asks respondents
where each of their parents and in-laws currently lives, with one
of the possible responses being in a nursing home. In addition, if a
parent or in-law has passed away since the prior survey wave, the
HRS asks respondents whether that parent or in-law ever lived in
a nursing home. We categorize parents or in-laws who currently
live or ever lived in a nursing home as having utilized nursing home
care.6
In order to measure informal care provided to the G1 generation,
the HRS asks each respondent if he or she and his or her spouse
spent a total of 100 or more hours in the last two years helping
each parent or in-law with basic personal activities like dressing,
eating, and bathing. The survey then asks who was helped and
how many hours of care the respondent, and separately his or her
spouse provided. The survey also asks which siblings and sibling
in-laws, if any, provide help with these tasks.7
4. Empirical strategy
In order to examine the effect of LTC use of the G1 generation on
the adult childs expectations about informal care provision from
his or her children, we estimate the following equation using a logit
model:
Pr(ExpectIC it = 1) = f (0 + 1 Ait + 2 Hit + 3 IC it + 4 Zit
+ 5 Fit + 6 wt + it ), (1)
where ExpectICt is an indicator variable for whether individual i at
time t expects that his or her children will be willing to provide
informal care in the future.
Ait is a vector of LTC received by parents and in-laws, measuring
if LTC was ever provided before or during time t. In our estimation,
we consider three separate vectors of LTC received by the G1 gen-
eration to delineate across types of information provided: (1) Any
parent or in-law receives any type of LTC; (2) Any parent or in-law
receives nursing home care versus any parent or in-law receives
informal care; (3) Own parent uses nursing home care; own parent
receives informal care from the respondent; own parent receives
informal care from the respondents siblings, and the three in-law
correlates. Hit is a vector of the adult childs health information,
included to capture risk factors for future LTC needs. This vector
includes the number of ADLs, the number of IADLs, an indicator for
a low cognition score, the number of chronic conditions, an indica-
tor for incontinence problems, and indicators for whether the adult
child has previously used nursing home, home health, or informal
care in the past two years. ICit represents the potential informal
care network available to the respondent (besides a spouse) includ-
ing the number of male and female children, the number of living
siblings, the number of co-resident children, and the number of
children living within 10 miles (excluding co-resident children).
The individual-level characteristics included in the vector Zit are
age, age squared, gender, race, and education categories (less than
high school, high school, and at least some college) for the respon-
dent and his or her mother and father. We also include an indicator
for being married, being divorced, and an interaction term between
6
The HRS does not contain information about formal home health care use of
parents or in-laws. As a result, unfortunately, we cannot analyze how G1 generation
use of formal home health care affects expectations or LTCI purchase.
7
We do not include help with chores, errands, or transportation in our denition
of informal care in order to keep the denition consistent between respondents and
siblings, for whom we only know about personal care provision.
126 N.B. Coe et al. / Journal of Health Economics 40 (2015) 122131

gender and divorced, as well as an indicator for whether all parents Table 1
Descriptive statistics.
and in-laws are deceased. Fit is a vector of nancial information.
This includes annual nancial wealth quartiles, annual housing Variable Average
wealth quartiles, whether the respondent and his or her spouse Outcomes: expectations and LTCI
have assets in a trust, whether the respondent has a will that ben- Expect informal care from children/children-in-law 0.505
ets his or her children, and whether the children hold the deed Probability of receiving inheritance 0.151
to the house. Last, we include survey wave indicators, wt . The Purchased LTCI 0.055
G1 Long-term care measures
standard errors are clustered at the individual level to account for
Any G1 received LTC 0.689
repeated observations. Any G1 received nursing home care 0.520
To analyze the effect of LTC use of the G1 generation on the Any G1 received informal care 0.422
adult childs reported probability of receiving an inheritance, we Parent used nursing home care 0.343
In-law used nursing home care 0.298
estimate the following fractional logit model:
Parent received informal care from respondent 0.156
exp(Xit ) Parent received informal care from respondents sibling(s) 0.219
E(ExpectInher it ) = , (2) In-law received informal care from respondent 0.084
1 + exp(Xit )
In-law received informal care from respondents sibling(s)-in-law 0.176
where ExpectInherit is individual is self-reported probability at time Health variables
# of ADLs 0.073
t of receiving an inheritance, taking on values from 0 to 1 inclusive,
# of IADLs 0.021
and Low cognition 0.067
Incontinent 0.148
Xit = 0 + 1 Ait + 2 Hit + 3 IC it + 4 Zit + 5 Fit + 6 wt . (3) # of chronic conditions 1.504
Ever used informal care 0.009
Ait , Hit , ICit , Zit , Fit , and wt are as dened above. The fractional logit
Ever used formal home health care 0.067
model developed by Papke and Wooldridge (1996) accounts for Ever used nursing home care 0.008
the fact that the dependent variable is a proportion and can take Financial variables
on values of 0 or 1. Has a trust 0.080
We then examine the impact of the G1 generation receiving LTC Has children on deed to house 0.033
Has children on will 0.488
on LTCI purchase, adding expectations about future informal care
Demographic variables
and inheritance receipt to the regression in order to disentangle Age 64.713
the direct effect of experience on LTCI purchase from the indi- Female 0.581
rect effects that operate via changes in expectations. Therefore, we Married 0.736
Divorced 0.123
estimate the following logit model:
Non-white 0.148
Pr(LTCI it = 1) = f (0 + 1 Ait + 2 ExpectIC it + 3 ExpectInher it Education:
Less than high school 0.227
+ 4 Hit + 5 IC it + 6 Zit + 7 Fit + 8 wt + it ), (4) High school graduate 0.372
Some college 0.401
where LTCIit is an indicator variable for reporting having an LTCI Mothers education:
policy in period t conditional on not having one in period t 1. Ait , Less than high school 0.539
High school graduate 0.278
ExpectICit , ExpectInherit , Hit , ICit , Zit , Fit , and wt are as dened above.
Some college 0.104
Missing 0.078
5. Results Fathers education:
Less than high school 0.553
High school graduate 0.226
Table 1 presents the descriptive statistics from our estimation
Some college 0.114
sample. About 50 percent of the sample expect their children to Missing 0.107
provide informal care in the future. The average expected prob- All parents and in-laws deceased 0.548
ability of receiving an inheritance in the future is 15 percent. Informal care network variables
Approximately 5 percent of the sample ever purchases LTCI during # of living siblings 2.850
# of male children 1.850
the period we consider. Almost 70 percent had some interaction # of female children 1.773
with the LTC system via their parents or in-laws. About 52 percent # of co-resident children 0.293
of the G1 generation used a nursing home while 42 percent used # of children within 10 miles 0.982
informal care (these are not mutually exclusive options). # of person-wave observations 26,979
The average age in the sample is almost 65. There are slightly
more females than males, and 74 percent of the sample is married.
Very few of the respondents have already used LTC themselves sample is married, and individuals have almost 3 living siblings,
likely reecting the sample selection criteria of being able pass the 2 sons, and 2 daughters, on average. In addition, one of the respon-
LTCI underwriting screening. Not even 1 percent of the sample has dents children lives within 10 miles, on average.
used informal or nursing home care, but almost 7 percent has used Table 2 presents the marginal effects and standard errors from
formal home health care in some capacity.8 the logit model estimating the impact of G1 LTC use on whether
On average, the sample has many potential informal care an individual expects his or her children will provide informal care
providers. As mentioned earlier, almost three-quarters of the in the future.9 Column 1 presents the marginal effects from the
model estimating the effect of any G1 member using any type of
LTC. Column 2 breaks down G1 LTC use between nursing home
8
While controlling for prior LTC use seems logical when estimating expectations and informal care. Column 3 presents the results when G1 LTC
about future informal care, using LTC oneself likely informs individuals about LTC
costs and probabilities above and beyond the information set learned from the G1
generation receiving LTC. We have rerun our baseline equations eliminating those
9
with their own experience with LTC, and we nd quantitatively similar results which The coefcients and marginal effects of the control variables are largely as
we discuss later. expected. Full results are available upon request.
 N.B. Coe et al. / Journal of Health Economics 40 (2015) 122131 127

Table 2 Table 3
Expectations about future informal care provision by children. Expectations about future inheritance receipt.

I II III I II III

Any G1 used LTC 0.0166 *

Any G1 used LTC 0.00525
(0.00888) (0.00522)
Any G1 used nursing home care 0.0223 ***
Any G1 used nursing home care 0.0116**
(0.00848) (0.00502)
Any G1 used informal care 0.00586 Any G1 used informal care 0.00179
(0.00851) (0.00493)
Parent used nursing home care 0.0267*** Parent used nursing home care 0.0294***
(0.00885) (0.00524)
In-law used nursing home care 0.00378 In-law used nursing home care 0.00681
(0.00943) (0.00564)
R provided care to parent 0.0154 R provided care to parent 0.0142**
(0.0124) (0.00708)
Sibling provided care to parent 0.00122 Sibling provided care to parent 0.00960
(0.0111) (0.00631)
R provided care to in-law 0.0197 R provided care to in-law 0.0116
(0.0156) (0.00867)
Sibling-in-law provided care to Rs in-law 0.00104 Sibling-in-law provided care to Rs in-law 0.0171***
(0.0118) (0.00632)

Observations 26,979 26,979 26,979 Observations 26,979 26,979 26,979

Notes: R denotes the HRS respondent. Marginal effects are shown and the standard Notes: R denotes the HRS respondent. Marginal effects are shown and the standard
errors of those marginal effects are in parentheses. Standard errors are clustered at errors of those marginal effects are in parentheses. Standard errors are clustered at
the individual level. Signicance levels: * p < 0.1, ** p < 0.05, *** p < 0.01. the individual level. Signicance levels: * p < 0.1, ** p < 0.05, *** p < 0.01.

Table 4
use is broken down into six categories: parent uses nursing home
Long-term care insurance purchase.
care, respondent provides informal care to parent, respondents
sibling provides informal care to parent, and the three in-law cor- I II III
relates. Any G1 used LTC 0.00302
We nd evidence that an individuals expectations about infor- (0.00315)
mal care from their children or children-in-law are impacted by Any G1 used nursing home care 0.00884***
(0.00299)
nursing home use of the G1 generation, and this effect is driven by Any G1 used informal care 0.00425
the respondents own parents nursing home use. If the respon- (0.00297)
dents parent uses a nursing home, he or she is 2.7 percentage Parent used nursing home care 0.00837***
points less likely to expect his or her child will provide care in (0.00319)
In-law used nursing home care 0.00631*
the future. This result is consistent with the demonstration effect.
(0.00326)
If an individuals parent uses a nursing home, that individual has R provided care to parent 0.00422
not signaled that he or she expects or prefers to receive informal (0.00429)
care from his or her children, which may decrease expectations Sibling provided care to parent 0.000931
about future informal care receipt. This result is also consistent (0.00401)
R provided care to in-law 0.00856
with some of the behavioral responses outlined above. For exam-
(0.00523)
ple, if the parents use of nursing home care is salient or can Sibling-in-law provided care to Rs in-law 0.000329
be brought to mind easily, individuals might give disproportion- (0.00409)
ate weight to the probability of using nursing home care in the Informal care expectations 0.00591** 0.00604** 0.00610**
(0.00290) (0.00290) (0.00290)
future, and consequently report that they do not expect infor-
Inheritance expectations 0.00320 0.00347 0.00399
mal care. However, it is harder to use salience or risk perception (0.00469) (0.00469) (0.00471)
as an explanation for the large difference in effects of in-law
Observations 26,979 26,979 26,979
versus parental use of nursing home care on informal care expec-
tations. Notes: R denotes the HRS respondent. Marginal effects are shown and the standard
errors of those marginal effects are in parentheses. Standard errors are clustered at
Next, we explore the effect of G1 LTC use on the self-reported
the individual level. Signicance levels: * p < 0.1, ** p < 0.05, *** p < 0.01.
probability of receiving an inheritance. Marginal effects and
standard errors from the fractional logit model are presented
in Table 3. We nd that the direction of the effect depends on inheritance, while sibling-in-law provision of informal care to an
the type of care the G1 generation utilized. After an individuals in-law decreases that likelihood suggests that individuals believe
parent uses nursing home care, he or she decreases expectations an exchange or bequest motive is operative (i.e. that individuals
about future inheritance receipt by 2.9 percentage points. If a who provide care will be rewarded with a bequest).
respondent provides informal care to his or her own parent, he or We analyze the effect of LTC use among the G1 generation on
she increases inheritance expectations by 1.4 percentage points, LTCI purchase, controlling for informal care and inheritance expec-
but if a sibling-in-law provides care to the respondents in-law, tations, to examine if there is a direct effect of experience on LTCI
he or she decreases inheritance expectations by 1.7 percentage purchase. Results are shown in Table 4. In Column 2, we nd having
points. The decrease in expected inheritance receipt after a parent a G1 member use a nursing home increases the probability of LTCI
uses nursing home care may be explained by the child observ- purchase by 0.9 percentage points. In particular, having a parent
ing that there is less wealth to inherit as a result of the parent use a nursing home increases the probability of LTCI purchase by
spending down to qualify for Medicaid or to pay for nursing home 0.8 percentage points and having an in-law use a nursing home
care out-of-pocket. The nding that informal care provided to a increases the probability of purchase by 0.6 percentage points,
parent by the respondent increases the likelihood of receiving an though the latter effect is marginally signicant (p = 0.053). Given
128 N.B. Coe et al. / Journal of Health Economics 40 (2015) 122131
that these effects are not statistically different from each other, our
results suggest that by observing someone in the G1 generation
receive nursing home care, the adult child changes his or her nan-
cial planning strategy to handle future LTC expenses. This change
might be driven by information learned about the costs of nursing
home care and the lack of public insurance coverage for nursing
home expenses. It is also consistent with behavioral responses. If
individuals typically follow the status quo of non-purchase, seeing
a G1 member use nursing home care may be the impetus to end that
inertia. It is important to note that informal care does not trigger the
same response. This could be due to different information learned
through informal caregiving versus nursing home care regarding
formal care expenses. Further, it is suggestive that the increase in
LTCI purchase probability is not only driven by an overreaction to
low-probability risks.
Interestingly, we nd individuals who expect their children
will provide informal care for them in the future are 0.6 per-
centage points more likely to purchase LTCI. While this result is
unexpected, it might reect that some individuals who know or
expect that their children will provide care want to avoid bur-
dening their children as diseases progress and purchase LTCI. For
example, qualitative work suggests that parents want their chil-
dren to provide support but not be burdened should the need
for LTC arise (Sperber et al., 2014). This result may also reect
that individuals who expect their children to provide informal
care have already had discussions about future LTC alternatives
and are relatively more forward-looking, leading them to purchase
LTCI.
We nd inheritance expectations have no signicant effect on
the probability of purchasing LTCI. Given the median inheritance
in the U.S. is less than one year of average nursing home costs, this
non-result is not surprising (Wolff and Gittleman, 2014).
Since G1 nursing home use decreases expectations about infor-
mal care receipt, and these expectations inuence LTCI purchase,
we run a simple counterfactual exercise to estimate the total effect
G1 nursing home use has on LTCI purchase probability. We use
the estimated models to predict a counterfactual case where no
one had a member of the G1 generation use a nursing home and
then a separate counterfactual where everyone had a G1 gener-
ation member who received nursing home care.10 We perform
the counterfactual on the informal care and inheritance expecta-
tions equations, and then on the LTCI purchase equation accounting
for the expectations as well as the direct effects of G1 nursing
home use on LTCI purchase. This counterfactual exercise suggests
that having either a parent or an in-law use nursing home care
increases LTCI purchase probability by 0.7 to 0.8 percentage points.
One interpretation of this nding is that most of the information
learned through these experiences is about LTC more generally.
This information could be about the costs of LTC, the incomplete
insurance for these costs, or about general aging processes. Given
that the effect is somewhat similar regardless of the relation-
ship with the G1 care recipient, new information about ones own
genetic predisposition in the aging process may play a role but not
an especially large one in the decision to purchase LTCI. Alterna-
tively, this nding could indicate that having a G1 member use
nursing home care provides the impetus to take action and to buy
LTCI. Individuals without this reminder may delay, possibly to
the point where they no longer meet the screening criteria for
coverage.
10
More specically, we separately predict two counterfactual cases where no one
versus everyone: (1) had a parent use nursing home care; (2) had an in-law use
nursing home care.
5.1. Heterogeneity
Brown et al. (2007) and Brown and Finkelstein (2008) highlight
that Medicaid, while means-tested, provides considerable insur-
ance coverage for LTC risks for those in the lower portion of the
wealth distribution, and thus the demand for LTCI may be crowded
out by public insurance. Thus, wealth may be an important source of
heterogeneity in the responses to learning about LTC. To test for this
heterogeneity, we reestimate the expectations and LTCI equations
only on the top half of the wealth distribution. In Table 5, we present
results for the more detailed breakdown of G1 LTC use.11 We nd
the impact of G1 LTC use on informal care expectations is similar
to that of the baseline. While we again nd that a parent using
a nursing home decreases an adult childs self-reported probabil-
ity of receiving an inheritance, we no longer nd the respondents
provision of care to a parent or a sibling-in-laws provision of care
to the respondents in-law signicantly impacts that probability.
Instead, we nd siblings providing care to the respondents parent
increases the respondents inheritance expectations. Individuals
may believe the exchange motive is operative, but informal care
provided by siblings might impact the relative size of the expected
inheritance between siblings, not the probability of receiving an
inheritance.12 Similar to the baseline results, nursing home care
received by the respondents parent increases the probability the
respondent purchases LTCI. We also nd the impact of an in-laws
use of nursing home care more signicantly increases the probabil-
ity of purchasing LTCI compared to the baseline estimates. Seeing
a parent (or in-law) use nursing home care may provide impor-
tant information about the costs and insurance coverage of such
care, which may be especially important for relatively wealthy
individuals who are less likely to qualify for Medicaid and more
able to afford the LTCI premiums, prompting them to purchase
LTCI. Interestingly, if a respondent provided informal care to an in-
law, he or she has a signicantly lower probability of purchasing
LTCI. Providing care for an in-law might signal that the individ-
ual places heavy weight on reciprocity and familial responsibility
norms, and thus, has a preference for informal care over formal
care, leading to a decreased probability of LTCI purchase. Those with
relatively more wealth who especially value responsibility norms
and are altruistic may be more willing to self-insure against LTC
in the hopes of leaving a large incidental bequest should LTC go
unused. Lockwood (2013) nds bequest motives, where bequests
are assumed to be luxury goods, increase savings and decrease LTCI
purchases, especially for individuals in the top half of the wealth
distribution.
We also explore heterogeneity in the effects by the age at which
the G1 generation member rst received LTC. We create an indi-
cator variable for the G1 member being 80 years old or younger
versus older than 80 years old when care was rst received.13 We
interact this indicator variable with our measures of G1 LTC use.
Results are available upon request but we provide a discussion
of our ndings for the more detailed breakdown of G1 LTC use
below.
Overall, we nd fairly consistent effects based on the age at
which a parent or in-law starts receiving care. We nd that if
a parent was over 80 years old when he or she rst received
11
The results for the broader breakdowns of G1 LTC use are available upon request.
12
It is important to note that we only observe expectations about receiving an
inheritance, not about the amount of the inheritance, and thus the identity of the
informal care provider may be less important for the extensive margin expectation.
Previous work suggests that parents often leave the exact same amounts to children
in a will, regardless of informal care provision (McGarry, 1999).
13
We selected 80 years old as the cutoff because that was the median age at which
LTC began for the G1 generation in our sample. The average age was 79.5 years old.
 N.B. Coe et al. / Journal of Health Economics 40 (2015) 122131 129

Table 5
Expectation and long-term care insurance purchase results for top half of wealth distribution.

Informal care expectations Inheritance expectations LTCI purchase

Parent used nursing home care 0.0301 **

0.0330 ***
0.0121**
(0.0118) (0.00795) (0.00486)
In-law used nursing home care 0.0166 0.00408 0.0103**
(0.0122) (0.00813) (0.00493)
R provided care to parent 0.0232 0.00674 0.00680
(0.0167) (0.0105) (0.00656)
Sibling provided care to parent 0.0127 0.0269*** 0.00239
(0.0153) (0.0101) (0.00645)
R provided care to in-law 0.0277 0.0169 0.0194***
(0.0197) (0.0124) (0.00735)
Sibling-in-law provided care to Rs in-law 0.0152 0.0121 0.00372
(0.0152) (0.00959) (0.00649)
Informal care expectations 0.00274
(0.00463)
Inheritance expectations 0.0116
(0.00704)

Observations 13,493 13,493 13,493

Notes: R denotes the HRS respondent. Marginal effects are shown and the standard errors of those marginal effects are in parentheses. Standard errors are clustered at the
individual level. Signicance levels: * p < 0.1, ** p < 0.05, *** p < 0.01.

nursing home care, the adult child is 3.2 percentage points less
likely to expect his or her own children will provide informal care
in the future, compared to no signicant effect if the parent was
young. However, the effects are not statistically different from each
other. Inheritance expectations decrease regardless of the age at
which the G1 member uses a nursing home or the age at which
sibling-in-laws provide informal care to in-laws. These estimated
effects are slightly larger for care that began when the parent
was 80 years old or younger, although the difference between the
effects is statistically insignicant. Intuitively, a parent who ini-
tiates LTC at a younger age has the potential for a longer care
episode and more expenses, decreasing the wealth available to
leave as a bequest. We nd no signicant age-related heterogene-
ity in the effect of parents using a nursing home on LTCI purchase
decisions.
There are a few cases where we nd evidence of heterogeneity
of effects based on the age of LTC onset. For example, a respon-
dent is more likely to expect an inheritance if his or her sibling
provided care to an older parent, compared to no signicant
effect if the sibling provides care to a young parent. This nd-
ing is consistent with the idea that the expected care episode
might be relatively short in duration, and thus there will be nan-
cial resources left to disperse. Regarding LTCI purchase, we nd
when a respondent provides informal care that started when the
parent was young, the respondent is 1.4 percentage points less
likely to purchase LTCI. This nding is consistent with the idea
that informal care provision to younger parents represents more
commitment to informal care and thus less demand for LTCI. It
is also consistent with a nancial drain story providing care to
younger parents, with potentially a longer disability spell, could
make individuals unable to afford LTCI. Further work eliciting indi-
vidual expectations about length of care spell and the nancial
burdens of care provision would be needed to tease out these expla-
nations.
5.2. Specication and robustness checks
To address the concern that families with G1 LTC use are
different from families without any G1 LTC use, we rerun the
regressions only on the subsample of individuals who have
someone in the G1 generation who accessed LTC at some
point. Thus, the identifying variation comes from pre- and post-
G1 LTC use among LTC-using families. The results are quite
robust, both quantitatively and qualitatively, to the baseline
estimates.14
One may have concerns about the selection of control vari-
ables. In particular, prior LTC use of the respondent may be
endogenous with respect to the expectations we consider as
well as LTCI purchase, which would bias our results. We rees-
timate all the specications excluding individuals who have
previously used LTC themselves.15 While we still nd a parents
use of nursing home care increases the individuals probability
of purchasing LTCI, the marginal effect is slightly smaller and
less precise than in the baseline. Overall, however, the results
are robust to the exclusion of individuals who used LTC them-
selves.
Finally, one might be concerned about the age criteria used
to select the sample. We have rerun all models restricting
the sample to individuals age 65 and older, a typical age at
which individuals may retire and begin LTC planning.16 Due
to a drop in sample size, some of the effects of G1 LTC use
on expectations and LTCI purchase are less precise, but are
of similar estimated size relative to the baseline results. Con-
sistent with the baseline results, we again nd a parents
nursing home use signicantly decreases expectations about
future informal care receipt and inheritance receipt, and increases
the probability of purchasing LTCI. Results are available upon
request.
14
The results are available upon request.
15
The results are available upon request.
16
One benet of this sample age restriction is that we can use the HRS question
concerning individuals assessment of their probability of entering a nursing home in
the next 5 years, which is only consistently asked across waves to individuals age 65
and over. We chose not to include this expectation in our baseline analysis because
it would have imposed a substantial sample restriction and severely limited the
age ranges for which we could analyze LTCI purchase behavior. However, when we
repeat our analysis on the older individuals, with or without the nursing home entry
expectation as an outcome and as an expectation that could affect LTCI purchase,
we nd qualitatively similar results to those of the baseline, although we lose some
signicance due to the drop in sample size.
130 N.B. Coe et al. / Journal of Health Economics 40 (2015) 122131
6. Conclusion
This paper analyzes the effects of experiencing parents and in-
laws use of LTC on an individuals expectations about informal
care and inheritance receipt as well as the probability of pur-
chasing LTCI. We nd that nursing home use has the largest and
most signicant impact on both individual expectations and LTCI
purchase decisions. We nd that individuals are more likely to pur-
chase LTCI if a close relative has used a nursing home, even after
controlling for expectations about the future. Supporting the evi-
dence of Medicaid crowd-out highlighted in Brown et al. (2007)
and Brown and Finkelstein (2008), the estimated effects of LTC
experience are stronger among the wealthier half of the sample,
who are less likely to be affected by Medicaid coverage for nursing
homes.
Our ndings are interesting in light of the standard asymmet-
ric information (AI) model and work by Finkelstein and McGarry
(2006) on LTCI. Standard AI models predict that in equilibrium
there will be a positive correlation between the amount of insur-
ance purchased and the occurrence of the risky event (Chiappori
and Salanie, 2000; Dionne et al., 2001; Chiappori et al., 2006).
Finkelstein and McGarry (2006) posit that two types of individ-
uals may lead to violations in the standard model: individuals with
private information that they are a higher risk than the insurance
company expects (for example, due to unobserved health status);
or individuals with private information that have a strong taste for
insurance and are lower risk than the insurance company expects
(i.e. risk averse individuals). Our analysis suggests that LTC use of
parents and in-laws, particularly nursing home care, makes indi-
viduals change their expectations about their own use of LTC in the
future, as well as the uptake of LTCI. This type of information has
not been explicitly considered in the AI models. Overall, the total
estimated effect of a parent using a nursing home on LTCI purchase
is about half of the previously estimated effects of potential policy
levers, such as having all states lower their Medicaid asset test to
the federal minimum (Brown et al., 2007), or the effects of state tax
deductions on LTCI holding (Goda, 2011). However, unlike these
policies, providing the information or the impetus to action that
is equivalent to having a parent or in-law use a nursing home has
no direct impact on state or federal revenues. Since the CLASS Act
has failed, policymakers could look to information campaigns as
potentially cheaper interventions if increasing LTCI coverage is a
priority.
Appendix.
Table A1
Table A1
Sample restrictions
# Person-wave
observations
All HRS respondents in waves 19982006 48,404
Age 50 and over 47,241
Has living children 44,745
Informal care and inheritance expectations questions 37,298
answered
Did not have LTCI in the prior wave 32,001
Healthy enough to pass underwriting screening as of 27,066
the prior wave
LTCI question answered 27,066
Full set of control variables available 26,979
References
AARP, 2006. The Costs of Long-Term Care: Public Perception Versus Reality in 2006.
AARP, Washington, DC.
Bacon, P.W., Gitman, L.J., Ahmad, K., Ainina, M.F., 1989. Long-term catastrophic care:
a nancial planning perspective. Journal of Risk and Insurance 56, 146154.
Bernheim, B.D., Shleifer, A., Summers, L.H., 1985. The strategic bequest motive. Jour-
nal of Political Economy 93, 10451076.
Brown, J.R., Coe, N.B., Finkelstein, A., 2007. Medicaid crowd-out of private long-term
care insurance demand: evidence from the Health and Retirement Survey. Tax
Policy and the Economy, vol. 21. MIT Press, pp. 134.
Brown, J.R., Finkelstein, A., 2007. Why is the market for long-term care insurance so
small? Journal of Public Economics 91, 19671991.
Brown, J.R., Finkelstein, A., 2008. The interaction of public and private insurance:
Medicaid and the long-term care insurance market. American Economic Review
98, 10831102.
Brown, J.R., Finkelstein, A., 2009. The private market for long-term care insurance
in the United States: a review of the evidence. Journal of Risk and Insurance 76,
529.
Brown, J.R., Goda, G.S., McGarry, K., 2012. Long-term care insurance demand limited
by beliefs about needs, concerns about insurers, and care available from family.
Health Affairs 31, 12941302.
Browne, M.J., Hoyt, R.E., 2000. The demand for ood insurance: empirical evidence.
Journal of Risk and Uncertainty 20, 291306.
Chiappori, P.-A., Jullien, B., Salani, B., Salanie, F., 2006. Asymmetric information in
insurance: general testable implications. The RAND Journal of Economics 37,
783798.
Chiappori, P.-A., Salanie, B., 2000. Testing for asymmetric information in insurance
markets. Journal of Political Economy 108, 5678.
Coe, N.B., Belbase, A., Wu, A., 2014. Overcoming barriers to life insurance coverage:
a behavioral approach. Mimeo.
Coe, N.B., Webb, A., 2009. Actual and Anticipated Inheritance Receipts, Center for
Retirement Research Working Paper 2009-32.
Courtemanche, C., He, D., 2009. Tax incentives and the decision to purchase long-
term care insurance. Journal of Public Economics 93, 296310.
Cox, D., Stark, O., 2005. On the demand for grandchildren: tied transfers and the
demonstration effect. Journal of Public Economics 89, 16651697.
Dionne, G., Gouriroux, C., Vanasse, C., 2001. Testing for evidence of adverse
selection in the automobile insurance market: a comment. Journal of Political
Economy 109, 444453.
Finkelstein, A., McGarry, K., 2006. Multiple dimensions of private information: evi-
dence from the long-term care insurance market. American Economic Review
96, 938958.
Finkelstein, A., McGarry, K., Su, A., 2005. Dynamic inefciencies in insurance mar-
kets: evidence from long-term care insurance. American Economic Review 95,
224228.
Gallagher, J., 2014. Learning about an infrequent event: evidence from ood
insurance take-up in the United States. American Economic Journal: Applied
Economics 6, 206233.
Goda, G.S., 2011. The impact of state tax subsidies for private long-term care insur-
ance on coverage and Medicaid expenditures. Journal of Public Economics 95,
744757.
Kahneman, D., Tversky, A., 1979. Prospect theory: an analysis of decision under risk.
Econometrica 47, 263291.
Kunreuther, H., 1996. Mitigating disaster losses through insurance. Journal of Risk
and Uncertainty 12, 171187.
Kunreuther, H., Ginsberg, R., Miller, L., Sagi, P., Slovic, P., Borkan, B., Katz,
N., 1978. Disaster Insurance Protection: Public Policy Lessons. Wiley,
New York.
Leventhal, H., Diefenbach, M., Leventhal, E.A., 1992. Illness cognition: using com-
mon sense to understand treatment adherence and affect cognition interactions.
Cognitive Therapy and Research 16, 143163.
Lindrooth, R.C., Hoerger, T.J., Norton, E.C., 2000. Expectations among the elderly
about nursing home entry. Health Services Research 35, 11811202.
Lockwood, L.M., 2013. Incidental Bequests: Bequest Motives and the Choice to Self-
Insure Late-Life Risks, Working Paper.
McCall, N., Mangle, S., Bauer, E., Knickman, J., 1998. Factors important in the pur-
chase of partnership long-term care insurance. Health Services Research 33,
187203.
McGarry, K., 1999. Inter-vivos transfers and intended bequests. Journal of Public
Economics 73, 321351.
Meier, V., 1999. Why the young do not buy long-term care insurance. Journal of Risk
and Uncertainty 18, 8398.
Meiners, M.R., 2001. Reections of a partnership insider on long term care nan-
cing. In: McCall, N. (Ed.), Who Will Pay for Long Term Care: Insights from the
Partnership Programs. , pp. 289299.
Norton, E.C., Van Houtven, C.H., 2006. Inter-vivos transfers and exchange. Southern
Economic Journal 73, 157172.
Palm, R.I., Hodgson, M.E., Blanchard, R.D., Lyons, D.I., 1990. Earthquake Insurance
in California: Environmental Policy and Individual Decision-Making. Westview
Press, Boulder, CO.
Papke, L.E., Wooldridge, J.M., 1996. Econometric methods for fractional response
variables with an application to 401 (K) plan participation rates. Journal of
Applied Econometrics 11, 619632.
 N.B. Coe et al. / Journal of Health Economics 40 (2015) 122131
Pauly, M.V., 1990. The rational nonpurchase of long-term-care insurance. Journal of
Political Economy 98, 153168.
Roth, D.L., Mittelman, M.S., Clay, O.J., Madan, A., Haley, W.E., 2005. Changes in
social support as mediators of the impact of a psychosocial intervention for
spouse caregivers of persons with Alzheimers disease. Psychology and Aging
20, 634644.
Sperber, N.R., Voils, C.I., Coe, N.B., Konetzka, R.T., Boles, J., Van Houtven, C.H., 2014.
How can adult children inuence parents long-term care insurance purchase
decisions? Gerontologist, PII: gnu082. PMID:25209446.
Stark, O., 1995. Altruism and Beyond: An Economic Analysis of Transfers and
Exchanges Within Families and Groups. Cambridge University Press, Cambridge,
UK.
Sunstein, C.R., Zeckhauser, R., 2011. Overreaction to fearsome risks. Environmental
and Resource Economics 48, 435449.
Tarlow, B.J., Wisniewski, S.R., Belle, S.H., Rubert, M., Ory, M.G., Gallagher-Thompson,
D., 2004. Positive aspects of caregiving contributions of the REACH project to the
development of new measures for Alzheimers caregiving. Research on Aging 26,
429453.
Taylor, D.H., Osterman, J., Will Acuff, S., stbye, T., 2005. Do seniors understand their
risk of moving to a nursing home? Health Services Research 40, 811828.
131
Taylor, S.E., Thompson, S.C., 1982. Stalking the elusive vividness effect. Psycholog-
ical Review 89, 155181.
Tversky, A., Kahneman, D., 1974. Judgment under uncertainty: heuristics and biases.
Science 185, 11241131.
Van Houtven, C.H., Norton, E.C., 2004. Informal care and health care use of older
adults. Journal of Health Economics 23, 11591180.
Viscusi, W.K., 1991. Economic theories of decision making under uncertainty: impli-
cations for policy analysis. In: Weimer, D. (Ed.), Policy Analysis and Economics.
Kluwer Academic Publishers, Boston, MA, pp. 85109.
Wiener, J.M., Tilly, J., Goldenson, S.M., 2000. Federal and state initiatives to jump start
the market for private long-term care insurance. Elder Law Journal 8, 57102.
Wolff, E.N., Gittleman, M., 2014. Inheritances and the distribution of wealth or what-
ever happened to the great inheritance boom? Journal of Economic Inequality
12, 439468.
Zarit, S.H., Reever, K.E., Bach-Peterson, J., 1980. Relatives of the impaired elderly:
correlates of feelings of burden. Gerontologist 20, 649655.
Zhou-Richter, T., Browne, M.J., Grndl, H., 2010. Dont they care? Or, are they just
unaware? Risk perception and the demand for long-term care insurance. Journal
of Risk and Insurance 77, 715747.
 Journal of Health Economics 40 (2015) 132140

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

Allergy test: Seasonal allergens and performance in school

Dave E. Marcotte
School of Public Policy, UMBC, 1000 Hilltop Circle, Baltimore, MD 21250, United States

a r t i c l e i n f o a b s t r a c t

Article history: Seasonal pollen allergies affect approximately 1 in 5 school age children. Clinical research has established
Received 18 April 2014 that these allergies result in large and consistent decrements in cognitive functioning, problem solving
Received in revised form 3 January 2015 ability and speed, focus and energy. However, compared to air pollution, the impact of pollen and seasonal
Accepted 6 January 2015
allergies on achievement in schools has received less attention from economists. Here, I use data on daily
Available online 16 January 2015
pollen counts merged with school district data to assess whether variation in the airborne pollen that
induces seasonal allergies is associated with performance on state reading and math assessments. I nd
JEL classication:
substantial and robust effects: A one standard deviation in ambient pollen levels reduces the percent of
I10
I20
3rd graders passing ELA assessments by between 0.2 and 0.3 standard deviations, and math assessments
by between about 0.3 and 0.4 standard deviations. I discuss the empirical limitations as well as policy
Keywords: implications of this reduced-form estimate of pollen levels in a community setting.
Allergies 2015 Elsevier B.V. All rights reserved.
Air quality
Achievement
Human capital

The effect of airborne allergens on the cognitive functioning of
allergy sufferers has been the subject of much attention among
medical researchers. Numerous clinical studies identify effects on
fatigue, mood, focus, the speed and accuracy of problem solving
and reasoning.1 There is also a substantial body of research on the
effects of common treatments for allergens on many of these same
outcomes.2 Indeed, the advantages of newer generation treatments
are their more limited side effects.3 However, the impacts of sea-
sonal allergies in community-based settings have not received as
much attention from economists. There is a fair amount of research
on the direct and indirect costs of allergic rhinitis in the clinical and
health services literatures. Reed et al. (2004) provide a comprehen-
sive summary of this work, and make clear the external and internal
validity limitations of this research.
Related to the topic of the current paper, economists have done
a substantial amount of research linking exposure to air pollu-
tion on child health outcomes, development and human capital
Tel.: +1 410 455 1455.
E-mail address: marcotte@umbc.edu
1
I review this evidence below.
2 human capital accumulation, but the immediate effect of allergy
See Bender (2005) for a review of the literature on both allergy symptoms and
treatments.
3
For example, so-called third-generation antihistamines such as fexofenadine
and desloratadine are marketed for their fewer side effects impeding daily func-
tioning.
accumulation. Air pollution is a combination of solids and gasses
suspended in the air that can harm human health.4 Much of the
focus is on pollution emitted as a consequence of human activ-
ity. But, natural sources like dust and pollen also contribute to
the level of ne particulate matter in the air. In general, air pol-
lution can affect cognitive performance of children in two ways:
Prolonged exposure, especially early in life, can harm develop-
ment; and exposure to high levels of pollution may have immediate
effects on health and thereby limit performance on cognitively
demanding tasks. While it is difcult to sort out these two mech-
anisms, research on pollution has identied impacts of in-utero
and neo-natal exposure on birthweight and early developmental
markers (Currie and Neidell, 2005). Pollen is unlikely to have these
effects in part because pollen allergies do not typically develop in
infancy, with age of onset usually around 810 years of age over-
all, and with a mean age of onset of 3 in one study of children
developing allergies and asthma before the age of 10 (Masuda et al.
(2008)). Research on pollen has focused on the effects of SAR on cog-
nitive functioning due principally from immediate changes in levels
of neurotransmitters that affect somnolence, mental alertness,
problem solving ability and mood. Naturally, inhibited cognitive
functioning during allergy season might have long-run effects on
4
U.S. Environmental Protection Agency: http://www.epa.gov/airquality/urbanair/.

http://dx.doi.org/10.1016/j.jhealeco.2015.01.002
0167-6296/ 2015 Elsevier B.V. All rights reserved.
 D.E. Marcotte / Journal of Health Economics 40 (2015) 132140
symptoms in clinical settings are decrements in performance on
cognitive tasks.
In this paper, I examine whether variation in the airborne pollen
that induces SAR is associated with poorer performance on reading
and math tests given to students in elementary and middle school.
I merge data on daily pollen counts at 16 reporting stations around
the country to data on test performance on mandatory state assess-
ments and student characteristics from school districts within 10
miles of these reporting stations. Students in grades 3 through 8
are tested every year in each of these districts. Math and reading
assessments are given during different weeks (and sometimes dif-
ferent months). Further, assessment dates differ by grade in many
districts. So the merged data set provides information on the per-
formance of students in a particular grade on a subject test in each
district by year, along with measures of ambient pollen levels, as
well as characteristics of the school and student body.
Using these combined data, I identify the impact of exposure to
allergens on test performance by making use of this variation in
testing dates across districts, subjects and grades, along with the
fact that the timing and volume of seasonal allergen peaks vary over
time and space. This sets up a clear natural experiment whereby
we can assess the direct effect of seasonal allergens on performance
on state tests. Further, because I know daily pollen counts, as well
as the dates on which tests were taken, I am able to estimate con-
temporaneous as well as leading and (lag effects) of pollen levels on
math and reading achievement. Later, I discuss the relevance of this
design compared to those involving student-level randomization.
Here, it is important to recognize the current design is relevant for
understanding aggregate impacts. In school settings, these aggre-
gate measures are widely used, and form the basis of accountability
systems in education.
The rst task of the paper is to describe the biological mech-
anisms through which pollen affects SAR and its subsequent
symptoms. In doing so, I summarize clinical evidence on the cog-
nitive effects of SAR. I then describe the data and the empirical
framework. I follow this with descriptive statistics and graphical
analyses of joint variation in ambient pollen and test performance,
along with the main empirical results. In the nal section, I dis-
cuss results and their limitations, and consider their value both for
understanding the effect of pollen on community level measures of
cognitive performance, as well as for understanding inter-temporal
variation in educational assessment systems.
1. Background
Seasonal allergic rhinitis (SAR) is the term clinicians use to refer
to pollen allergies, often referred to as hay fever.5 Pollen is a ne
powder, comprised of grains produced by plants as part of the
reproduction cycle (NIAID, 2012). Many trees and grasses produce
large quantities of pollen and rely on the wind to spread it from one
plant to another for reproduction. SAR is a reaction induced in the
bodies of some people to these pollen grains that are otherwise not
harmful to humans (NIAID, 2012). The allergic reaction is due to the
combination of antibodies that target allergens with receptor cells,
releasing chemicals to combat the perceived threat. These chem-
icals include histamine and cytokines that cause inammation of
tissue and increased secretion of mucus membrane (Janeway et al.,
2001). These are the common symptoms of SAR including nasal
congestion, watery eyes, and irritated throat.
These chemicals and symptoms can also affect levels of fatigue,
cognitive function, and mood. The most obvious mechanism
5
While SAR is commonly called hay fever, fever is not a symptom. However, fever
can occur due an infection in the sinuses, a common complication of SAR.
133
through which an allergic response to allergens affects cognitive
function is through effects on sleep. A very common problem
suffered by allergy sufferers is interrupted sleep and daytime som-
nolence (Santos et al., 2006). Cytokines as well as histamines
are involved in brain function, affecting cognition, and memory
(McAfoose and Baune, 2009; Tashiro et al., 2002). Additionally,
cytokines appear to affect mood, and have been linked to mood
disorders, such as major depression (Kronfol and Remick, 2000;
Dowlati et al., 2009).
While SAR is a chronic condition, the task of estimating its
prevalence in the population is difcult because many sufferers
do not seek treatment, and a conrmed diagnosis requires a skin
test (NIAID, 2012). The estimate from the National Health Interview
Survey is that 7.3% of Americans have been diagnosed by a physi-
cian with hay fever in the 12 months prior to interview.6 Meltzer
et al. (2012) estimate the prevalence of SAR in the community at
16%, while the Agency for Health Care and Quality estimates that
prevalence ranges between 10 and 30%. By all accounts, prevalence
is higher among children than adults, with some estimates as high
as 40%. There is also evidence that prevalence is rising (Linneberg
et al., 2000).
Perhaps because of its chronicity and generally mild symptoms,
SAR has not received much attention from economists studying
the consequences of disease. The exception to this is in the area of
estimating direct and indirect costs of illness. Much of this work
is analytic, deriving estimates of direct costs by estimating med-
ication and physician costs from National Health Accounts data,
and indirect costs by surveying allergy sufferers to ask how much
productivity in the workplace is inhibited by symptoms (Malone
et al., 1997; Gallup Organization Inc., 1989). A review of this litera-
ture by Reed et al. (2004) makes clear that empirical, observational
estimates of the effects of SAR on day to day functioning of sufferers
is scarce.
There is a sizeable literature in medicine on the effects of SAR
on functioning. Much of this work is based on clinical lab research,
comparing subjects with a history of SAR in various settings. For
example, Wilken et al. (2002) randomly divided subjects with SAR
into a group exposed to pollen and a control group, and found
that exposed subjects scored lower on measures of computation
and reasoning ability, and had longer response times and more
difculty with attention. Marshall et al. (2000) nd similar pat-
terns for subjects with SAR when comparing tests administered
during allergy season to those administered when pollen levels
were essentially zero. Regardless of the design for establishing
the treatmentcontrol comparison, clinical studies overwhelm-
ingly nd lower measured cognitive processing abilities and speed
among symptomatic SAR subjects (e.g., Bender, 2005; Druce, 2000;
Marshall and Colon, 1993; Fineman, 2002). It also appears that typ-
ical medical treatments do not offer much protection from fatigue
and decrements in cognitive functioning (Bender, 2005; Kay, 2000).
The only evidence of which I am aware of the effects of SAR on
school children in a community setting comes from a case-control
study of nearly 2000 British teenagers. Walker et al. (2007) com-
pare students in one region of the UK who had a history of SAR
with students with no such history as they sat for the General Cer-
ticate of Secondary Education (GCSE) exams, which are used to
determine post-secondary placement. Importantly, practice CGSE
exams are administered in winter when pollen counts are negli-
gible, and then the actual exams in June, a period of high grass
pollen in the region. The authors used a type of difference in dif-
ference analysis by comparing practice scores to nal exam scores,
6
http://www.cdc.gov/nchs/data/series/sr 10/sr10 256.pdf.
134 D.E. Marcotte / Journal of Health Economics 40 (2015) 132140
and nd that students with SAR are 40% more likely than compari-
son students to score one grade lower in one of three core subjects
of the nal than the practice CGSE, and 70% more likely to score
lower if they reported taking antihistamine treatment at the time
of the nal exam (Walker et al., 2007).
Research on other forms of air pollution on child development is
mainly thought to operate through irritating lungs, causing respi-
ratory problems, increasing risk of infection, and slowing growth
and development in early childhood (Currie and Neidell, 2005). Fur-
ther, Lavy et al. (2014) make the case that in theory pollution can
affect cognition because small particulate matter can penetrate the
lungs and inhibit the ow of oxygen into the bloodstream and hence
the brain. While the importance of this link has yet to be estab-
lished, it is clear that acute respiratory response to high levels of
pollutants can cause breathing problems and asthma attacks and
thereby inhibit performance. For example, Graff Zivin and Neidell
(2012) and Chang et al. (2014) illustrate that poor air quality lowers
productivity of piece rate daily farm workers and produce pack-
ers, respectively. While they are unable to identify mechanisms,
Lavy et al. (2014) illustrate that high levels of ne particulate pol-
lution have a negative effect on performance of Israeli high school
students on exams that determine admission to selective post-
secondary schools.
In sum, compared to pollution in general, we know less about
the effects of ambient pollen on cognitive functioning in com-
munity settings, even though this have been well established in
clinical settings, and there is some evidence that seasonal allergens
adversely affect performance on high school end-of-course exams.
2. Empirical framework
To examine the relationship between seasonal allergens and
cognitive problem solving on state assessments of math and read-
ing achievement, I compiled a data set from sources of two types.
First, I obtained daily pollen counts from January 2003 through
October 2012 from the National Allergy Bureau (NAB), of the Amer-
ican Academy of Allergy, Asthma and Immunology. The NAB data
provide daily records of total (and constituent) pollen counts at
recording stations around the country.7 The NAB stations report the
accumulated number of pollen grains per cubic meter (grains/m3 )
of air in a 24-h period. Pollen counts of this type are the basis of
reports in weather reports on television news broadcasts, newspa-
pers and websites.
The second sources were state and school district accountability
data systems, websites and staff. The school systems included here
are those for which some portion of the district was within 10 miles
of the site at which pollen levels were measured. So, for each NAB
reporting station, I merge data for at least one school district. Many
allergen reporting stations are located within geographically large
school districts, and only data from one district is relevant. In other
cases, reporting stations are matched to two or three proximate
districts. One reporting station (Cook County, IL) lies in an area
with very small (geographically) districts and is matched to seven
school districts. In total, I use data from 48 school districts, so that
the average reporting station is matched to 3 school districts.
From each school district, I obtained data on the percent of
students who scored at the level of procient or advanced on
state math and English language arts (ELA) assessments, as well
as student characteristics such as race and free/reduced price
meal (FARM) eligibility, and teacher to student ratio. While dis-
tricts across the country report the rates at which students are
7
Some but not all stations report during the winter, with stations in more tem-
perate zones more likely to report 365 days per year.
procient (or advanced) on state assessments as part of the repor-
ting requirements under the NCLB, state assessments differ on their
rigor. Indeed, these differences and critiques of low standards (e.g.
Peterson and Hess, 2008) have been an impetus for the recent push
(and push-back) for states to adopt standards based on the Com-
mon Core. Because the rigor of tests varies by state during this
period, all models include xed effects to capture across-state mean
differences in prociency rates.
For each district in the sample, I also obtained information on
the dates state assessments were administered. Using these dates,
I calculate the mean level of pollen in the atmosphere during the
testing period for each grade/subject in a district. I compile these
data into a district/grade/subject/year panel measuring test per-
formance, school characteristics and ambient levels of pollen, with
records for all students, and for students by race and FARM status.
Using these data, I estimate a series of models of the following
type
Pgdt = a + b1 Pollengdt + b2 Xdt + g + f (d, t) + gdt
where Pgdt measures the proportion of students scoring procient
or advanced on their state assessment in grade g, in district d dur-
ing year t.8 Pollengdt is a measure of pollen in the atmosphere at the
time the test was administered. I discuss the measure of pollen
in more detail below. Xdt measures the year-specic percent of
students in the district eligible for free/reduced price meals; the
percent of student who are black; the percent of students who
are Hispanic; the pupilteacher ratio; and per-pupil expenditures
on instruction. g is a grade xed effect and f(d,t) represents var-
ious parameterizations of district/time effects. I estimate models
that include district and year xed effects; district xed effects and
common time trends, and; district xed effects and district-specic
linear trends. All of these models are estimated separately for math
and ELA assessments. In all models, standard errors are clustered
at the district-grade level.
Insight into how to model the relationship between pollen and
test outcomes can be obtained from the medical and epidemiologic
literature on the relationship between levels of atmospheric pollen
and symptomatic response. That literature is robust and commonly
nds that SAR symptoms or complications rise quickly with pollen
count, and then plateau. This type of doseresponse pattern is
common in studies regardless of how outcomes are measured,
including: (1) Doctors visits for symptoms of SAR; (2) Emergency
clinic visits for complications like asthma, or; (3) Sales or consump-
tion of drugs used for treating SAR (Caillaud et al., 2014a). In studies
using these measures, pollen counts of less than 100 grains/m3 and
as low as 30 grains/m3 are associated with response, and higher
pollen counts do not appear to result in proportionally higher lev-
els of symptomotology (e.g. Erbas et al., 2007; Johnston et al., 2009;
Caillaud et al., 2014b). This doseresponse pattern is reected in
a common metric for reporting pollen levels: the Padgett Pollen
Indicator (PPI). The PPI distinguishes between days with low, mod-
erate, high and extreme levels of pollen in the air. The PPI denes
days with pollen counts in excess of 15 grains/m3 as moderate;
counts in excess of 90 grains/m3 as high; and pollen counts above
1500 grains/m3 as extreme.
Because of this dose response pattern, I include measures of
pollen in log form. On its own, the log transformation is consistent
with the clinical doseresponse pattern, since it is more sensitive to
variation at the bottom of the distribution. Fig. 1 further makes clear
8
In three of the states in which sample districts were located, assessment scores
were changed or re-scaled the study period: South Carolina, Tennessee, and New
York. In these cases I include indicators to pick up mean shifts associated with the
re-scaling.
 D.E. Marcotte / Journal of Health Economics 40 (2015) 132140 135

Table 2
Descriptive statistics of analytic sample.

Variable Mean Std. dev. Minimum Maximum

Math Prof. rate 68.13 21.62 2 99

ELA Prof. rate 70.54 22.39 14.1 99.4
Pollen Ct. during test 467.7 635.3 0 3494
Students FARM eligible (%) 45.55 25.19 0 97.7
Af. Amer. students (%) 20.29 20.50 0 83.7
Hispanic students (%) 19.85 18.47 1 68.7
Pupil/teacher ratio 14.39 2.41 5.9 21.7

n = 1608. Observations are at the student group/grade level within districts.

Fig. 1. Distribution of district/year obs. by pollen during tests.

Table 1
Locations of NAB reporting stations and pollen during state tests.

Site Mean Std. dev. Minimum Maximum

Albany, NY 409.3 818.6 0 2046.6

Atlanta, GA 781.5 512.8 282.4 1307.0
Austin, TX 508.1 326.0 202.7 809.1
Baltimore, MD 69.6 58.5 8.7 166.7
Chicago, IL 8.9 25.4 0 71.7
Colorado Springs, CO 235.7 252.4 0.7 654.3
Dallas, TX 251.5 66.2 183 322.3
Fig. 2. Scatterplot and t of prociency rates and pollen count.
Dayton, OH 337.4 195.0 61.5 599.7
Erie, PA 201.9 10.0 190.8 210.4
Greenville, SC 105.8 44.4 64.4 214.0
3. Results
Kansas City, MO 784.1 979.9 80.9 2860.7
Knoxville, TN 475.5 390.4 233.7 926.0
Rochester, NY 141.0 258.3 0 640.8 In order to provide context for the analyses to follow, rst con-
San Jose, CA 349.9 423.9 103.9 1383.1 sider descriptive statistics for the districts under study here. In
Seattle, WA 127.4 54.5 84.2 207.7 Table 2, I present summary statistics of prociency rates and demo-
Waco, TX 1140.7 700.9 226.6 2097.7
graphic and economic characteristics of the analytic sample. Two
features of the sample are worth noting. First, the districts included
are generally urban or suburban. This is reected in the proportion
of students of color, as well as rates of FARM eligibility. Second,
the advantage of the log transformation. The gure displays the there is substantial heterogeneity within the study sample, with
distribution of pollen counts during testing periods, which is highly wide variation on prociency rates as well as student characteris-
skewed.9 In addition to continuous measures of pollen count, I also tics.
estimate models using the PPI to measure variation in pollen levels. As an initial way to assess the correlation between levels of
To begin to understand how changes in the levels of pollen in pollen in the atmosphere and prociency rates on state assess-
the atmosphere might affect performance on state assessments, ments, consider the scatterplot and t of the relationship between
in Table 1, I present descriptive statistics on pollen levels at the the two in Fig. 2. Each data point in the plot represents a com-
weather reporting stations for which NAB data are available dur- bination of the prociency rate and pollen levels for students in
ing test dates at proximate school districts. In addition to the a district/year for 3rd graders in math, net of district (and group)
mean level of pollen during test dates observed over the panel, xed effects, and district trends. Both visual analysis and the linear
this includes the standard deviation and extreme values. It is clear trend indicate that years when students score below expectations
that mean levels of pollen vary from site to site. More interest- for a district, given the districts trend (i.e. below 0 on the y-axis),
ing, though, is the within-site variation. Several sites administer tend to be years when students in those district are exposed to
tests when the pollen level is very low (<15) in some years, and higher levels of pollen than typical in the district (i.e. above 0 on
extremely high in others. It is this sort of within district variation the x-axis).10
in pollen levels that I exploit in the models below. However, a lim- An alternative way to assess the joint variation in atmospheric
itation of the data used here is that in some areas, pollen levels are pollen and test performance is to plot mean prociency rates and
persistently high during testing periods. This limits the power of allergen levels within a district over time, net of district means and
these data to test whether the impact of additional pollen on cog- trends. Fig. 3 displays time series of mean prociency rates of 3rd
nitive functioning plateaus beyond 100 grains/m3 , as suggested in graders in math in all school districts (measured on the left axis)
the clinical literature. and pollen levels (measured on the right axis). The two time series
do appear to move in opposite directions from year to year. For
example, the two years with the highest test scores over district
9
The empirical results are not sensitive to the decision to measure pollen counts
in log form. Nonetheless, goodness-of-t tests conrm the superiority of the use of
logs. 10
The coefcient on the linear t is 1.81, with a standard error of 0.406.
 136
Table 3
Estimates of ambient pollen and state assessment results.

Variable Model 1 Model 2 Model 3

ELA Math ELA Math ELA Math

Coef. SE Sign Coef. SE Sign Coef. SE Sign Coef. SE Sign Coef. SE Sign Coef. SE Sign

ln(pollen 1.242** 1.792*** 1.138*** 1.76*** 1.308*** 2.23***

count) 0.546 0.683 0.532 0.672 0.397 0.613
0.303 0.31 0.289

D.E. Marcotte / Journal of Health Economics 40 (2015) 132140

ln(pollen p = 0.005 0.445 p = 0.019 p = 0.006 0.437 p = 0.017 p = 0.001 0.451 p = 0.004
count) 4th 0.297 0.604 0.294 0.599 0.282 0.596
grade
ln(pollen 0.422 p = 0.002 1.059 p = 0.134 0.471 p = 0.002 1.014 p = 0.120 0.398 p < 0.001 1.102** p = 0.006
count) 5th 0.35 0.565 0.353 0.562 0.321 0.553
grade
ln(pollen 0.342 p = 0.094 1.22** p = 0.297 0.334 p = 0.124 1.204** p = 0.307 0.35 p = 0.013 1.222** p = 0.031
Count) 6th 0.297 0.569 0.298 0.569 0.264 0.556
grade
ln(pollen 0.163 p = 0.041 1.483** p = 0.594 0.154 p = 0.056 1.468** p = 0.615 0.163 p = 0.003 1.496** p = 0.144
count) 7th 0.298 0.614 0.305 0.618 0.287 0.592
grade
ln(pollen 1.623*** p = 0.462 1.803*** p = 0.983 1.573*** p = 0.390 1.755*** p = 0.993 1.646*** p = 0.390 1.866*** p = 0.475
count) 8th 0.399 0.676 0.391 0.672 0.401 0.672
grade
District xed Yes Yes Yes Yes Yes Yes
effects?
Year xed Yes Yes No No No No
effects?
Common No No Yes Yes No No
trend?
District- No No No No Yes Yes
specic
trends?
n 1608 1576 1608 1576 1608 1576
R2 0.915 0.862 0.912 0.858 0.938 0.892

All models control for student demographic and economic characteristics, as described in the text, enrollment and pupilteacher ratios. Standard errors (below, in italics) clustered at the district-grade level.
**
p < 0.05.
***
p < 0.01.
p-Value from F-test of joint signicance with main effect.
 D.E. Marcotte / Journal of Health Economics 40 (2015) 132140
Fig. 3. Time series of pollen and 3rd grade math scores.
averages and trends (2009 and 2011) are also the years with the
lowest levels of atmospheric pollen. It is also clear, however, that
any inverse relationship between the two time series is certainly
not perfect.
3.1. Regression results
To better assess whether any relationship exists between ambi-
ent pollen levels and performance on state assessments, I turn
next to the results of the multivariate models discussed above. In
Table 3 I present the results from three separate sets of regressions,
each using different strategies to control for underlying changes in
test scores over time. All models are estimated in linear-log form,
because of the highly skewed distribution of ambient pollen lev-
els during testing periods. In Model 1 (columns 1 and 2), I present
results for ELA and math, respectively, in models with common
year xed effects. In columns 3 and 4, rather than time shifters,
I include a common linear trend, and in columns 5 and 6 are the
results from models using district specic linear trends. The rst
two models are robust to unmeasured changes in cohort quality or
environmental factors that affect all students, either as year specic
effects or in linear form. The last model is robust to unmeasured,
linear changes in cohort quality or environment that are specic to
each school district. In all models, the rst coefcient measures the
effect of pollen on the rate at which students in a district pass their
states ELA/math assessment, and the only coefcient of interest for
3rd graders. Below this are interaction terms for additive effects for
other grades.
Regardless of how underlying trends are modeled, higher levels
of pollen during test windows have a negative effect on the percent
of students passing state assessments. These are largest for younger
students, since the coefcients on the direct (3rd grade) effects are
uniformly negative, and the additive effects for 4th and 5th grades
are typically negative and for 8th graders the overall effect is always
indistinguishable from zero. Across all models the effects for stu-
dents in grades 4 through 7 are statistically indistinguishable from
those of 3rd graders in ELA. For math, however, it appears the rela-
tionship between pollen and performance is strongest for students
in lower grades. Overall, while the negative effects of pollen on test
performance are concentrated on students in the lowest grades,
effects are qualitatively similar for both subjects and across grades.
To make sense of the magnitude of the estimates, a 100%
increase in pollen levels is associated with a decline in 3rd graders
passing state assessments by between 1.1 and 1.3% for ELA and
between 1.8 and 2.2% for math. To help interpret these esti-
mates, recall that 100% increase in pollen levels (at the mean) is
smaller than one standard deviation. Scaling this up slightly, a one
137
standard deviation in ambient pollen levels reduces the percent
of 3rd graders passing ELA assessments by between 0.2 and 0.3
standard deviations, and math assessments by between about 0.3
and 0.4 standard deviations. These are sizeable effects compared
to other changes in education settings often discussed, such as
changes in class size or teacher quality. However, they are a bit
smaller than effect sizes on memory, problem solving speed and
reasoning measured in clinical settings, which range from 0.30 to
0.65 (Wilken et al., 2002).
In Table 4, I present results of the Model 1 estimated for two
important subgroups of students: black students and economically
disadvantaged students, as proxied by FARM eligibility. I present
results for these two subgroups because they are important in edu-
cational accountability, and because disadvantaged groups are less
likely to receive diagnosis and treatment for many health condi-
tions, so could represent groups for which policy levers to improve
access to care may be especially benecial. For both groups, the
patterns observed in Table 3 for all students generally persist. The
deleterious effects of ambient pollen are seen especially in math,
and they seem to be less severe in 8th grade. Importantly, how-
ever, the point estimates for these subgroups in math are larger (in
absolute value) than those observed for all students in Table 3.
3.2. Specication checks
While the results suggest high levels of pollen reduce perfor-
mance on primary school math and reading exams, they are surely
open to questions about functional form and model specication.
The results in Tables 3 and 4 are from models that restrict the effect
of ambient pollen on achievement to be linear in logs. Recall, that
research in clinical settings has often found the effects of pollen
levels on allergy symptoms, physician visits or drug sales to rise
quickly and plateau, and that pollen levels of 30100 grains/m3
are typically associated with substantial symptom response. To
allow for this type of relationship, I estimate models in which the
pollen levels are measured using the PPI index of severity. In addi-
tion to being widely used, employing the PPI allows for a test of
non-linearity in models of test performance.
In the rst panel of Table 5 I present results from models in
which pollen levels are measured as a series of indicators of the
PPI index, where the omitted category is the lowest level of pollen.
Because indicator models with multiple measures of pollen inter-
acted with six grades would have too many coefcients of interest,
I estimate this model restricting the effects of pollen to be com-
mon across grade, and including grade xed effects.11 In order to
conserve space, I present estimates for Model 3, which includes
district-specic linear trends. This is the most exible and robust
specication, providing the clearest prepost comparison. In any
case, as in the previous tables, results are highly similar across
specications. For both the ELA and math exams, the rate at which
students pass assessments is lower at any ambient pollen level
other than the lowest level (<15/m3 ). I estimate that when tested
on days with pollen in the moderate rather than low range, about
3.5% fewer students pass ELA exams and 4.9% fewer pass math
exams. Notably, for neither subject do higher levels of pollen (high
or extreme) generate lower levels of achievement. Recall, however,
that at several NAB pollen monitoring sites pollen levels during
testing were persistently above 90 m3 . As a result the within-site
differences in student performance between years with the lowest
11
To the extent that the weak evidence from Tables 3 and 4 are consistent with
different effects by grade, this across-grade average may understate effects in lower
grades and overstate effects in higher grades.
138 D.E. Marcotte / Journal of Health Economics 40 (2015) 132140

Table 4
Estimates of ambient pollen and state assessment results: by subgroup.

Variable Black students FARM eligible students

ELA Math ELA Math

Coef. SE Sign Coef. SE Sign Coef. SE Sign Coef. SE Sign

Pollen count 0.959 0.729 2.615 0.847

***
1.066 0.694 2.597 0.838
***

Pollen count 4th grade 0.529 0.358 p = 0.044 0.777 0.679 p = 0.011 0.695** 0.329 p = 0.01 0.812 0.699 p < 0.01
Pollen count 5th grade 0.681* 0.413 p = 0.020 1.983*** 0.641 p = 0.319 0.769 0.406 p = 0.006 1.859*** 0.656 p = 0.212
Pollen count 6th grade 0.379 0.33 p = 0.423 1.463** 0.627 p = 0.091 0.216 0.305 p = 0.209 1.266* 0.657 p = 0.041
Pollen count 7th grade 0.197 0.331 p = 0.288 1.97*** 0.687 p = 0.385 0.016 0.357 p = 0.133 1.592** 0.701 p = 0.148
Pollen count 8th grade 1.954*** 0.461 p = 0.155 2.128*** 0.714 p = 0.464 1.575*** 0.448 p = 0.445 1.863*** 0.76 p = 0.260
District xed effects? Yes Yes Yes Yes
Year xed effects? Yes Yes Yes Yes
Common trend? No No No No
District-specic trends? No No No No
n 1204 1171 1309 1282
R2 0.8727 0.828 0.893 0.854

All models control for student demographic and economic characteristics, as described in the text, enrollment and pupilteacher ratios. Standard errors (below, in italics)
clustered at the district-grade level.
*
p < 0.10
**
p < 0.05
***
p < 0.01
p-Value from F-test of joint signicance with main effect.

Table 5
Estimates of ambient pollen and state assessment results: specication checks.

Variable Effects for all grades Model 3 w/controls for pollution

ELA Math ELA Math

Coef. SE Coef. SE Coef. SE Sign. Coef. SE Sign.

PPI pollen level: low is reference

Moderate 3.471** 1.515 4.911*** 1.585
High 2.586** 1.075 2.686** 1.245
Extreme 2.937** 1.29 2.428* 1.46
Pollen count 1.193*** 0.362 1.61*** 0.547
Pollen count 4th grade 0.288 0.287 p < 0.001 0.451 0.575 p < 0.001
Pollen count 5th grade 0.406 0.324 p < 0.001 1.047 0.535 p = 0.078
Pollen count 6th grade 0.352 0.267 p = 0.01 1.058 0.543 p = 0.147
Pollen count 7th grade 0.165 0.286 p = 0.002 1.332** 0.579 p = 0.50
Pollen count 8th grade 1.639*** 0.408 p = 0.20 1.659** 0.657 p = 0.91
Sulfur dioxide (aqi) 0.251*** 0.081 0.035 0.084
Grade xed effects? Yes Yes Yes Yes
District xed effects? Yes Yes Yes Yes
Common trend? Yes Yes No No
District-specic trends? Yes Yes Yes Yes
n 1537 1506 1608 1576
R2 0.9374 0.857 0.938 0.894

All models control for student demographic and economic characteristics, as described in the text, enrollment and pupilteacher ratios. Standard errors (below, in italics)
clustered at the district-grade level.
*
p < 0.10.
**
p < 0.05.
***
p < 0.01.

and moderate levels of pollen are being driven by a subset of all
sites.
An alternative threat to validity for the ndings presented in
Tables 3 and 4 is that high levels of ambient pollen could occur
during periods of relatively high levels of other forms of air pollu-
tion, and hence be picking up the effects of low air quality more
generally. To provide some evidence of the impact of pollutants
separate from airborne pollen, I make use of data collected by the
U.S. Environmental Protection Agency (EPA) at air pollution mon-
itoring sites throughout the country. The EPA collects data on a
variety of pollutants. A frequently used measure is airborne par-
ticulate matter (APM), but this includes organic particles including
dust and pollen. So, I use data on sulfur dioxide (SO2 ), which is
generated by fossil fuel combustion in power plants, industrial
production and automobiles and is a commonly used measure of
air quality.12 Unfortunately, the EPA monitoring sites do not over-
lap with the NAB reporting stations well. In some cases, there are
no EPA sites near NAB sites, and in others EPA monitoring data is
available for only a few years. In total, I am able to obtain EPA data
for about 43% of site/years for which pollen data are available.
In the second panel of Table 5, I present results of Model 3
from Table 3, but include the SO2 air quality index (AQI) as a
control, and an indicator variable for whether or not SO2 data
12
The combustion processes that emit SO2 also emit constituents of APM, so these
measures or air quality are clearly related. The correlation between SO2 levels and
measures of ne APM for the EPA monitoring sites during this period is 0.2051.
 D.E. Marcotte / Journal of Health Economics 40 (2015) 132140
are missing for a site/year. So the effect of pollution is estimated
off of within site variation in pollution for the years when these
data are available. The results in Table 5 indicate that SO2 pol-
lution levels do affect performance on state reading tests, but
not math tests. To better understand the relative effects here,
the standard deviation of SO2 AQI is about 9, so coefcient of
SO2 on reading scales to an effect size of 0.108, which is less
than half of the effect size of pollen on performance on read-
ing assessments. Importantly, when I include measures of air
pollution, the effects of pollen levels on both math and read-
ing tests are essentially identical to the estimates obtained in
Table 3.
4. Discussion and conclusions
In this paper, I have taken up a straightforward, but important
question on the effect of ambient pollen on cognitive functioning
in the community. Researchers in clinical settings have long
understood that the symptoms of seasonal allergic rhinitis affect
concentration, problem solving and even mood in ways that
impede cognitive functioning. While the evidence from clinical
settings is reason to anticipate that natural variations in ambient
pollen could affect measures of cognitive performance like state
reading and math assessments, there is no way to predict popula-
tion level impacts without observational data of the type employed
here. Of course, the empirical strategy employed here falls short
of the randomized control trial design that might be employed by
clinical researchers, and the inferences we can draw are subse-
quently not as strong. The panel models estimated here control for
typical measures of student characteristics and school resources,
and all inference is drawn from within-district variation and is
robust to unmeasured changes in cohort quality or environmental
changes that happen to be coincident with atmospheric pollen.
The threats to inference that remain are due to any other fac-
tors that happen to vary with seasonal pollen and are not measured
here. The research design employed here largely rests on the
assumption that annual variation in pollen levels during testing
periods is as good as random, driven by natural variation that is
unrelated to schooling policies. There is no way to put to rest all
potential threats to this assumption. But, it is possible to set up a test
to assess whether the variation driving the results is due to pollen
per se, or something that just happens to be coincidental. This fal-
sication test is to re-estimate models from Table 3, while also
including measures of ambient pollen in weeks before and/or after
state tests are administered. Pollen levels in the weeks before tests
are administered could affect test performance if symptoms lag
pollen levels, but pollen levels after tests are administered should
have no affect on outcomes. But, to the extent that inter-temporal
variation in pollen is simply correlated with other unmeasured fac-
tors that inuence test scores, then pollen levels after testing should
also be correlated with these omitted factors, leading us to believe
in an after-test pollen effect.
In Fig. 4, I present the main coefcients of interest from regres-
sions testing whether there is a before- and/or after-test pollen
effect. These coefcient are estimated in separate regression which
include controls and specication as in Table 3, including a con-
temporaneous measure of pollen during the test week(s), as well
as pollen in the weeks leading up to the test period, and then after
the test period. It is important to control for pollen during the test
period, since there is substantial serial correlation in pollen levels.
The bars in Fig. 4 indicate the magnitude of the coefcient from
these regressions on ELA and math prociency rates: Bars in regu-
lar shade indicate statistical signicance at the 5% level, while bars
139
Fig. 4. Prociency effects of pollen in weeks before, during and after test period.
in light shade indicate no signicant relationship between pollen
in a given week and test scores.
For both ELA and math exams, the same pattern emerges
between pollen levels leading into, and out of the testing period
and student performance. Pollen levels in the weeks leading up to
the test window are negatively related to prociency rates. But, the
effect of pollen levels in advance of the test is not as large as the
effects of pollen during the period when tests are administered.
Further, in no case are pollen levels after the test window related
to test performance.
While this falsication test helps provide assurance that
the relationship between ambient pollen and test performance
observed here is the result of the effects of pollen and thereby
SAR on the ability of students to perform on state assessments,
there are important limitations of the current study. Most notably,
the reduced form approach of the current design cannot tell us
about the impact of SAR on the test performance of those affected.
Unfortunately, there is simply no community level data available
to measure test performance among individuals suffering from SAR
compared to those who do not. Diagnosis of SAR requires a skin-
prick test that is not possible in any reasonable survey design. In any
case, the reduced form strategy employed here is informative about
an average treatment effect of atmospheric pollen on test perfor-
mance. Necessarily, this ATE is likely a poor estimate of effects on
individuals. Most students do not suffer from SAR, and elevated lev-
els of ambient pollen likely have little or no effect for them. Indeed,
the ATE estimated here is likely comprised of a treatment effect for
SAR sufferers, but muted by non-response among those without
allergies.
Another limitation of the current paper is that I can not dis-
tinguish between effects of pollen levels on how well students do
on state achievement tests versus effects on learning and cogni-
tive performance more generally. Identifying effects on learning
and cognitive development would require a longer panel. How-
ever, the negative effects of pollen levels in the weeks prior to
state tests in Fig. 4 suggest that SAR symptoms inhibit learning and
not just contemporaneous performance on math and reading tests.
Further, since the performance on tests can be used in student abil-
ity tracking or for school accountability tracking, even if the effect
of SAR on cognition was mainly transitory, poor test performance
could affect subsequent schooling and human capital development.
Indeed, Lavy et al. (2014) illustrate that ambient pollution levels
can affect performance on high stakes tests in Israel, and argue
that because of this pollution can limit schooling and labor market
outcomes. These two pathways are clearly important, and more
140 D.E. Marcotte / Journal of Health Economics 40 (2015) 132140
work remains to be done to distinguish between transitory and
permanent effects of pollen and pollution in general.
The results of the current paper are relevant for economists
studying both health and education. Health economists have made
real contributions to understanding the role that factors like pol-
lution have on child health and development. This paper examines
the immediate effects of naturally occurring particulate matter on
cognitive functioning. While this relationship has been established
in clinical research, the current ndings suggest that allergens limit
functioning in community settings. This has broader implications
than school performance. For example, productivity in the work-
place may be limited by exposure to pollen among SAR sufferers
engaging in cognitively demanding work.
For economists who study education, the average treatment
effects identied here are relevant for understanding whether and
how year-to-year changes in test results can be used to measure
school or district performance. I estimate that commonly observed
variation in ambient pollen levels serve as an important imped-
iment to improvement on state math and reading assessments.
The point estimates reported here suggest that prociency rates
for 3rd graders on ELA exams could be 2.54.5% lower if testing
occurs during a period of extremely high pollen, compared to typ-
ical pollen levels. For math, I estimate the prociency rates to be
4.57% lower. Effects of this type have direct relevance to under-
standing the learning or performance effects of a condition that
effects a large number of students annually, even as it can be treated
at low cost. Further, better understanding the role of extraneous
factors like air quality in inuencing variation in performance on
state assessments is important as administrators continue to rene
accountability systems.
Acknowledgements
I am grateful for the help of many people at a variety of organiza-
tions for help in acquiring data for this project. These include: the
National Allergens Bureau and the American Academy of Allergy
Asthma and Immunology, and in particular Jerome Schultz; Melissa
Ogden of the Knox County (TN) Public Schools; Carolyn Hayward
and Janeen Demi-Smith of the Colorado Springs School District 11;
Liz Jones from the South Carolina Department of Education; and
Liliane Patel from the San Jose (CA) Unied School District. I ben-
eted from the able research assistance of Aynur Saygin, and from
the comments and suggestions of Peter Hinrichs, John Cawley, Dave
Frisvold, Bryant Kim and seminar participants at Cornell, the 2014
AEFA meetings in San Antonio, TX, and the 2015 APPAM research
conference in Albuquerque, NM.
References
Bender, B.G., 2005. Cognitive effects of allergic rhinitis and its treatment. Immunol-
ogy and Allergy Clinics of North America 25 (May (2)), 301312.
Caillaud, D., Toloba, Y., Raobison, R., Besancenot, J.P., Thibaudon, M., Martin, S.,
Segala, C., 2014a. Health impact of exposure to pollens: a review of epi-
demiological studies. Revue des Maladies Respiratoires 31 (February (2)),
142149.
Caillaud, D., Martin, S., Segala, C., Besancenot, J.P., Clot, B., Thibaudon, M., 2014b.
Effects of airborne birch pollen levels on clinical symptoms of seasonal allergic
rhinoconjunctivitis. International Archives of Allergy and Immunology 163 (4),
4350.
Chang, T., Graff Zivin, J.S., Gross, T., Neidell, M.J., 2014. Particulate Pollution and the
Productivity of Pear Packers, NBER working paper 19944.
Currie, J., Neidell, M., 2005. Air pollution and infant health: what can we learn
from Californias recent experience. Quarterly Journal of Economics 120,
10031030.
Dowlati, Y., Hermann, N., Swardfage, W., Liu, H., Sham, L., Reim, E.K., Lanctot, K.L.,
2009. A meta-analysis of cytokines in major depression. Biological Psychiatry
67 (5), 446457.
Druce, H.M., 2000. Seasonal allergic rhinitis and cognitive function. Annals of Allergy,
Asthma, and Immunology 84 (April (4)), 371373.
Erbas, B., Chang, J.H., Dharmage, S., Ong, E.K., Hyndman, R., Newbigin, E., Abramson,
M., 2007. Do levels of airborne grass pollen inuence asthma hospital admis-
sions? Clinical and Experimental Allergy 37, 16411647.
Fineman, S.M., 2002. The burden of allergic rhinitis: beyond dollars and cents. Annals
of Allergy, Asthma, and Immunology 88 (April (4 Suppl.)), 27.
Gallup Organization Inc., 1989, September. Survey of Attitudes and Experiences of
Allergy Sufferers. The Gallup Organization, Inc., Washington, DC.
Graff Zivin, J.S., Neidell, M.J., 2012. The impact of pollution on worker productivity.
American Economic Review 102 (7), 36523673.
Janeway, C.A., Travers, P., Walport, M., 2001. Immunobiology: The Immune System
in Health and Disease, 5th ed. Garland Science, New York.
Johnston, F.H., Hanigan, I.C., Bowman, D.M.J.S., 2009. Pollen loads and allergic rhinitis
in Darwin, Australia: a potential health outcome of the grass-re cycle. Eco-
Health 6, 99108.
Kay, G.G., 2000. The effects of antihistamines on cognition and performance. Journal
of Allergy and Clinical Immunology 105 (6), S622S627.
Kronfol, Z.D., Remick, D.G., 2000. Cytokines and the brain: Implications for clinical
psychiatry. American Journal of Psychiatry 157, 683694.
Lavy, V., Ebenstein, A., Roth, S., 2014. The Long Run Human Capital and Economic
Consequences of High-stakes Examinations, NBER working paper 20647.
Linneberg, A., Jorgensen, T., Nielsen, N.H., Madsen, F., Frolund, L., Dirksen, A., 2000.
The prevalence of skin-test-positive allergic rhinitis in Danish adults: two cross-
sectional surveys 8 years apart. The Copenhagen allergy study. Allergy 55,
767772.
Malone, D.C., Lawson, K.A., Smith, D.H., 1997. A cost of illness study of allergic rhinitis
in the United States. Journal of Allergy and Clinical Immunology 99, 2227.
Marshall, P.S., Colon, E.A., 1993. Effects of allergy season on mood and cognitive
function. Annals of Allergy 71 (3), 251258.
Marshall, P.S., OHara, C., Steinberg, P., 2000. Effects of seasonal allergic rhinitis on
selected cognitive abilities. Annals of Allergy, Asthma, and Immunology 84 (4),
403410.
Masuda, S., Fujisawa, T., Katsumata, H., Atsuta, J., Iguchi, K., 2008. High prevalence
and young onset of allergic rhinitis in children with bronchial asthma. Pediatric
Allergy and Immunology 19 (6), 517522.
McAfoose, J., Baune, B.T., 2009. Evidence for a cytokine model of cognitive function.
Neuroscience and Biobehavioral Reviews 33 (3), 355366.
Meltzer, E.O., Blaiss, M.S., Naclerio, R.M., Stoloff, S.W., Derebery, J.M., Nelson, H.S.,
Boyle, J.M., Wingertzahn, M.A., 2012. Burden of allergic rhinitis: allergies in
America, Latin America and Asia-Pacic adult surveys. Allergy and Asthma Proc-
cedings 33, s113s141.
National Institute of Allergy and Infectious Diseases (NIAID), 2012. Pollen Allergy.
http://www.niaid.nih.gov/topics/allergicDiseases/understanding/pollenallergy/
Documents/pllenAllergy.pdf (Retrieved 31.12.14).
Peterson, P., Hess, F., 2008. Few states set world class standards. Education Next 8
(3), 7073.
Reed, S.D., Lee, T.A., McCrory, D.C., 2004. The economic burden of allergic rhinitis: a
critical evaluation of the literature. Pharmacoeconomics 22 (6), 345361.
Santos, C.B., Pratt, E.L., Hanks, C., McCann, J., Craig, T.J., 2006. Allergic rhinitis and
its effect on sleep, fatigue, and daytime somnolence. Annals of Allergy, Asthma,
and Immunology 97 (November (5)), 579587.
Tashiro, M., Mochizuki, H., Iwabuchi, K., Sakurada, Y., Itoh, M., Watanabe, T., Yanai,
K., 2002. Roles of histamine in regulation of arousal and cognition. Life Sciences
72, 409414.
Walker, S., Khan-Wasti, S., Fletcher, M., Cullinan, P., Harris, J., Sheikh, A., 2007.
Seasonal allergic rhinitis is associated with detrimental effect on examina-
tion performance in United Kingdom teenagers: case-control study. American
Academy of Allergy, Asthma and Immunology 3, 381387.
Wilken, J.A., Berkowitz, R., Kane, R., 2002. Decrements in vigilance and cogni-
tive functioning associated with rag-weed induced allergic rhinitis. Annals of
Allergy, Asthma, and Immunology 89 (4), 372380.
 Journal of Health Economics 40 (2015) 141158

Contents lists available at ScienceDirect

Journal of Health Economics

journal homepage: www.elsevier.com/locate/econbase

The role of marriage in the causal pathway from economic conditions

early in life to mortality
Gerard J. van den Berg a,b,c,d , Sumedha Gupta e,
a
University of Mannheim, Mannheim, Germany
b
Institute for Labor Market Policy Evaluation (IFAU), Uppsala, Sweden
c
VU University Amsterdam, Amsterdam, The Netherlands
d
Institute for the Study of Labor (IZA), Bonn, Germany
e
Indiana University-Purdue University Indianapolis, United States

a r t i c l e i n f o a b s t r a c t

Article history: This paper analyzes the interplay between early-life conditions and marital status, as determinants of
Received 3 November 2011 adult mortality. We use individual data from Dutch registers (years 18152000), combined with business
Received in revised form 22 January 2014 cycle conditions in childhood as indicators of early-life conditions. The empirical analysis estimates bivari-
Accepted 21 February 2014
ate duration models of marriage and mortality, allowing for unobserved heterogeneity. Results show that
Available online 13 March 2014
conditions around birth and school going ages are important for marriage and mortality. Men typically
enjoy a protective effect of marriage, whereas women suffer during childbearing ages. However, having
JEL classication:
been born under favorable economic conditions reduces female mortality during childbearing ages.
I12
J14 2014 Elsevier B.V. All rights reserved.
E32
N33
N13
C41

Keywords:
Death
Recession
Life expectancy
Lifetimes
Marital status

1. Introduction
Socioeconomic conditions during infancy and childhood are
well-established predictors of adult health and mortality. Poor
socioeconomic status, leading to poor nutrition, and greater
exposure to diseases in-utero and during childhood are associated
with increased vulnerability to a whole range of health problems
and lower life expectancy (see e.g. Case et al., 2003; Doblhammer,
2004; Van den Berg et al., 2006).
Marriage also affects mortality, with married men exhibiting
lower susceptibility to health problems like cardiovascular dis-
eases and consequently exhibiting higher life expectancy. Unlike
the clear direction of the association between early life conditions
and mortality, the relationship between marriage and mortality
could be bidirectional. Marriage could involve direct material1 or
emotional transfers and care giving from the spouse, which could
increase the well-being of a married individual. More indirectly,
married individuals may experience lower psychological stress
owing to a favorable societal attitude towards married individuals
(Van Poppel, 1992), lead healthier lifestyles (lower consumption of
alcohol or smoking; see Modin, 2003 and references therein) and
adopt less risk taking behavior, which may decrease risk of mortal-
ity. Consequently, marriage may be protective against mortality
(Johnson et al., 2000; Lund et al., 2004; Gardner and Oswald, 2004;
Murray, 2000). But several studies also nd support for a selection

Corresponding author at: Department of Economics, Indiana University-Purdue

University Indianapolis, 425 University Boulevard, Indianapolis, IN 46202-5140,
1
United States. Tel.: +1 312 480 9677. For example by means of economies of scale and specialization within the
E-mail address: sugupta@iupui.edu (S. Gupta). household (Gardner and Oswald, 2004).

http://dx.doi.org/10.1016/j.jhealeco.2014.02.004
0167-6296/ 2014 Elsevier B.V. All rights reserved.
142 G.J. van den Berg, S. Gupta / Journal of Health Economics 40 (2015) 141158
into marriage (e.g. Goldman, 1993a,b) according to which a range
of factors like age, health, social background, income, occupation,
education and race may affect marriage as well as mortality.2 Bet-
ter health, for instance, of married persons is then a consequence
of the selection of healthy persons into and unhealthy individ-
uals out of marriage. Any such positive selection into marriage
would overstate the average causal effect of marriage on mortal-
ity at the individual level. The causal and the selection effect could
both explain the lower mortality of married individuals (for e.g. see
Hu and Goldman, 1990; Lillard and Panis, 1996). The difculty to
control for confounding factors makes it challenging to distinguish
between these effects.
This study analyzes the interplay between early life conditions
and marriage as determinants of high-age mortality. Specically,
we focus on the extent to which early-life conditions inuence the
positive or negative effect of marriage on mortality, and we con-
sider the extent to which early-life conditions inuence the entry
rate into marriage (which we call the marriage rate). To date,
there is no extensive literature on the effect of early-life conditions
on marriage or on the issue whether these conditions affect the
extent to which marriage postpones mortality.
The results of the analysis shed light on the extent to which the
effect of early life conditions on mortality later in life is channeled
through marital status. More specically, the results shed light on
the use of marriage as a compensatory device in case of adverse
early-life conditions. We are interested in the question whether
individuals born in adverse conditions benet from a larger pro-
tective effect of marriage on the mortality rate. Moreover, we want
to know whether individuals born under adverse conditions marry
earlier. If the protective effect is larger for them then this pro-
vides an incentive to marry earlier. Conversely, they may marry
later because they may have worse opportunities in the marriage
market. As such, our paper is informative on the degree to which
an economically bad start in life can be compensated by deliberate
social behavior at prime age, and it is informative on the degree
to which such a pathway is used by those for whom the gains are
highest.
A joint study of conditions early in life, marriage and mortal-
ity requires individual data covering full lifetimes for a reasonably
large sample of individuals. This necessitates the use of historical
data. In this paper we use records from Dutch registers of birth,
marriage and death certicates, covering 18152000. These indi-
vidual level data are then combined with indicators of historical
macro conditions. The individual register data have been used in a
wide range of studies (see Mandemakers, 2000, for an overview)
including studies on the effect of early-life conditions on mortality
(Van den Berg et al., 2006, 2009). These studies have ignored mari-
tal status as an intermediate outcome between birth and death. The
data have also been used in descriptive studies of determinants of
the marriage rate and in studies on the association between marital
status and mortality (see e.g. Van Poppel and Joung, 2001), though
not addressing the issues we focus on.
Section 2 provides a conceptual framework, along with a brief
discussion of the past literature, on the different ways in which
early life conditions inuence selection into marriage, and in
turn how marriage and early life conditions together inuence
mortality. Section 3 describes the individual data along with
some institutional context of marriage over our observation
window. This section also contains a discussion of the historical
context used to identify the causal effects, and of the relevance for
2
For instance, factors like job security, high social background, high income, and
education (Gardner and Oswald, 2004) increase ones marriage prospects and also
favorably affect the individuals survival.
contemporary societies. The empirical approach and the identi-
cation of causal effects in the presence of selectivity are discussed
in Section 4. Section 5 presents the estimation results along with
sensitivity analyses and implications. Section 6 concludes.
2. Conceptual framework and background literature
Fig. 1 presents potential pathways between early life, marriage
and later-life mortality. It contains four arrows, where the arrow
that originates at birth and targets the arrow from marriage to
mortality represents the interaction effect of early-life conditions
and marriage on mortality. The arrow from early-life conditions
to mortality and the arrow from marriage to mortality have been
studied extensively. We focus on the other two arrows: the impact
of economic conditions early in life on the individual rate of getting
married, and the extent to which the protective effect of marriage
depends on conditions early in life. In the empirical analysis, all
four arrows need to be assessed simultaneously. Fig. 1 also displays
boxes representing potential underlying mechanisms through
which early-life conditions may inuence marriage and mortality.
The actual economic conditions in the household early in life and
all the events later in life may jointly depend on shared unobserved
confounders. Such unobserved factors may, for example, inuence
parental earnings as well as adult health of the offspring. As we
explain in detail in Section 4, we capture the actual economic con-
ditions early in life by way of the business cycle in early years of life.
(We postpone a discussion of the empirical model specication and
the identication of models that permit selection on unobservables
to Section 4.)
In general, there may be factors that jointly determine marriage
and mortality, and clearly, one such common determinant of mar-
riage and mortality concerns the conditions in early life. Phillips
et al. (2001) suggest that prenatal growth provides a link between
marital status and mortality. Using a Finnish data set of male births,
the authors note that the marriage rate is positively related to
birth conditions like weight. The authors hypothesize that fetal
conditions may affect an individuals personality, socialization, sex-
uality and emotional responses in later life, consequently affecting
their marital status and mortality later in life. This is challenged by
Vger and Modin (2002) who, using a Swedish longitudinal panel
covering individuals from birth to death, nd no differences in
mortality between ever married and never married women before
versus after adjustment for prenatal conditions, once socioeco-
nomic background (like marital status of mother, social class at
birth) and adult conditions (occupation, education and income)
have been controlled for. The same result is obtained for men: mar-
ried men are less likely to suffer from heart diseases and stroke
than unmarried men, and the risk ratios are unaffected if social
factors at birth are adjusted for. The authors conclude that there is
no evidence for early health status in life being the factor explain-
ing differences in mortality across different marital statuses. These
studies are innovative, but they do not focus on selection on unob-
servables (see below).
Box 3 in Fig. 1 lists mechanisms through which marriage could
causally affect mortality, favorably or adversely. Marriage may
result in transfers of resources and care between spouses that might
positively inuence individual health and life expectancy. Other
potential mechanisms of a favorable causal effect of marriage on
mortality could be lower stress (due to spousal and societal sup-
port), adoption of low risk lifestyle (e.g. reduction of smoking and
alcohol consumption, increased use of safety belts, active lifestyle,
etc.; see Modin, 2003, and references therein) and children as a
means of old-age insurance. Marriage could also have a negative
impact on longevity, for instance due to maternal mortality during
childbirth for women, increased disease exposure in a household
 G.J. van den Berg, S. Gupta / Journal of Health Economics 40 (2015) 141158
Fig. 1. Pathways between early life conditions, marriage, and mortality, over an individuals lifetime.
Fig. 2. Log annual real per capita GNP.
with small children and decline in per capita household resources
due to more resource-sharing.
Early-life conditions may interact with marriage to exert an
additional indirect causal effect on individual mortality. Box 4
in Fig. 2 points to this pathway and lists potential underlying
mechanisms.3
3
The framework in this section covers large strands of literature. We mention a
few additional specic studies and issues. Several studies have used business cycles
as an indicator of early-life economic conditions; see e.g. Van den Berg et al. (2006,
2011). Labor and marriage market opportunities of adults may also be inuenced by
cohort size which in turn may depend on negative shocks in society early in life. Blau,
Kahn and Waldfogel (2000) nd that better female and worse male labor markets
imply lower marriage rates. Brandt et al. (2008) suggest that marriage rates could be
143
3. Data
3.1. Individual records
The Historical Sample of the Netherlands (HSN),4 dataset project
Utrecht-Zeeland-Friesland 02 (UZF.02), is created by merging indi-
vidual data from Dutch registers of birth, marriage and death
certicates. It includes a random sample of individuals born in one
of the three provinces of Utrecht, Friesland or Zeeland5 between
1812 and 1902. The last day of observation of the sample is
December 31, 1999. It records key events in an individuals life
birth, marriage and death and also includes information on
occupation of parents, gender and geographical location. As noted
above, the data have been used in numerous other studies, so the
present exposition can be brief.
Individual lifetime durations are noted in days. If the individual
is still alive at the end of 1999, the date of death is not observed.
inuenced by (1) decreased attractiveness in the marriage market owing to poorer
health related to adverse early-life conditions; (2) increased economic attractive-
ness due to higher labor market demand for the otherwise smaller cohort and (3)
increased marriage market demand due to relative scarcity of famine-cohorts from
younger (women) and older (men) cohorts due to customary gender-differences
in ages of couples. The authors nd no evidence for a net effect on marriage rates.
Direct and indirect effects of early-life conditions can also be driven by conditions
in childhood years (i.e. years beyond the birth year). Some interest has arisen on
the role of conditions during other critical stages of development, like in years after
birth, on health outcomes in later years of life. In the diagram, the conditions in
childhood years can be represented by a similar box as box 1.
4
See Mandemakers (2000) for further details of the sampling structure, goals of
the HSN-database and more information on the HSN.
5
The Netherlands, at the time, had 11 provinces. In terms of economic activity
the three provinces included in this study were jointly very representative of the
country. The same is true for aggregate mortality rates in our data which closely
resemble patterns at the national level.
144 G.J. van den Berg, S. Gupta / Journal of Health Economics 40 (2015) 141158

Table 1 Table 2
Sample sizes of three categories of individuals who reach age 16: those who die Summary statistics of duration variables.
before marriage, those whose marriage and death date are observed, and those
whose longevity is right-censored at marriage. Percentages by gender. Never-married Married-dead Married-censored

Age of marriage
Number of % Never % Married-dead % Married-censored
Full sample 28.32* 27.90*
individuals married
[26.34] [26.21]
Total 5593 21.38 52.91 25.71 (7.48) (6.67)
Men 2709 23.77 52.60 23.62 Men 29.59* 28.97*
Women 2884 19.14 53.19 27.67 [27.32] [26.99]
(7.90) (7.14)
Women 27.14* 27.06*
[25.42] [25.51]
The lifetime durations of individuals whose death dates are not
(6.86) (6.14)
observed are right-censored at their last day of observation, i.e. at Age of death
birth or at marriage. As the legal age of marriage at the time was 16 Full sample 57.10* 70.82*
years and the people who died before this age cannot contribute [65.89] [74.76]
to the likelihood of marriage, these people are excluded from the (25.22) (16.36)
Men 53.87* 71.27*
model estimation. Further loss of observations owing to missing [61.02] [74.34]
values of explanatory variables results in a nal sample of 5593 (24.91) (15.07)
individuals. Women 60.88* 70.40*
Social class is captured by a hierarchal index based on Van [70.36] [75.10]
(25.09) (17.47)
Tulder (1962)s mapping from parental occupation into a six-layer
(1 being the lowest and 6 the highest) hierarchical scale.6 This has *
: Mean, [] : median, () : standard deviation.
been used in many other studies with the HSN data. In general, our
observed explanatory variables for individual marriage and mor-
tality are realized at birth as opposed to those acquired later in
life, as the latter may be endogenous or confounded. The place of system. Instead, people without work could try to obtain some
residence at birth is included as a binary choice urbanization indi- income support from charities and churches. The absence of an UI
cator which takes a value of 1 if the person is born in a city and system means that the business cycle is a relatively powerful deter-
0 otherwise. One disadvantage of the unprecedented observation minant of household income. This makes an empirical approach in
window is the absence of several explanatory variables commonly which economic conditions early in life are instrumented by the
used in the mortality literature but unobserved in population regis- business cycle relatively informative if data from such older cohorts
ters. Notably, we do not observe the individuals adult health status are used.
and access to public health services and medical technology. It is also important to point out that recent medical research has
To provide summary statistics of outcomes of interest, we dis- shown that not just fetal malnourishment is associated with long-
tinguish between 3 types of people. The rst group consists of run effects on mortality, but that discrepancies between early-life
people who never marry during their lifetime and therefore no conditions in utero and shortly after birth on the one hand, and
marriage date is observed for them. The second group comprises of later lifestyle on the other hand, lead to long-run effects (see e.g.
individuals for whom marriage as well as death dates are recorded. Mogren et al., 2001; Holemans et al., 2002; see also Fogel, 1997, for
Those in the third category are not observed after their marriage an overview). So in modern as well as historical societies, individu-
and therefore their death dates are missing. Table 1 presents the als born in low-income households with higher nutritional intakes
sizes of the three groups and their composition in terms of gender. later in life may be at risk at higher ages. Note that the virtual dis-
Table 2 gives further details for these three groups of people, once appearance of infant mortality implies that those who would have
again by gender, for the duration of marriage and mortality, where died if born under adverse conditions in the nineteenth century
observed. nowadays survive into adulthood. This can be seen as a factor that
Life-course study of late-life mortality risks of individuals more contributes to the potential relevance of long-run effects in modern
or less mandates the need for historical data with full individual societies.
lifetimes of a reasonable sample size. However, use of historical It is possible that the causal pathway from economic conditions
data requires some understanding of the institutions of the time. to mortality at higher ages is amplied by stress at the household
The following subsection aims to provide some context. level shortly before and/or shortly after birth. Adverse aggregate
conditions may induce stress by way of (a higher perceived prob-
ability of) income loss. Clearly, if stress is driven by exposure to
3.2. Historical context and current relevance
adverse information and by relative deprivation compared to the
relevant population, then stress effects may be as important today
All births and most marriages in the data took place more
as they were say 100 years ago. Moreover, stress may exacerbate
than 50 years ago. This calls for a discussion of the relevance
the effects of adverse nutrition and health infrastructure by the
of studies with such historical data for contemporary long-run
perception that nutrition and health care will be insufcient. Farah
health effects of early-life conditions. Clearly, at the time, economic
et al. (2008) nd that poverty in the household leading to stress
and nutritional conditions in the Netherlands were different from
causes neurological damage among children in the rst years of
current conditions. During the birth cohort observation window,
their life, resulting in a lower IQ.
the Netherlands did not have an unemployment insurance (UI)
For current developing countries, which in certain aspects may
be regarded as similar to or worse off than the Netherlands in the
period evaluated in the present paper, the existing literature has
6
Examples of common 19th century occupations in Van Tulder (1962) six social focused on inequalities in infant and child mortality by household
classes are as follows: (1) diker, day laborer, dock worker (2) beer brewer, farm
laborer, orist, sherman (3) barber, baker, shoe smith, carpenter (4) bailiff, mer-
socioeconomic status, since there are typically no long run data
chant, innkeeper (5) factory manager, headmaster, infantry captain (6) lawyer, registers (see Sastry, 2004). In this sense, our paper aims to com-
pharmacist, surgeon. plement these studies by studying long-run mortality effects. In
 G.J. van den Berg, S. Gupta / Journal of Health Economics 40 (2015) 141158
Section 5 we will see that our results suggest some policy recom-
mendations for developing countries.
To assess the relevance of historical evidence of marriage effects,
it is useful to cite some more information on the institution of
marriage in Northwestern Europe within our observation window.
Petersen (1960), Matthijs (2003), Hartman (2004) and Dribe and
Lundh (2010) provide overviews, and the latter study also cites
other recent overviews. In most of the 19th century, the average
age at marriage was rather high, a substantial fraction of individ-
uals remained bachelor, and the average age difference between
men and women was about 2 years. These features are also visible
in our data; see Table 2. Socio-economic origin, age, and geograph-
ical origin have been found to be particularly strong determinants
of marriage choice. Perhaps more importantly for our purposes is
the fact that many women left the labor force upon marriage, and
instead bore children and took care of the household. In the absence
of societal social safety nets, the economic and care dependence of
spouses was high.
One may therefore expect that marriage was even more impor-
tant for survival than nowadays, in the Netherlands. Historical
records also note a societal contempt for unmarried women beyond
the age of 30 (Van Poppel, 1992). And once married, divorces were
quantitatively unimportant throughout our observation window.
Reporting on cohorts born 18151922, Van Poppel (1997) shows
that although divorce rates gradually increased over the 19th cen-
tury in the Netherlands but that they nevertheless remained very
low and at values below 2.5%. Widow(er)s frequently re-married
unless they were old (Van Poppel, 1998).
In the 19th and early 20th century, the only thing probably
worse than being a spinster was the stigma attached to unwed
motherhood (see Modin, 2003; Modin et al., 2009). Despite this, in
some parts of Northwestern Europe, at the turn of the nineteenth
century the share of illegitimate births could be as high as 11% on
average and up to 20% in rural areas. In addition to the disadvantage
of being born to mothers usually hailing from the lowest stratum
of society these children were often deeply neglected, receiving
little sympathy or support from the general public. Infant mor-
tality rates amongst children born-out-of wedlock were therefore
high. In light of these societal norms, women who conceived out-of-
wedlock would have all the incentive to marry quickly if possible.
In the absence of reliable fertility records for the individuals in
our sample we are unable to comment on the timing of their mar-
riage in relation to start of fertility. While rates of illegitimacy in
the Netherlands are noted to be lower compared to Sweden at the
time, conception-out-of-wedlock could be another determinant of
marriage.
On the other hand, for the majority of the women, the associa-
tion of marriage and childbirth means that women were exposed to
life-threatening experiences during childbearing ages post matri-
mony. This, again, also applies to the current situation in developing
countries. According to Dasgupta (2011), maternal mortality is the
largest single cause of death among women in their reproductive
years, in several poor countries. In Sub-Saharan Africa, maternal
mortality is 90 times larger than in high-income countries (World
Bank, 2010).
4. Empirical methodology
4.1. Outline of the approach to identify causal effects
Consider again Fig. 1 in Section 2. We estimate a joint model for
all four causal effects, taking into account that all variables may be
affected by unobserved determinants. To deal with the potential
endogeneity of the actual economic conditions in the household
145
early in life, we use business cycle conditions in early years of life
as an indicator of individual economic well-being. To explain how
we deal with the endogeneity of marital status and the marriage
date, it is useful to think of our data as containing two subsam-
ples: one with individuals born in adverse years and one with
individuals born in favorable years. Suppose that for each of these
two subsamples, one can estimate the full model separately (this
requires identication for given early-life conditions that are equal
for all subsample members). The estimation would give two sets
of results, including two different effects of marriage on mortality.
Contrasting these results then reveals the effects of the indicator of
early-life conditions on all parameters in the model, including the
interaction effect of the indicator of early-life conditions and the
effect of marriage on mortality.7 Thus, to disentangle the long-run
effects of conditions at birth and during childhood, we exploit the
cross-individual variation in the timing of the business cycle early
in life.
What remains is to demonstrate that each of these two sepa-
rate models is identied for given early-life conditions, even if the
marriage date is endogenous in the sense that marriage and mor-
tality have shared unobserved determinants. For this we rely on
the Timing-of-Events approach for the identication of the effect
of the occurrence of one event on the rate that another event
occurs (see Abbring and Van den Berg, 2003). In our context, the
former event is marriage and the latter is death. For given early-
life conditions, our model reduces to a bivariate duration model
for the duration until marriage and the duration until death, with
a causal effect of marriage on the mortality rate and with possi-
bly shared unobserved determinants of the marriage rate and the
mortality rate. If marriage were not endogenous, conditional on
observable characteristics, then the causal effect of matrimony on
mortality could be estimated in a straightforward way as the coef-
cient of a time varying regressor that takes a value of 0 prior
to marriage and 1 thereafter. However, as discussed above, sev-
eral unobserved factors, like health, could drive the timing of both
marriage and mortality. So for instance, a healthy person might
marry earlier due to increased attractiveness in the marriage mar-
ket but might also die later. Inability to take into account this
potential correlation between unobserved factors driving marriage
and mortality would then bias the estimates of the causal effect of
marriage on mortality upwards. Identication of the causal effect of
potentially endogenous marriage on mortality therefore requires
explicit accounting for this correlation structure. The Timing-of-
Events approach allows for the identication of such models,
with correlated unobserved heterogeneity terms for marriage and
mortality,8 provided that (i) the unobservables are independent of
the observed explanatory variables (i.e. the random effect assump-
tion holds), and (ii) the marriage and mortality rates satisfy Mixed
Proportional Hazard (MPH)-type specications. It is important to
point out that for given conditions early in life, the identication
of the marriage effect does not require an instrumental variable
(in the usual sense of the word) for marriage. Identication effec-
tively exploits the fact that the causal marriage effect on mortality
7
Notice that the empirical inference produces estimates of effects of the business-
cycle indicator of early-life conditions; not of the effects of all the early-life
conditions (such as parental household income) themselves. The extent to which
one can generalize the conclusions about the former effects to conclusions about
the latter effects depends on the extent to which the effects of cyclical uctuations
in economic conditions run through the actual conditions early in life. We return to
this issue in detail in Section 4.5.
8
See Section 4.3 for details. By now, this approach has been used in dozens of
studies. Richardson and Van den Berg (2013) contains an empirical example in a
labor economics setting. They discuss in detail which assumptions are required to
hold for the method to be applicable.
146 G.J. van den Berg, S. Gupta / Journal of Health Economics 40 (2015) 141158
Fig. 3. The distribution of the cyclical indicator of GNP over birth years.
only works from the entry into marriage onwards, whereas the
association between marriage and mortality due to shared deter-
minants works at all ages. This distinction between local and
global dependence requires a minimal amount of randomness in
the moment of marriage, which in turn can be justied by the
random element in the arrival of opportunities in the marriage
market.
The set-up of the remainder of this section is as follows. First
we examine the data on the business cycle. Next, we discuss the
Timing-of-Events approach and the parameterization of the model.
Finally, we discuss the conditions required for the use of business
cycles as indicators of transitory economic household conditions
early in life.
4.2. Business cycle indicators
In principle, one would like to compare cohorts born in
economic booms to those born during economic troughs with oth-
erwise identical conditions over life. This however is not possible
due to the steady secular improvements in living conditions, med-
ical innovations and health care over time. As Van den Berg et al.
(2006) point out, this issue can be dealt with in a conservative way
by comparing a cohort born in a specic boom to that born in the
subsequent recession, because the latter benet more from secu-
lar developments. In this case any reduction in longevity can be
attributed to the cycle. However, given the small sample size of
5593 individuals, it is impossible to compare outcomes from two
consecutive stages of a cycle, especially if we wish to consider the
two genders separately. Therefore, following Van den Berg et al.
(2006) we identify intervals of successive years of booms and
recessions.
To obtain cyclical indicators, we decompose the historical time
series on log annual real per capita GNP over the observation win-
dow. The GNP series, instead of the conventionally used GDP, is
chosen purely for reasons of availability of mutually consistent
observations over the years of interest. Due to the unavailability
of GNP data for the years 18121814, we only consider individuals
who were born in and after 1815. Fig. 3 presents log annual real
per capita GNP, measured in 1000 Euros with 1995 as a base year.
The graph highlights both the upward trend and the many cyclical
uctuations.
We perform a trend/cycle decomposition of the log annual real
per capital GNP using the Hodrick-Prescott lter with smoothing
parameter 500. Fig. 3 presents the cycle and trend of the GNP
series as function of calendar time. The plot shows that periods of
economic booms and recessions are clearly identiable in the data
and Van den Berg et al. (2006) show that these are robust with
respect to the choice of the decomposition method and smoothing
parameter. Fig. 4 provides a histogram of the distribution over time
of the cyclical term (or indicator) which will be our main explana-
tory variable. For most of the analysis we round-off the value of the
cyclical term to a binary outcome representing economic upturns
or downturns.
We additionally control for incidence of epidemics9 and war.
World War II (194045) has been the only war and occupation
on Dutch soil since Napoleon. It included the severe famine of the
winter of 1944/45, where mortality rates peaked because of malnu-
trition (Jewish genocide victims were excluded from the data). The
period of the World War II is represented using a separate dummy
variable as no reliable macro-economic data is available for this
interval.
4.3. A bivariate duration model for marriage and mortality
For convenience, we use lower case symbols for random vari-
ables as well as their realizations, and we abstract from potential
outcome variables. The outcomes of interest are the duration until
marriage tm and the duration up to death, td . We assume that
all individual differences in the joint distribution of these two
processes is conditional on calendar time , other socioeconomic
and demographic factors x, current macro-economic conditions
z(), trend components and cyclical indicators ztr ( t + i) and
t + i) of macro-economic
zc (   conditions in earlier years of life
(i 0, . . ., t 1 ), various interaction terms, and the unobserved
characteristics v. The vector x contains covariates that are observ-
able at birth, and v is independent of x. Let tm be the moment at
which an individual gets married. The indicator function I(tm < t)
is used to denote whether an individual is married at age t or
not. The parameter captures the causal effect of marriage on the
individual mortality rate. This parameter may in turn depend on
various variables, including of course the early-life conditions. The
studies mentioned in the previous sections suggest that the effect
strongly depends on the age of the individual under considera-
tion.
For ease of exposition, we may normalize  for a given individual
such that  = 0 at birth, and we may subsume the variables z(),
ztr ( t + i) and zc ( t + i) in x, which accordingly becomes a time-
varying explanatory variable x(t). The unobserved determinants of
the individual mortality rate and the marriage rate are denoted by
vd and vm , respectively. The mortality rate at age t, conditional on
x(t), vd , and tm is denoted by d (t, x(t), vd , tm ) and is assumed to
have the following specication
d (t, x(t), vd , tm ) = d (t) exp(x (t)d + I(tm < t) + vd ) (1)
For z() we take log annual real per capita GNP at t, as well as
dummy variables for years with epidemics and World War II.10
Since the focus is on the impact of marriage on mortality, we
only consider the set of individuals who reach age 16, i.e. the legal
age of marriage. The individual is exposed to the possibility of
marriage thereafter and we only consider outcomes after this age.
However, we include information about the conditions prevailing
in years of early childhood (16 years of age), leading right up to
9
See e.g. Almond and Mazumder (2006) and Almond (2006) for discussion on
adverse effects of in-utero exposure to an epidemic on long term individual out-
comes.
10
This takes care of the fact that GNP series is missing for the years of World
War II.
 G.J. van den Berg, S. Gupta / Journal of Health Economics 40 (2015) 141158
Fig. 4. Impact of marriage at age 30 on mortality rate, by gender and by state of the business cycle at birth.
16 years (ages between 712 and 1315 years). These are summa-
rized by four binary indicators. A binary boom/recession dummy is
used to record a favorable (cyclical component > 0)/adverse (cycli-
cal component < 0) period of the business cycle in the year of birth.
3 additional indicators of average cyclical macro-economic condi-
tions during the age intervals of 16, 712 and 1315 years are
also included by means of dummies for whether or not the aver-
ages of the cyclical element of the GNP series is positive or negative
between these ages.
The trend component ztr ( t + i) of the GNP series in the years
of birth and childhood, obtained from the Hodrick-Prescott decom-
position, captures the secular long-run effects. Van den Berg et al.
(2006) discuss in detail the empirical challenges in distinguishing
the effects of these trend components from the effects of current
log GNP z() and age t, due to multicollinearity. All these variables
are by and large increasing over time, and at the individual level
current log GNP can be captured fairly well by the sum of the trend
component of the GNP in early years of life and an increasing func-
tion of age. Consequently, the trend component is omitted from
the model specication till age 16, only after which the individual
faces the risk of marriage and enters our sample. For similar rea-
sons calendar time  is also left out. We do, however, include the
contemporaneous trend conditions in the macro-economy for each
year of life beyond age 16. The trend component of the series cap-
tures all secular effects from birth to the current age and therefore
inuence marriage and mortality. We also include the contempora-
neous cyclical conditions in the macro-economy which are also an
indicator of current employment opportunities and the resulting
marriage market conditions. Inclusion of the cyclical component of
the GNP controls for the impact of contemporaneous employment
conditions which have been noted to inuence the marriageability
of the individual (for instance Blau et al., 2000).
147
For an individual of age t years who is not yet married, the
marriage rate at t conditional on observed and unobserved char-
acteristics x(t) and vm is denoted by m (t|x(t), vm ) and is assumed
to have the MPH specication given by,
m (t|x(t), vm ) = m (t) exp(x (t)m + vm ) (2)
where once again x(t) is independent of vm and the individuals
background characteristics x are constant over time.
Now consider the joint distribution of td and tm . Conditional on
x(t), vd and vm , the only possible relation between the variables td
and tm is the relation by way of the direct/causal effect of a mar-
riage on the mortality rate. In case of independence of vd and vm ,
we would have a standard duration model for td |x(t), tm in which
I(tm < t) can be treated as a time-varying regressor that is orthogo-
nal to the unobserved heterogeneity term vd . However, if vd and vm
are not independent, inference on td |x(t), tm has to be based on td ,
tm |x(t). This means that we need to integrate the individual likeli-
hood contribution over the distribution G(vd , vm ) which is the joint
distribution function of the unobserved characteristics vd and vm
among 16 year olds.
We estimate the bivariate duration model with full maxi-
mum likelihood. See Abbring and Van den Berg (2003) for the
identication proof for the full model including what in our case
is the causal effect of marriage on mortality, in the presence of
selection. As noted above, crucial underlying assumptions for point
identication are that unobserved determinants are independent
of observed determinants and that the model structure is correct.
Without the former, the model is not identied. This is of course
unfortunate; for example, it is conceivable that unobserved genetic
determinants affect both the social class of the parents and the
health outcomes of the child. With data on siblings one could
hope to deal with this by exploiting within-family variation in
148 G.J. van den Berg, S. Gupta / Journal of Health Economics 40 (2015) 141158

outcomes. However, such data are not at our disposal. The best born during an economic recession (Iboom = 0) the total causal effect
one can do in random effects analyses is to include as many is
covariates as possible. We follow this approach. As we shall see,
this brings us to the boundary of what is computationally feasible. 
4
ape
recession = i Ui (t) (4)
The assumption that the model structure is correct entails that
i=0
there must be a minimal amount of residual randomness in the
timing of marriage and that the observed covariates do not have The time dependence of the hazard functions is similarly
identical quantitative effects both on marriage and on mortal- expressed using Chebyshev polynomials of the second kind of
ity. In our setting, the residual randomness naturally follows degree 4 in age of the person. Thus, the duration dependence of
from the randomness in the arrival of prospects in the marriage the marriage and mortality rates are respectively given by:
market.    
Strictly speaking, the effects of zc ( t) in the model (including 
4

4

any effect on ) capture the causal effects of the business cycle at m (t) = exp m
i
Ui (t) and d (t) = exp di Ui (t) (5)
birth. Suppose that the actual economic well-being of the parents i=0 i=0
household at birth equals a permanent component plus a transitory
component. If the permanent component is captured by social class These specications contain 10 parameters in total (m i
and di ,
and the transitory component equals a proportion of the business with i = 0, 1, . . ., 4).
cycle indicator, then the quantitative effects of the business cycle We take the joint distribution of the unobserved heterogene-
indicator may be translated into quantitative effects of the actual ity terms vd and vm to be bivariate discrete, with two unrestricted
economic conditions at birth. mass-point locations for each term. Let v1d , v2d , v1m , and v2m denote
the points of support of vd and vm , respectively. The associated
probabilities are denoted as follows:
4.4. Parameterization
Pr(vd = v1d , vm = v1m ) = q1 , Pr(vd = v2d , vm = v1m ) = q3
We use exible specications for the causal effect of mar- Pr(vd = v1d , vm = v2m ) = q2 , Pr(vd = v2d , vm = v2m ) = q4
riage on mortality rate, the duration dependence functions, and
the bivariate unobserved heterogeneity distribution. In order to with 0 qi 1 for i = 1, . . ., 4, and q4 = 1 q1 q2 q3 .
allow the causal effect of marriage on mortality to vary over the The covariance of vd and vm is given by, cov(vd , vm ) = (q1 q4
individuals lifetime, is expressed using Chebyshev polynomials, q2 q3 ) (v1d v2d ) (v1m v2m ). We note that cov(vd , vm ) = 0 would
for instance of degree 4, in the age of the married individual. This imply independence of vd and vm and q1 = q4 = 0 or q2 = q3 = 0
polynomial could be specied simply as a sum of terms i ti , i = 0, would mean perfect correlation. In a regression framework, the
1, . . ., 4 where t is the age of the individual. However, since the set-up would be similar to a non-parametric, joint specication
terms of ti are not orthogonal, estimation of the parameters i is of the error structures of a triangular system of two simultaneous
aficted by multicollinearity. We take care of this problem by using data generating processes marriage and mortality. Identication
Chebyshev polynomials of the second kind. In this case, the poly- would follow therefrom.
nomial is specied as a sum of terms i Ui (t), i = 0, 1, . . ., 4 where
U0 (t), U1 (t), U2 (t), U3 (t) and U4 (t), 11 are mutually orthogonal poly-
4.5. Business cycles as indicators of transitory economic
nomials of indexed degree. Furthermore, we allow this effect of
conditions early in life
marriage on the mortality rate to vary by conditions in early life of
the individual by including interaction terms of and the dummy
In this study, business cycle conditions in early years of life
of being born in a boom or not as determinants of the mortality
are used as indicators of individual economic conditions only in
rate.
those years. This approach is similar to the use of the season of
This set up implies that the total age dependent causal effect of
birth (Doblhammer, 2004, although seasons are perfectly antici-
marriage on the mortality rate for a married individual would be
pated and hence fertility can be ne-tuned to seasonal conditions)
given by
or the use of food price uctuations early in life (Bengtsson and
 4  Lindstrm, 2003). The business cycle is not used as an instrumen-

4
ape
 tal variable in a conventional IV analysis, since after all we do not
= i Ui (t) + a,int
i
Ui (t) Iboom (3)
observe actual household income or actual nutritional intake in
i=0 i=0
the early years of life. In the terminology of the evaluation liter-
ape ature, we do not observe the treatment status. Nevertheless, a
where i are the parameters corresponding to the mutually
reduced-form analysis of the effect of the indicator on the outcome
orthogonal polynomials Ui (t) of degree i capturing the age depen-
of interest reveals whether a causal effect of the treatment on the

dence of the causal on individual death rate. The
effect of marriage
4 outcome exists. Like with IV, we require two conditions to be sat-
second term a,int Ui (t) Iboom
i=0 i
takes into account any effect ised: informativeness and the exclusion restriction. We discuss
of early-life conditions, represented by the dummy (Iboom ) for being each in detail here.
born in a year of economic boom or not. The parameters a,inti
are In our setting, informativeness means that the state of the
the coefcients of this interaction term. Consequently, for a person business cycle must have an effect on economic conditions in
the household. In the absence of observations of the latter, this
is not trivial to demonstrate at the individual level. In the 19th
century, individual data on income, consumption or assets were
11
To start, the domain of the ages t where t [0, 103] is linearly transformed to not collected in a systematic fashion, and tax records from that
the domain of the orthogonal Chebyshev polynomials such that now
t [1, 1]. era are not digitized and were not collected centrally. Census
This is done in our case by using the simple rule
t = 2(t t0 )/(nt 1) 1 where
nt is the year of the individuals life that is being considered. Then our orthogo-
data collected in the Netherlands throughout the 19th century
nal polynomials are: U0 (
t) = 1, U1 (

t) = 2

t, U2 (

t) = 4

t 2 1, U3 (

t) = 8

t 3 4
t, U4 (

t) = are uninformative on income and assets. Ultimately, this reects

16
t 4 12
t 2 + 1. the trade-off between being able to observe household-specic
 G.J. van den Berg, S. Gupta / Journal of Health Economics 40 (2015) 141158
transitory conditions around birth and being able to observe
individual mortality in late life.
In our observation window of birth years, there are also no
aggregate time series on food intake or average household income.
At the same time, as we have seen in Section 3.2, GNP does reect
economic production and, ultimately, household income. Historical
records do provide an uninterrupted macro time series on the mean
real per capita wage. To proceed, we estimate the relation between
uctuations in this series and the concurrent business cycle. We
nd a strong and signicant positive association.12 This means that
disposable income among employed workers did indeed uctuate
with the business cycle. Note that in our birth cohort years there
was no welfare safety net, so that effects of changes in GNP on
household income across the normal range should be stronger than
in modern times. Other uninterrupted national time-series of vari-
ables that may depend on the business cycle are largely absent in
our birth years window.
Lindeboom et al. (2010) nd a strong negative correlation in
those years between the business cycle on the one hand and yearly
uctuations in real food prices (rye, potatoes and wheat) on the
other hand. This suggests that business cycle uctuations do indeed
coincide with uctuations in nutritional intake (and this is fur-
ther reinforced by our own nding above that real wages move
in tune with the cycle). The authors use both the business cycle at
birth and the uctuations in real regional food prices at birth as
explanatory variables for mortality at high ages and conclude that
the business cycle has more explanatory power. This is not surpris-
ing given that food intake comprises more than just rye, potatoes
and wheat. Jacobs and Smits (2001) show that the business cycle
in the Netherlands closely follows the business cycle in the U.K. in
those years.
The informativeness requirement may be violated for high
social-class parents who are able to use their long-run economic
status to insure against shocks. Ignoring this will typically lead to
an under-estimation in absolute value of all the long-run effects of
adverse economic conditions early in life. We deal to some extent
with these issues by including social class in the set of explanatory
variables in our analyses. Van den Berg et al. (2009), using a com-
bination of our data set and additional data sources from the 18th
and 19th century in the Netherlands, demonstrate that the social
class measure is strongly correlated to the permanent income of the
father. Additionally, they nd that for the 20% of parents with the
highest social classes, the business cycle at birth does not have long
term mortality effects on the children. This could be due to selection
on unobservables, but it is also possible that the 20% richest house-
holds were insensitive to business cycle shocks. In that case the
instrument is not informative for them. We therefore re-estimate
our full model excluding the subsample of individuals whose par-
ents were in the 20%13 with the highest social class. The results
are discussed in Section 5. They reinforce the view that the busi-
ness cycle was an important determinant of household economic
conditions.
We now turn to the exclusion restriction. In our context this
means (i) the business cycle is exogenous at the individual level,
and (ii) the business cycle around birth has no additional long-run
effects on marriage and mortality later in life on top of the effects
we are after, namely the effects running through the parental
households living conditions. Part (i) is trivially satised. Part (ii)
is violated if a particular cyclical uctuation has a long-lasting
12
The regression coefcient for the regression of log per capita real wages on the
cyclical component of log per capita real GNP, between 1815 and 1902, is 1.08 with
a p-value of 0.00).
13
This reduces our sample size to 2160 men and 2317 women.
149
secular effect on society in such a way that only specic cohorts are
affected in the long run. This seems very unlikely. For example, it is
unlikely that a temporary downturn in governmental investments
would cause a particular birth cohort to suffer in the long run
from adverse societal conditions. Upswings and downturns occur
with a certain frequency and as the individual ages, the previous
uctuations wash out, and so do their possible secular effects.
Indeed, in our birth cohort observation window, the frequency of
cyclical uctuations was substantially higher than nowadays.
Part (ii) is also violated if the composition of newborns system-
atically varies with the business cycle. In particular, the fertility
of low social class parents may be negatively affected by reces-
sions, causing individuals born in recessions to come on average
from higher social classes. Ignoring this will typically lead to an
under-estimation in absolute value of all the long-run effects of
adverse economic conditions early in life. Past literature however
does not nd evidence supporting the presence of such endoge-
nous fertility in birth cohorts from before 1940. Van den Berg et al.
(2009), in a sample resembling ours, found that the size of birth
cohorts and their composition by social class do not change with
the business cycle at birth. They also relate the fraction of new-
borns per year whose father has an occupation in the countryside
to the business cycle at birth. Again, the relation turns out to be
weak and insignicant. Other studies with data from Northwest
Europe from around 1900 also fail to nd that the composition of
newborns is systematically related to uctuations in macro indi-
cators early-life conditions. Kreholt (2001) studies Swedish birth
cohorts from 1897 to 1938 and examines whether the fraction
of newborns whose father had a blue collar (versus white collar)
occupation varies with the state of the business cycle as measured
by the annual change in the inow into poor relief. The results
show that there is no signicant variation, neither among male
nor among female newborns. Van den Berg et al. (2011) conclude
that in Denmark in the years 18731906 there was no relation
between the business cycle on the one hand and the birth rate and
the composition of newborns in terms of a whole range of personal
characteristics on the other.
Dehejia and LLeras-Muney (2004) report evidence that in recent
years the birth rate among individuals with higher education or
status may be higher in recessions. At rst sight this contrasts to our
ndings. To explain their ndings they mention that opportunity
costs are lower in recessions. However, the Netherlands in our birth
cohort years did not have unemployment benets or welfare, so
that a job loss would immediately lead to time-consuming efforts
to obtain alternative sources of income. Moreover, macroeconomic
uctuations had a higher frequency than nowadays, complicating
the ne-tuning of births in the 19th century, and fertility control
was less effective than nowadays.
Even if the cohort composition of newborns is invariant over
the cycle, there may be a long run effect of cohort size on the
marriage rate. This is suggested by some of the literature discussed
in Section 2. If cohort size is pro-cyclical then the marriage rate in
the cohort may move in tune with the cycle at birth. This would
violate the exclusion restriction in the sense that we have an effect
of the business cycle at birth on marriage that does not run through
individual economic conditions at birth. However, such a violation
only affects the interpretation of an association between the cycle
at birth and the marriage rate. It does not affect the interpretation
of the other ndings. To proceed, we examine national time series
of birth rates by gender.14 Regressing total births on the cyclical
component of the GNP series, rst in the year of birth and then
14
These cover the whole of the Netherlands and are from the Human Mortality
Database (www.mortality.org).
150 G.J. van den Berg, S. Gupta / Journal of Health Economics 40 (2015) 141158
separately in the year prior to birth, we nd no signicant
correlations15 and hence little evidence of endogenous fertility. To
get a closer look at the availability of potential marriage partners
we consider sex ratios in birth cohorts over time and nd that these
remained stable in the Netherlands at the time. Regressing sex
ratios on cyclical component of GNP series we nd no signicant
correlation at age 0 or at age 16.16 In sum, birth cohort sizes do
not vary with the business cycle. To examine whether any residual
variation in birth cohort sizes affects marriage rates we would
need to take into account that men on average marry women who
are a few years younger. In our data, information on ages of grooms
and brides is only available for the province of Utrecht. Using this
subset of the data, we nd that the age difference between the
groom and the bride is approximately normally distributed with
the mean close to 2 years and a standard deviation of 5.84 years.
It turns out that the ratio of men to 2-year younger women shows
only marginal variation over time. In sum, we do not nd evidence
for a pathway from the business cycle to birth cohort sizes to the
rate of available partners, so that any association between the
cycle at birth and the marriage rate is not driven by a violation of
the exclusion restriction.
In general, in the study of long-run effects of early-life condi-
tions, there is a certain tension between the two requirements of
informativeness and the exclusion restriction. Dramatic exogenous
shocks at birth, such as famines and epidemics, are very informa-
tive. However, they also give rise to a strong fertility reduction,
especially among lower social classes, leading to a violation of the
exclusion restriction. For instance, Stein et al. (1975) showed that
during the Dutch hunger winter 19441945 the fertility reduction
was lower among groups of higher socioeconomic status. Dra-
matic shocks may induce a restructuring of society, giving rise
to an additional violation of the exclusion restriction. Hence, the
best indicator is not necessarily the strongest indicator. The evi-
dence to date suggests that in this sense, business cycles have some
major advantages. For one thing, they do exert long-run effects;
their long-run effects on mortality, cause-specic mortality, and
cognitive ability are well-documented for older birth cohorts, in
a range of countries, as mentioned in Section 1 (see e.g. Van den
Berg and Lindeboom, 2014, for an overview; see also Woitek, 2003,
for extensive historical evidence relating the business cycle around
birth to adult height which is a marker of physical tness later in
life). This suggests informativeness for at least a substantial part of
the population. Moreover, as we discussed, there is a substantial
amount of evidence that among pre-1940 cohorts, fertility and the
composition of newborns are not sensitive to the business cycle.
5. Results
5.1. Estimation results by gender
Table 3 presents the estimation results for the full model. For
the estimates concerning entry into marriage, a positive value is
associated with an earlier marriage. For the mortality rate, positive
values of estimates signify a shorter lifetime.
The rst crucial result is a signicant causal effect of being mar-
ried on the mortality rate for both men and women. Gender differ-
ences however are vital, and so are age and the conditions around
the individuals birth. The estimates of the size of the effects of the
latter are hard to read from the separate coefcients in Table 2. Fig. 5
therefore presents, in a compact way, the impact of marriage on the
instantaneous mortality rate by gender and also separately for birth
15
p-Values of 0.37 and 0.84, respectively.
16
p-Values of 0.13 and 0.33, respectively.
during an economic boom versus a recession, as implied by our esti-
mation results in Table 2. For sake of exposition, the instantaneous
hazard rates are plotted for marriage at the age of 30. Similar results
(not presented) are obtained for other ages of marriage. The gures
also plot the mortality rates in the absence of marriage. To show the
effect of marriage without contamination from other explanatory
variables, each panel averages over the other explanatory variables.
Note however that the model includes a time-varying explanatory
variable (GNP trend) whose effect runs against the age dependence
of the mortality rate. Therefore, care should be taken when inter-
preting the increasing shapes of the mortality rates in the gure.
The four panels of Fig. 5 clearly show a sudden jump in the
mortality rates at the moment of marriage. For women born in a
recession, the mortality rate shows a large and long lasting increase
in response to marriage. Favorable early life conditions help mit-
igate the adverse effect of marriage for women. For men, on the
other hand, the impact of marriage on the mortality rate is favor-
able irrespective of cyclical conditions pertaining in the economy
at the time of birth. This result goes against the view that there is
no protective effect of marriage for women. Instead, the degree of
protection depends on early-life conditions and changes with age.
In any case, part of the association of marriage and life expectancy
is causal for each gender.
We aim to determine whether the causal effect of marriage on
mortality is statistically signicant or not. Fig. 6 displays the esti-
mates of the age-dependent marriage effect interacted with the
indicator of early-life conditions on mortality, along with the 95%
condence intervals, on the log mortality rate (that is, is plotted
rather than exp ). Once again, given large gender differences and
the vital role of the economic conditions in early years of life, the
plots are presented separately by gender and condition of the busi-
ness cycle in the birth year ( from Eq. (3) above for those born in
an economic boom and from Eq. (4) instead for those born during
economic recessions). The interpretation of the plots is as follows.
Marriage signicantly increases the hazard of death during the
years in which the lower 95% condence interval of the age depen-
dent causal effect of marriage on mortality becomes positive (i.e.
goes above 0 on the Y axis). Likewise, negative values of the upper
condence interval of the marital causal effect imply a signicant
lowering of the hazard of death due to marriage (i.e. a protective
effect of marriage on death). In both Fig. 6(a) and (b) the causal
effect of marriage on death hazard changes continuously with age
reiterating the presence of an age-dependent causal effect of mar-
riage on mortality for both genders. The relevance of the early life
conditions is suggested by the analysis of the instantaneous mor-
tality rate around the time of marriage above (Fig. 5). We proceed
to discuss the results for women and after that for men.
For women (Fig. 6a) we nd that marriage instantaneously sig-
nicantly raises the mortality rate in the childbearing ages. This
adverse effect of marriage is much stronger and persists for a
longer time in case of birth under adverse economic conditions.
By comparing the top and bottom graphs of Fig. 6(a), we can see that
women born during recessions experience a strong adverse effect
of marriage on their mortality hazards during ages 1947 years.
Their counterparts born during booms, on the other hand, experi-
ence an increase in hazards of death due to marriage for a much
shorter duration between ages 1629 years. These differences in
the statistical signicance of the marital impact on mortality risks
for the two groups emerge despite the periodic overlaps in their
condence intervals. So women born in favorable economic times
appear to cope better with problems during the childbearing ages
than women born in a year of recession (in Section 5.3 we provide
the implied quantitative estimates of average effects on longevity).
It is plausible that the increased mortality during childbearing
ages, as compared to the mortality when being single, reects the
 G.J. van den Berg, S. Gupta / Journal of Health Economics 40 (2015) 141158 151

Table 3
Parameter estimates for the full model with interactions between age and conditions at birth in the causal effect of marriage on the mortality rate (rst out of three parts of
the table).

Variable Full sample Men Women

Est. t-Stat. Est. t-Stat. Est. t-Stat.

Individual background characteristics affecting the marriage rate

Female 0.45 9.46
Social class father at birth 0.08 4.35 0.11 4.00 0.04 1.55
Father is literate 0.11 1.52 0.03 0.27 0.13 1.38
Born in urban area 0.07 1.30 0.09 1.15 0.20 2.53
Born in province Utrecht 0.04 0.75 0.12 1.42 0.02 0.27
Born in province Zeeland 0.17 3.24 0.10 1.27 0.24 3.40
Business cycle conditions early in life affecting the marriage rate
Boom (instead of recession) at birth 0.04 0.70 0.01 0.16 0.08 0.98
Cycle indicator for age 1 up to 6 0.01 0.19 0.12 1.50 0.03 0.46
Cycle indicator for age 7 up to 12 0.11 1.97 0.21 2.46 0.08 0.98
Cycle indicator for age 13 up to 15 0.04 0.22 0.14 0.54 0.28 1.29
Exposure to epidemics early in life affecting the marriage rate
1849 Cholera in Utrecht during age 16 0.12 0.09 0.05 0.02 0.08 0.05
1870/1 smallpox during age 16 0.89 1.98 0.12 0.21 1.82 2.77
1849 Cholera in Utrecht during age 712 2.37 1.58 1.90 0.93 2.98 1.26
1870/1 smallpox during age 712 0.03 0.06 0.11 0.15 0.02 0.02
1849 Cholera in Utrecht during age 1315 3.43 2.95 3.01 1.73 3.17 1.85
1870/1 smallpox during age 1315 1.15 2.44 0.42 0.63 1.27 2.02
Contemporaneous macro conditions affecting the marriage rate
1849 cholera in Utrecht 0.42 1.01 0.28 0.48 0.56 0.95
1870/1 smallpox 0.02 0.11 0.13 0.56 0.11 0.50
1918 inuenza 0.14 0.92 0.16 0.72 0.14 0.69
World War II (GNP missing) 3.12 4.07 2.34 2.19 4.45 3.84
Current Trend (log annual real p.c. GNP) 0.23 2.71 0.17 1.44 0.33 2.94
Current Cycle (log annual real p.c. GNP) 0.20 0.63 0.16 0.60 0.06 0.15
Age effect on marriage rate
m0
9.39 7.10 9.56 5.58 15.98 5.07
m1
8.88 5.81 8.05 4.40 19.41 4.69
m2
13.39 8.59 14.27 7.17 22.95 5.81
m3
3.83 4.51 3.63 3.46 9.04 4.25
m4
4.72 10.15 5.30 8.15 7.15 7.32
Unobserved heterogeneity terms for marriage
vm1
1.73 15.78 1.81 11.99 1.40 9.56
v2m 0.60 23.73 0.67 17.41 0.77 16.53
Unobserved heterogeneity terms for death
v1d 0.16 5.68 0.14 3.511 0.17 5.618
v2d 1.24 2.66 1.25 1.65 1.41 2.36
Individual background characteristics affecting the mortality rate
Female 0.16 3.96
Social class father at birth 0.00 0.17 0.02 0.68 0.02 0.86
Father is literate 0.02 0.31 0.07 0.79 0.05 0.54
Born in urban area 0.07 1.42 0.21 2.89 0.06 0.81
Born in province Utrecht 0.25 4.74 0.23 3.12 0.28 3.67
Born in province Zeeland 0.13 2.67 0.08 1.10 0.17 2.49
Business cycle conditions early in life affecting the mortality rate
Boom (instead of recession) at birth 0.12 1.63 0.32 3.08 0.17 1.40
Cycle indicator for age 1 up to 6 0.02 0.44 0.03 0.47 0.01 0.19
Cycle indicator for age 7 up to 12 0.14 2.74 0.17 2.36 0.09 1.24
Cycle indicator for age 13 up to 15 0.18 0.93 0.14 0.54 0.28 1.14
Exposure to epidemics early in life affecting the mortality rate
1849 Cholera in Utrecht during age 16 0.57 0.55 0.38 0.27 0.98 0.68
1870/1 smallpox during age 16 1.61 3.50 3.31 2.58 1.70 2.57
1849 Cholera in Utrecht during age 712 1.80 1.84 1.38 2.14 0.14 0.09
1870/1 smallpox during age 712 1.55 3.00 1.06 1.51 2.37 3.04
1849 Cholera in Utrecht during age 1315 0.88 1.30 1.20 1.03 0.96 1.12
1870/1 smallpox during age 1315 0.44 1.02 0.63 0.99 0.08 0.15
Contemporaneous macro conditions affecting the mortality rate
1849 cholera in Utrecht 1.14 2.52 1.55 3.07 0.31 0.31
1870/1 smallpox 0.57 2.41 0.35 0.97 0.74 2.37
1918 inuenza 0.08 0.40 0.02 0.08 0.12 0.42
World War II (GNP missing) 4.86 11.38 3.55 5.57 6.62 10.66
Current Trend (log annual real p.c. GNP) 0.60 11.98 0.45 6.05 0.80 10.93
Current Cycle (log annual real p.c. GNP) 0.28 1.13 0.00 0.00 0.31 1.03
Effect of marital status on the mortality rate
ape
0 0.16 0.84 0.18 0.50 0.33 1.38
ape
1 0.23 0.35 0.05 0.11 0.54 1.06
ape
2 0.06 1.17 0.06 0.88 0.08 1.03
ape
3 0.68 0.34 1.04 2.03 0.43 0.91
ape
4  0.04 0.77 0.09 1.07 0.02 0.26
Age boom at birth a,int
0
0.02 0.07 0.08 0.18 0.09 0.31
152 G.J. van den Berg, S. Gupta / Journal of Health Economics 40 (2015) 141158

Table 3 (Continued)

Variable Full sample Men Women

Est. t-Stat. Est. t-Stat. Est. t-Stat.


Age boom at birth a,int 0.18 0.60 0.12 0.19 0.23 0.64

1

Age boom at birth a,int 0.36 0.98 0.22 0.30 0.55 1.16
2 
Age boom at birth a,int 0.38 1.60 0.29 0.64 0.46 1.54
a,int
3 
Age boom at birth 4 0.23 1.43 0.18 0.68 0.30 1.33
Age effect on the mortality rate
d0 1.29 3.17 0.24 0.39 2.62 4.41
d1 1.94 24.65 1.83 14.19 2.13 19.94
d2 0.83 7.97 0.71 3.91 1.01 7.20
d3 0.05 0.73 0.12 1.09 0.03 0.31
d4 0.14 1.78 0.23 1.82 0.02 0.16
Joint probabilities of unobserved heterogeneities
q1 0.39 14.38 0.42 10.24 0.48 12.12
q2 0.41 15.69 0.40 11.26 0.33 9.36
q3 0.11 5.10 0.11 3.16 0.13 4.98
q4 0.09 4.28 0.07 2.53 0.07 3.38

Log likelihood 33,374.03 16,941.33 16,299.30

Number of individuals 5593 2709 2884

health hazards of pregnancy, delivery, childbearing, and taking care
of small children. Indeed, in the absence of advanced contracep-
tives, pregnancies would follow soon after marriage. Additional
risk factors for married women in childbearing ages are (i) a possi-
ble reduction of nutrition upon the arrival of children, as the total
available amount needs to be shared among a larger number of
household members, and (ii) an increased exposure to infectious
diseases brought into the household by small children. Investiga-
tion into the precise driving forces would require data on causes
of death (for instance during childbirth). Maternal mortality (i.e.
mortality at birth of a child) statistics are unavailable for the
Netherlands for our observation window. However, studies report
gures of maternal mortality in pre-industrial western societies
ranging up to 1600 deaths per 100,000 live births (De Brouwere
et al., 1998). Maternal mortality rates in the Netherlands can be
expected to be similar to those of Sweden (250300) and England
and Wales (400450) around 1870. Results from studies on trends
in age and gender specic mortality declines in the Netherlands
(for e.g. Wolleswinkel-Van den Bosch et al., 1998) nd a decline
in female mortality towards the end of our observation window
for the age interval of 2049. The authors link this to a decline in
marital fertility that started around the same time.
Our results are then consistent with the view that early-life con-
ditions inuence the resilience of women during the childbearing
stage of the life course. For lack of data, the epidemiological liter-
ature has not addressed the effect of early-life conditions on the
female mortality associated with pregnancy and delivery. How-
ever, a number of studies have demonstrated early-life effects on

Fig. 5. (a) Causal effect of marriage on the mortality rate interacted with economic conditions at birth, as a function of age, for women. (b) Causal effect of marriage on the
mortality rate interacted with economic conditions at birth, as a function of age, for men.
 G.J. van den Berg, S. Gupta / Journal of Health Economics 40 (2015) 141158
Fig. 6. Age dependence of marriage and death (baselines) by gender.
pregnancy complications and poor reproductive success, where the
data are collected retrospectively among those who survive the
delivery of their children. Lumey and Stein (1997) show that among
women who had been exposed to a famine in utero and who sur-
vived up to age 43, the offspring was more likely to suffer from
perinatal death. Innes et al. (1999), using birth weight as a proxy
for early-life conditions, observe an effect of low birth weight on
the risk of developing pre-eclampsia during pregnancy. The lat-
ter is a potentially deadly condition. Hackman et al. (1983) nd
an inverse association between the mothers birth weight and the
babys need for intensive care. See also the surveys by Lummaa
(2003) and Sloboda et al. (2011). The latter reports that women with
low birth weight show a marked reduction in ovarian and uter-
ine size in adolescence. Under the assumption that neonatal health
problems for the offspring and maternal health problems during
childbirth may have common determinants, these studies are of
relevance for us. Specically, they reinforce our interpretation of
the role of early-life conditions on health during the childbearing
stage of the life course.
Potentially, our empirical nding that early-life conditions inu-
ence the length of the age interval during which marriage is not
protective against mortality could also be attributed to an effect of
early-life conditions on the length of the fertility age interval (age
at menopause minus age at menarche). However, according to the
epidemiological literature, there is no clear association between
nutritional conditions early in life and the size of this interval or on
the age at menopause (see Sloboda et al., 2011, for a survey).
Notice that the long-run effect of early-life conditions on female
mortality is entirely driven through marriage, in the following
sense. Early-life conditions affect the extent to which marriage is
protective for mortality. However, they do not affect the mortality
rate if marriage does not take place in ones lifetime. It is impor-
tant to note that our results are not contradictory to results in some
earlier studies. For instance Osmond et al. (1993), focusing on asso-
ciations between death rates and weight at birth and end of rst
153
year of life, also nd signicant gender differences. In a Finnish sam-
ple born between 1911 and 1930, they nd a signicant negative
association between birth weight and weight at 1 year and all-cause
mortality for men. However, no evidence of a similar signicant
trend is found for women, the only exception being death due to
cardiovascular disease which was negatively correlated with birth
weight and weight at age 1 year. Moreover, men and women also
differ in how conditions at birth interact with conditions thereafter
to jointly inuence adult mortality. Considering the simultane-
ous correlation between birth weight and weight at year 1 with
death due to cardiovascular disease, they nd that women who
are underweight at birth but above average in weight at year 1
(in our case married in adulthood) are at a higher risk of cardio-
vascular related death. For men on the other hand, below average
weight at birth and at the end of year 1 (unmarried as adults in our
study) jointly is associated to death due to cardiovascular disease.
In a subsequent study (Eriksson et al., 2009), exploring underlying
mechanisms for such observed gender differences in the impact
of early-life conditions on adult mortality, the authors nd that
intra-uterine growth strategies differ signicantly between men
and women. Poor nutrition of the mother during pregnancy results
in low birth weight of her child. However, for a given low birth
weight male placentas are smaller than of their female counter-
parts. Smaller placentas more frequently result in altered placental
diameters in male fetus. Such compensatory expansions in boys,
while beneting brain development, is at the cost of other organs
like the kidneys. The authors suggest that in its extreme form this
dangerous growth strategy of boys in response to adverse uterine
conditions may explain the higher prevalence of hypertension and
shorter life expectancies of men compared to women. Our result
that early-life conditions do not directly affect female adult all-
cause mortality reiterates these ndings. Furthermore, our ndings
shed light on how adverse early-life conditions may still indirectly
lower female life expectancy by interacting with later life events
like marriage.
154 G.J. van den Berg, S. Gupta / Journal of Health Economics 40 (2015) 141158
At the same time, female marriage rates are unaffected by con-
ditions in the years of birth and early childhood (early childhood
being ages 16 years, which the medical literature identies as cru-
cial for long term development; see e.g. Power et al., 2003). This
shows that any strategic considerations to enter or postpone mar-
riage in response to favorable or adverse economic conditions early
in life do not play a major role as determinants of marriage. At the
very least, any such behavior does not result in visible effects on the
marriage rates. Interestingly, the aforementioned study of Lumey
and Stein (1997) nds that age at rst delivery does not depend on
whether the woman was in-utero during a famine or not.
Turning now to men (Fig. 6b), we nd a protective causal impact
of marriage on the mortality rate. The effect is signicant over the
typical ages at death of about 5685. This nding is intuitive in
light of social observations like growing loneliness owing to shrink-
ing social circles for single individuals. Marriage on the other hand
could offer support from a wife during older ages and improve qual-
ity of life by means of better housekeeping and personal care. The
HSN does not record the time of death of spouses of the individu-
als in the sample. The absence of a consistently positive protective
effect of marriage beyond the ages of 85 years probably reects a
limitation of our polynomial specication. We explore the role of
marital transitions further in Section 5.2.
As for the role of conditions in early years of life, in contrast
to women, for men we nd no signicant effect of being born in
a boom as opposed to being born in a recession on the protective
effect of marriage on the mortality rate. Though men born dur-
ing adverse economic conditions seem to be enjoying the positive
marital effect for marginally longer than married men born during
favorable conditions, this difference is not statistically signicant.
Of course, men born under favorable conditions live on average
longer (we discuss this in the next paragraph). Finally, for both
men and women our results again show that marital status does
not have a time constant impact on the mortality rate throughout
life.
Considering the direct impact of early life conditions on mortal-
ity, we nd a signicantly lower mortality rate amongst men born
during a boom. This reiterates the results of Van den Berg et al.
(2006) who estimate mortality models without including marital
status as an explanatory or intermediate variable. Another striking
result is a signicant negative effect of average cyclical conditions
during the age interval of 712 years on the marriage as well as
the mortality rate of men. We postulate two explanations for this.
The rst one concerns education. The age interval 712 coincided
with primary school in the Netherlands, which prior to 1901 was
not compulsory and only free for the very poor. Many individu-
als did not get more formal education than this. Better economic
conditions at these ages could facilitate the entry into secondary
education; moreover it could lead to larger professional involve-
ment and consequent delays in marriage. In recessions, nancial
adversity could increase temporary absences from school affecting
negatively at least the quality of education (Dunn et al., 2003).
Better educational opportunities could also explain the signi-
cant negative impact of favorable conditions at ages 712 on male
mortality. Studies such as Cutler and Lleras-Muney (2006) and
Maccini and Yang (2009), using more contemporary data, nd a
mediating effect of education on the mortality rate for both gen-
ders. The fact that we do not nd a signicant effect of average
cyclical conditions during the age interval of 712 years on the
marriage and mortality rates for women is not inconsistent with
this. School attendance of women, in our historical sample, was
often lower and shorter than for men, and career concerns were
mostly absent.
The second explanation, which is especially relevant for the
direct effect of conditions at ages 712 on male mortality, is that
these ages include a critical or sensitive period in boys develop-
ment towards their adult health status. Specically, Van den Berg
et al. (2014) identify such a period at around age 9 in the sense that
if conditions at this age are adverse then adult height is lower.
Class differences are very important for nuptiality. Male mem-
bers of the three lower social classes exhibit a much larger marriage
rate relative to the upper classes. This is in line with past obser-
vations of more frequent and younger marriages for the lower
social classes resulting from increased economic opportunities fol-
lowing the industrial revolution (see the literature mentioned in
Section 3). Possibilities of social class upward mobility by means
of marriage may also arise for women. Investigation into marriage
market prospects is left to future work.
Finally, considering age dependence of exit rates into marriage
and mortality we nd the expected inverted U-shape for the former
and a monotonically increasing one for the latter (see Fig. 6).17 We
observe that the marriage rate increases until the age of 32 for men
and 29 for women after which it consistently declines though at a
slower rate after the mid 40s. This sudden slow-down in the declin-
ing marriage rate could indicate second marriages. In the absence of
information about multiple marriages and continued marital status
of individuals we are unable to fully comment on this. However, in
the following section we try to explore the impact of higher order
marriages.
5.2. Sensitivity analysis
Under the current specication, marriage is a fully absorbing
state i.e. once married the individual remains in the state of mar-
riage until death. In principle, this assumption can be violated for
a number of reasons and is maintained entirely due to lack of data
on marital transitions. Re-marriage may also end divorce or wid-
owhood, and in either case there may be a subsequent exit from
the state of marriage due to mortality, divorce or widowhood, etc..
As mentioned in Section 3.2, even though divorce was relatively
uncommon during our observation window, widowhood was fre-
quently followed by re-marriage except at high ages.
We explore the relevance of these marital transitions, using
a different nationally representative historical dataset which
also falls under the umbrella of the Historical Sample of the
Netherlands. This is the dataset release 2001.01 Beta. This data
is also put together from administrative records. Its main advan-
tage is that it includes all marital transitions at the individual level.
Its main disadvantage is that its observation window is on aver-
age more recent than in the data used in the paper. Specically,
the dataset only includes individuals born in 18501922. This only
partially overlaps with our sample of cohorts born in 18151902.
Relatively more observations are right-censored due to individuals
lifetimes still being ongoing. Additionally, the emergence of better
social security nets in recent years implies that these individuals
are on average better insured against macroeconomic shocks than
those in the original sample. This could weaken business cycles in
early years of life as indicators of economic conditions during these
years. The new sample consists of 1416 men and 1486 women.
We re-analyzed our model with the 2001.01 Beta release data,
now estimating the transition rates from marriage to widowhood
and from widowhood into remarriage. In this model, the mortality
rate is allowed to depend on whether the individual is currently
17
For sake of completeness we depict the age dependence over the full age range
as observed in the data. However, no unexpected patterns are observed in marriage
and mortality baselines when shorter age intervals are considered. Note that the
scales of values on the vertical axes are determined by normalizations. Hence, the
values as such are not informative.
 G.J. van den Berg, S. Gupta / Journal of Health Economics 40 (2015) 141158
married or not, where the latter can vary over time as the individual
status may switch back and forth between being married or not.
Even after controlling for all marital transitions, our main results
remain the same.18
The last robustness check we perform to see how sensitive our
results are to the denition of the marital effect, is the re-estimation
of our model while right-censoring longevity and marriage dura-
tions at age 70. Once again, we nd that our results remain
qualitatively unchanged.
We also estimate several alternative specications of the causal
effect of marriage on the individual mortality rate. First, we try a
simpler model with only the age dependent marital effect on mor-
tality without the interactions with the dummy for being born in
a boom or not. Another specication replaces the age of the mar-
ried individual with the number of years married as a determinant
of the causal effect of marriage on mortality. This specication is
tested with and without additional interaction terms between early
life conditions and the number of years married. Finally, following
past literature we estimate a basic model where the causal effect of
marriage is specied as a time varying regressor i.e. takes a value of
0 before the person is married and 1 after his marriage. Our results
are robust to all these alternative specications. Tests conrm that
our full model as estimated above is a better model specication.
For sake of brevity, detailed discussion of these alternative speci-
cations and full parameter estimates are not presented here but
can be requested from the authors.
In the Online Appendix we also consider alternative model spec-
ications in which the business cycle indicators are actual average
deviations per age interval instead of binary variables. Again, the
results are robust to this. As noted in Section 4.5, we also estimate
models while excluding the parents in the highest social class. For
ease of comparison with our main results we report results aver-
aged over the other explanatory variables. Looking separately at
men and women born during economic recessions and booms we
nd that all our results essentially remain the same.
5.3. Additional discussion and implications
Using our parameter estimates in Table 2 and the actual ages of
marriage of the individuals in the sample we nd that on average19
marriage implies a 2.16% decrease in longevity for women. How-
ever, once we control for business cycle conditions in the year of
birth these gures look very different. For women born during an
economic boom, marriage in fact has favorable impact on their life
expectancy (0.61% increase in expected lifetime). Therefore, the
result of an adverse effect of marriage on the mortality rate for
women aged less than 52 years is driven only by women born
during economic downswings. For this group marriage reduces
expected longevity by 5.40%. For men, marriage leads to a 4.77%
increase in life expectancy on an average with no signicant dif-
ference between those born during economic boom versus births
during economic recessions.
Given the non-constant protective effect of marriage over an
individuals lifetime, it is also useful to consider by how much life
expectancy changes if you marry at any given age. Fig. 7 presents
the change in life expectancy as a consequence of marriage for every
possible age of being married (16103 years), calculated using the
parameter estimates from our model and averaging over all other
explanatory variables. Results are once again presented separately
18
For sake of brevity, results for this subsection are not presented here but are
available in the Online Appendix of this paper. The latter includes details on data
and models used here.
19
Averaging over individuals marrying at various ages.
155
by gender and economic boom or not in the year of birth. Getting
married at young ages (below about 25 years) shortens longevity
regardless of early-life conditions, as compared to staying single.
This is most likely because marriage at young ages increases the
length of the childbearing age interval. The adverse effect reduces
with increasing age at marriage. Married teenagers are in the worst
position. Beyond the mid twenties the adverse effect of marriage
on longevity only holds for women born in recessions. This nding
suggests that women who might have already suffered a health set-
back early in life are less able to cope with future health strains. For
men on the other hand, the protective effect is generally higher the
younger they marry. We cannot identify the reasons for this result
from the data we have. However, early onset of healthier lifestyle
and safe habits are plausible causes. Finally, once again we see that
the role of early life conditions via marriage is less relevant for men.
The presence of adverse health consequences of marriage for
women raises the question why women choose marriage, espe-
cially since decisions involve expected present values of various
options, and any long-run benets of marriage are discounted more
heavily than the short-run costs. A similar concern has arisen in
the family pay gap literature that nds large wage penalties for
women choosing to have children. It appears that not only do
women planning to have children self select themselves into lower
earning occupations and jobs, they also accept large wage losses
on return to labor market after child-related employment breaks
(see e.g. Beblo et al., 2006). Recent social-psychological literature
on female well-being provides some insight into the forces driving
these choices. Abele-Brehm (2006) in a recent study using longi-
tudinal data on about 2000 individuals nds higher levels of life
satisfaction amongst women living with partners, with women
on maternal leave being especially satised. This eld of research
seems to indicate that while women pay a large cost in terms of
income and even health by choosing to have a family, family life is
crucially important for their happiness making having a family
the preferred choice for a lot of women. Wong (2007) structurally
estimates an equilibrium matching model of the marriage mar-
ket and the labor market, and she concludes that women derive
a large non-pecuniary utility ow from being married and having
children. Without such non-pecuniary utility, the t of the model
is signicantly worse. Of course, evolutionary theory provides a
more fundamental explanation for why women prefer to engage
in an institution such as marriage that provides a setting within
which to obtain and raise offspring.
Several obvious similarities exist between 19th century Dutch
society and the current-day developing world. Therefore results
drawn from this study could help policy makers in less developed
economies in their struggle against child deprivation and high mor-
tality rates (see De Brouwere et al., 1998). Supporting children born
in adverse economic times, by means of extra provisions of food,
housing and health care, will not only help decrease infant and child
mortality but in fact increase over all male life expectancy. Special
care could be taken in supporting female children born in economic
downturns for whom health consequences of adverse early eco-
nomic conditions are indirect (via marriage) and not immediately
visible. In societies where the girl child is frequently considered
only secondary to her male siblings, in absence of immediate con-
sequences of female child deprivation, her needs are likely to be
easily ignored during poor economic times leading to life long
adverse health consequences. Further support could be provided
to women during their child bearing ages. This could be done by
discouraging marriages of very young women especially in rural
areas or within lower social classes and religious communities by
setting and enforcing suitable legal age of marriage. Additionally,
access to modern contraceptives could help curb quick succes-
sive pregnancies that lead to high infant and maternal mortality
156 G.J. van den Berg, S. Gupta / Journal of Health Economics 40 (2015) 141158

Fig. 7. Expected changes in life expectancy as a causal effect of marriage at different ages of getting married.

rates. Active family planning programs could also allow couples to
enjoy the benets of partnership without suffering from its nega-
tive health consequences. Over a longer time horizon, urbanization,
economic development of the country (reected in high per capital
real GNP) and more education would help discourage early mar-
riages and consequently its adverse consequences, especially for
women.
6. Conclusion
The empirical analyses with historical data generate a range of
insights into causal pathways between economic conditions early
in life, marriage, and mortality. These are identied while taking
account of selection effects due to unobserved confounders. The
results are markedly different between men and women.
For women, the extent to which marriage protects against mor-
tality depends on the economic conditions early in life. Birth under
adverse conditions implies that longevity is reduced upon mar-
riage, as compared to staying unmarried. Conversely, birth under
favorable conditions implies that longevity is extended upon mar-
riage. In each case, the effect of marital status on mortality is not
uniform as a function of age. The health costs of marriage are borne
early on, whereas the protective benets occur after the childbear-
ing ages. But for women born under favorable conditions, the costs
 G.J. van den Berg, S. Gupta / Journal of Health Economics 40 (2015) 141158
during childbearing ages are smaller, and the age at which benets
outweigh costs occurs earlier in life.
We provide a range of reasons for why the health costs of mar-
riage during childbearing ages are due to the health hazards of
pregnancy, delivery, childbearing, and taking care of small children.
Our results are consistent with the view that early-life conditions
inuence the resilience of women during this stage of the life
course. In current times, in Western societies, maternal death dur-
ing deliveries and during childbearing ages in general is very rare.
However, it is plausible that the various stages of childbearing are
still a health burden for the mother, and it is obviously useful to
know that this burden is larger in case the mother was exposed
to adverse conditions in her own birth year. As we have seen, in
poor countries, maternal mortality is still the dominant death cause
during reproductive years.
The analysis for women also shows that early-life conditions do
not inuence the marriage rate. This shows that strategic consid-
erations to enter or postpone marriage in response to favorable or
adverse economic conditions early in life do not play a major role as
determinants of marriage. The decision to marry is driven by other
behavioral motivations.
For men the results are different. Marriage is protective at all
ages, and the size of the protective effect on the mortality rate does
not depend on conditions early in life. At the same time, conditions
around birth have an autonomous long-run effect on mortality.
For men, conditions around the ages 7 to 12 are important for the
marriage rate as well as the mortality rate. Presumably, these are
the ages in our sample during which decisions are made regarding
education and career. Favorable conditions at these ages lead to a
postponement of marriage and to lower mortality later in life.
The long-run causal pathway from early-life economic condi-
tions to mortality late in life differs by gender. For women, the
long-run effect of early-life conditions on mortality is driven 100%
through marriage. Early-life conditions affect the extent to which
marriage is protective for mortality. However, they do not affect
the mortality rate if marriage does not take place in ones lifetime.
For men, the long-run effect is present regardless of the marital
status during ones adult life. Its size is not affected by marriage at
all.
Acknowledgements
We are grateful to the Editor Richard Frank, an anonymous
Referee, Andrew Chesher, James Banks, Marcus Eliason, Michle
Tertilt, Ronald Wolthoff, Marta Lpez, Erik Plug, Maarten Lin-
deboom, Kees Mandemakers, and Manuel Arellano, for valuable
comments on earlier versions. We also thank participants of PAA,
ESEM, and ESPE conferences, the IZA Summer School, the Micro-
data meeting at CEMFII, a gender conference at IAB/University of
Nrnberg, seminars at the Tinbergen Institute Amsterdam and UCL
London, and workshops of the Max Planck Institute for Demo-
graphic Research on Causal inference and endogeneity in event
history analysis and Early-life vs. later-life effects on late-life
mortality for their useful input. The Historical Sample of the
Netherlands Dataset Release UZF.02 was kindly provided by the
International Institute of Social History (IISG), Amsterdam. The
authors acknowledge nancial support from the EU.
Appendix A. Supplementary data
Supplementary data associated with this article can be found,
in the online version, at http://dx.doi.org/10.1016/j.jhealeco.
2014.02.004.
157
References
Abbring, J.H., Van den Berg, G.J., 2003. The non-parametric identication of treat-
ment effects in duration models. Econometrica 71, 14911517.
Abele-Brehm, A., 2006. Who is happiest? Career and family as determinants of pro-
fessionals life satisfaction in the Life Rush Hour between thirty and forty:
Findings from the Erlangen longitudinal study.
Almond, D., 2006. Is the 1918 inuenza pandemic over? Long-term effects of in
utero inuenza exposure in the post-1940 U.S. population. Journal of Political
Economy 114 (4), 672712.
Almond, D., Mazumder, B., 2006. The 1918 inuenza pandemic and subsequent
health outcomes: an analysis of SIPP data. The American Economic Review:
Papers and Proceedings 95 (2), 258262.
Beblo, M., Bender, S., Wolf, E., 2006. The wage effects of entering motherhood a
within-rm matching approach. ZEW Discussion Paper No. 06-053.
Bengtsson, T., Lindstrm, M., 2003. Airborne infectious diseases during infancy and
mortality in later life in southern Sweden, 17661894. International Journal of
Epidemiology 32, 286294.
Blau, F.D., Kahn, L.M., Waldfogel, J., 2000. Understanding young womens marriage
decisions: the role of labor and marriage market conditions. Industrial and Labor
Relations Review 53, 624647.
Brandt, L., Siow, A., Vogel, C., 2008. Large shocks and small changes in the marriage
market for famine born cohorts in China. Working paper University of Toronto
Department of Economics, No. 334.
Case, A., Fertig, A., Paxson, C., 2003. From cradle to grave? The lasting impact of
childhood health and circumstance. Working paper, NBER and Princeton Uni-
versity.
Cutler, D.M., Lleras-Muney, A., 2006. Education and Health: Evaluating Theories and
Evidence. NBER Working Papers 12352. National Bureau of Economic Research,
Inc.
Dasgupta, P., 2011. Personal histories and poverty traps. In: Lin, J., Plescovik, B. (Eds.),
Proceedings of the Annual World Bank Conference on Development Economics.
World Bank, Washington, DC.
De Brouwere, V., Tonglet, R., Van Lerberghe, W., 1998. Strategies for reducing mater-
nal mortality in developing countries: what can we learn from the history of the
industrialized West? Tropical Medicine & International Health 3 (10), 771782.
Dehejia, R., LLeras-Muney, A., 2004. Booms, busts, and babies health. The Quarterly
Journal of Economics 119 (3), 10911130.
Doblhammer, G., 2004. The late life legacy of very early life. Springer, Berlin.
Dribe, M., Lundh, C., 2010. Marriage choices and social reproduction: the inter-
relationship between partner selection and intergenerational socioeconomic
mobility in 19th-century Sweden. Demographic Research 22, 347382.
Dunn, M.C., Kadane, J.B., Garrow, J.R., 2003. Comparing harm done by mobility and
class absence: missing students and missing data. Journal of Educational and
Behavioral Statistics 28 (3), 269288.
Eriksson, J.G., Kajantie, E., Osmond, C., Thornburg, K., Barker, D.J.P., 2009. Boys live
dangerously in the womb. American Journal of Human Biology 22, 330335.
Farah, M.J., Noble, K.G., Hurt, H., 2008. Poverty, privilege, and brain development:
empirical ndings and ethical implications. Working paper. University of Penn-
sylvania.
Fogel, R.W., 1997. New ndings on secular trends in nutrition and mortality: some
implications for population theory. In: Rosenzweig, M.R., Stark, O. (Eds.), Hand-
book of Population and Family Economics. North-Holland, Amsterdam.
Gardner, J., Oswald, A.J., 2004. How is mortality affected by money, marriage and
stress? Journal of Health Economics 23 (6), 11811207.
Goldman, N., 1993a. Marriage selection and mortality patterns: inferences and fal-
lacies. Demography 30, 189208.
Goldman, N., 1993b. The perils of single life in contemporary Japan. Journal of Mar-
riage and the Family 55, 191204.
Hackman, E., Emanuel, I., Van Belle, G., Daling, J., 1983. Maternal birth weight and
subsequent pregnancy outcome. Journal of the American Medical Association
250, 20162019.
Hartman, M.S., 2004. The Household and the Making of History: A Subversive View
of the Western Past. Cambridge University Press, Cambridge.
Holemans, K., Caluwaerts, S., Andr Van Assche, F., 2002. Unravelling the fetal origins
hypothesis. Lancet 360, 2073.
Hu, Y., Goldman, N., 1990. Mortality differentials by marital status: an international
comparison. Demography 27, 233250.
Innes, K.E., Marshall, J.A., Byers, T.E., Calonge, N., 1999. A womans own birth weight
and gestational age predict her later risk of developing preeclampsia, a precursor
of chronic disease. Epidemiology 10, 153160.
Jacobs, J., Smits, J.-P., 2001. Business cycles in the Netherlands, 18151913. Working
paper. Groningen University.
Johnson, N.J., Backlund, E., Sorlie, P.D., Loveless, C.A., 2000. Marital status and mor-
tality: the national longitudinal mortality study. Annals of Epidemiology 10 (4),
224238.
Kreholt, I., 2001. The long shadow of socioeconomic conditions in childhood: do
they affect class inequalities in mortality? In: Jonsson, J.O., Mills, C. (Eds.), Cradle
to Grave. Life Course Change in Modern Sweden. Sociologypress, York.
Lillard, L.A., Panis, C.W.A., 1996. Marital status and mortality: the role of health.
Demography 33 (3), 313327.
Lindeboom, M., Portrait, F., Van den Berg, G.J., 2010. Long-run effects on longevity of
a nutritional shock early in life: the Dutch Potato famine of 18461847. Journal
of Health Economics 29 (5), 617629.
158 G.J. van den Berg, S. Gupta / Journal of Health Economics 40 (2015) 141158
Lumey, L.H., Stein, A.D., 1997. In utero exposure to famine and subsequent fertil-
ity: the Dutch famine birth cohort study. American Journal of Public Health 87,
19621966.
Lummaa, V., 2003. Early developmental conditions and reproductive success in
humans: downstream effects of prenatal famine, birthweight, and timing of
birth. American Journal of Human Biology 15, 370379.
Lund, R., Holstein, B.E., Osler, M., 2004. Marital history from age 15 to 40 years and
subsequent 10-years mortality: a longitudinal study of Danish males born in
1953. International Journal of Epidemiology 33 (2), 389397.
Maccini, S., Yang, D., 2009. Under the weather: health, schooling, and economic con-
sequences of early-life rainfall. American Economic Review 99 (3), 10061026.
Mandemakers, K., 2000. The Netherlands. Historical sample of the Netherlands.
In: Kelly Hall, P., McCaa, R., Thorvaldsen, G. (Eds.), Handbook of International
Historical Microdata for Population Research. Minnesota Population Center
Minneapolis, pp. 149177.
Matthijs, K., 2003. Demographic and sociological indicators of privatisation of
marriage in the 19th century in Flanders. European Journal of Population 19,
375412.
Modin, B., 2003. Born out of wedlock and never married it breaks a mans heart.
Social Science and Medicine 57, 487501.
Modin, B., Koupil, I., Vger, D., 2009. The impact of early twentieth century ille-
gitimacy across three generations. Longevity and intergenerational correlates.
Social Science and Medicine 68, 16331640.
Mogren, I., Hgberg, U., Stegmayr, B., Lindahl, B., Stenlund, H., 2001. Fetal expo-
sure heredity and risk indicators for cardiovascular disease in a Swedish welfare
cohort. International Journal of Epidemiology 30, 853862.
Murray, J.E., 2000. Marital protection and marital selection: evidence from a
historical-prospective sample of american men. Demography 37, 511521.
Osmond, C., Barker, D.J.P., Winter, P.D., Fall, C.H.D., Simmonds, S.J., 1993. Early
growth and death from cardiovascular disease in women. British Medical Journal
307, 15191524.
Petersen, W., 1960. The demographic transition in the Netherlands. American Soci-
ological Review 25 (3), 334347.
Phillips, D.I.W., Handelsman, D.J., Eriksson, J.G., Forsen, T., Osmond, C., Barker, D.J.P.,
2001. Prenatal growth and subsequent marital status: longitudinal study. British
Medical Journal 322, 771.
Power, C., Li, L., Manor, O., Davey Smith, G., 2003. Combination of low birth weight
and high adult body mass index: at what age is it established and what are
its determinants? Journal of Epidemiology and Community Health 57, 969
973.
Richardson, K., Van den Berg, G.J., 2013. Duration dependence versus unobserved
heterogeneity in treatment effects: Swedish labor market training and the tran-
sition rate to employment. Journal of Applied Econometrics 28, 325351.
Sastry, N., 2004. Trends in socioeconomic inequalities in mortality in develop-
ing countries: the case of child survival in Sao Paulo, Brazil. Demography 41,
443464.
Sloboda, D.M., Hickey, M., Hart, R., 2011. Reproduction in females: the role of the
early life environment. Human Reproduction 17, 210227.
Stein, F., Susser, Z., Saenger, M., Marolla, G., 1975. Famine and Human Development.
Oxford University Press, The Dutch Hunger Winter of 1944/45.
Vger, D., Modin, B., 2002. Prenatal growth, subsequent marital status, and mor-
tality: longitudinal study. British Medical Journal 324, 398.
Van den Berg, G.J., Doblhammer, G., Christensen, K., 2011. Being born under adverse
economic conditions leads to a higher cardiovascular mortality rate later in life:
evidence based on individuals born at different stages of the business cycle.
Demography 48, 507530.
Van den Berg, G.J., Lindeboom, M., 2014. Aging: conditions early in life and health
and mortality at advanced ages. In: Culyer, T. (Ed.), Encyclopaedia of Health
Economics. North-Holland, Amsterdam.
Van den Berg, G.J., Lindeboom, M., Lpez, M., 2009. Inequality in individual mortal-
ity and economic conditions earlier in life. Social Science and Medicine 69 (9),
13601367.
Van den Berg, G.J., Lindeboom, M., Portrait, F., 2006. Economic conditions
early in life and individual mortality. American Economic Review 96 (1),
290302.
Van den Berg, G.J., Lundborg, P., Nystedt, P., Rooth, D.-O., 2014. Critical periods during
childhood and adolescence: a study of adult height among immigrant siblings.
Journal of the European Economic Association, forthcoming.
Van Poppel, F., 1992. Trouwen in Nederland: Een Historische-Demograsche Studie
van de 19e en Vroeg-20e Eeuw. Working paper. NIDI, The Hague.
Van Poppel, F., 1997. Family breakdown in nineteenth-century Netherlands: divorc-
ing couples in The Hague. History of the Family 2 (1), 4972.
Van Poppel, F., 1998. Nineteenth-century remarriage patterns in the Netherlands.
Journal of Interdisciplinary History XXVIII (3), 343383.
Van Poppel, F., Joung, I., 2001. Long-term trends in marital status mortality differ-
ences in the Netherlands 18501970. Journal of Biosocial Sciences 33, 279303.
Van Tulder, J.J.M., 1962. De beroepsmobiliteit in Nederland van 1919 tot 1954. In:
Een sociaal-statistische studie. Stenfert Kroese, Leiden.
Woitek, U., 2003. Height cycles in the 18th and 19th centuries. Economics and
Human Biology 1, 243257.
Wolleswinkel-Van den Bosch, J.H., Van Poppel, F., Tabeau, E., Mackenbach, J.P., 1998.
Mortality decline in the Netherlands in the period 18501992: a turning point
analysis. Social Science and Medicine 47 (4), 429443.
Wong, L.Y., 2007. Labor market development and marriage. Working paper. Bing-
hamton University.
World Bank, 2010. World Development Indicators. World Bank, Washington, DC.