EconometriaAplic - Aula6 (Diferencas em Diferencas) Ferraz PUC

Econometria Aplicada
Aula 6: Diferenas-em-Diferenas e
Mtodo de Controle Sinttico
Claudio Ferraz
PUC-Rio
PUC-Rio
Abril 2017
Dados em painel Extensoes do LATE
Usando variacao temporal
Suponha que estamos interessados em estimar o efeito de um

programa (Di ) usando:
Yi = + Di + Xi0 + i
Professor Claudio Ferraz (1o semestre 2009)Econometria AplicadaDados em Painel, Efeito-Fixo, e Dif-em-Dif Aula 10 3/9

Yi = + Di + Xi0 + i
Na impossibilidade de obter variacao aleatorizada em Di e sabendo

que ha caractersticas nao observaveis que afetam tanto Di quanto
Yi , usamos uma variavel instrumental zi para identificar .

Yi = + Di + Xi0 + i

Na pratica, encontrar uma variavel instrumental valida nao e uma

tarefa facil. O que mais podemos fazer caso nao tenhamos um
instrumento?

Yi = + Di + Xi0 + i

Na pratica, encontrar uma variavel instrumental valida nao e uma

tarefa facil. O que mais podemos fazer caso nao tenhamos um
instrumento?
Nestas aulas iremos explorar variacao temporal, ou entre coortes, para

identificar quando a variavel nao observavel esta fixa no tempo.
Efeitos fixos individuais
Comecamos com um exemplo. Trabalhadores que tem seus salarios

determinados por um processo de barganha sindical ganham mais ou
eles tem caractersticas que os fariam ganhar mais de qualquer
maneira?

maneira?
Seja Yit Log da renda e Dit um indicador que e igual a 1 se a pessoal

pertence a um sindicato. Para cada pessoa i observamos somente
Y0it ou Y1it .

maneira?
Seja Yit Log da renda e Dit um indicador que e igual a 1 se a pessoal

pertence a um sindicato. Para cada pessoa i observamos somente
Y0it ou Y1it .
Temos um vetor de variaveis Ai (habilidade) que nao observamos,

mas que esta fixa no tempo. Vamos assumir que:
E [Y0it |Ai , Xit , t, Dit ] = E [Y0it |Ait , Xit , t]
Podemos escrever o resultado de Y0it como:
E [Y0it |Ai , Xit , t] = + t + A0i + Xit0
Por ultimo, iremos assumir que o efeito causal e aditivo e constante:
E [Y1it |Ai , Xit , t] = E [Y0it |Ai , Xit , t] +
Estes supostos implicam num modelo dado por:
E [Yit |Ai , Xit , t, Dit ] = + t + Dit + A0i + Xit0
Estes supostos implicam num modelo dado por:
E [Yit |Ai , Xit , t, Dit ] = + t + Dit + A0i + Xit0
E uma equacao estimada na forma:
Yit = i + t + Dit + Xit0 + it
onde i = + A0i .

O modelo de efeitos fixos pode ser estimado colocando uma dummy
para cada observacao i e cada observacao t. Isso pode ser muito
intensivo computacionalmente quando i e grande.

Uma alternativa, que nos da um coeficiente identico, e utilizar desvios

de medias estimando:
(Yit Yi ) = (t t ) + (Dit Di ) + (Xit Xi )0 + (it i )


Uma terceira alternativa para a eliminacao do efeito-fixo i e tirar a

primeira diferenca:
Yit = t + Dit + Xit + it


Uma terceira alternativa para a eliminacao do efeito-fixo i e tirar a

primeira diferenca:
Yit = t + Dit + Xit + it
Com dois perodos todos estimadores sao iguais. Com t > 2, desvio
de medias e mais eficiente. Tomar cuidado com erro padrao quando
utilizar diferencas.
Efeitos fixos e erro de medida
Apesar de controlar por caractersticas nao-observaveis fixas, o

estimador de EF esta muito sujeito a atenuacao por eero de medida.

A utilizacao de EF para estimar implica que a identificacao esta

vindo de variacao do mesmo i ao longo do tempo. Ou seja, variacoes
em Dit .


em Dit .
Se na cross-section temos erro de medida ou de codificacao (Di = 1

quando deveria ser 0), este problema e magnificado quando
estimamos EF ja que a variacao usada para identificacao pode ser
rudo puro. Coeficientes estimados com EF serao menores.


em Dit .
Se na cross-section temos erro de medida ou de codificacao (Di = 1

quando deveria ser 0), este problema e magnificado quando
estimamos EF ja que a variacao usada para identificacao pode ser
rudo puro. Coeficientes estimados com EF serao menores.
O que podemos fazer nestes casos? Usar IV ou utilizar informacao

externa para corrigir o erro de medida (correlated random effects). FE
elimina variacao ruim e boa.
Water for Life: The Impact of the Privatization of
Water Services on Child Mortality
Sebastian Galiani
Universidad de San Andres
Paul Gertler
University of California, Berkeley and National Bureau of Economic Research
Ernesto Schargrodsky
Universidad Torcuato di Tella
While most countries are committed to increasing access to safe water

and thereby reducing child mortality, there is little consensus on how
to actually improve water services. One important proposal under
discussion is whether to privatize water provision. In the 1990s Ar-
gentina embarked on one of the largest privatization campaigns in
the world, including the privatization of local water companies cov-
ering approximately 30 percent of the countrys municipalities. Using
the variation in ownership of water provision across time and space
generated by the privatization process, we find that child mortality
fell 8 percent in the areas that privatized their water services and that
the effect was largest (26 percent) in the poorest areas. We check the
robustness of these estimates using cause-specific mortality. While pri-
vatization is associated with significant reductions in deaths from in-
fectious and parasitic diseases, it is uncorrelated with deaths from
causes unrelated to water conditions.
This paper has benefited from comments by Joshua Angrist, John Cochrane, Alberto
Chong, Jonathan Gruber, David Levine, Steve Levitt, Florencio Lopez-de-Silanes, Sebastian
Martinez, Ted Miguel, Rohini Pande, Manisha Shah, Pablo Spiller, and Maximo Torero.
We are also grateful to Pablo Cafiero, MD, Liliana Clara, MD, and Enrique Calderon, MD,
who provided expert information on water-related diseases in Argentina, and to Matias
Cattaneo and Juan Pantano, who provided excellent research assistance.
[Journal of Political Economy, 2005, vol. 113, no. 1]

2005 by The University of Chicago. All rights reserved. 0022-3808/2005/11301-0006$10.00
83
Fig. 1.Evolution of mortality rates for municipalities with privatized vs. nonprivatized water services
water for life 89
TABLE 1
Change in Ownership of Water Systems, 199099
Number of
Ownership Municipalities Percentage
Always public 196 39.7
Always private not-for-profit cooperative 143 28.9
Transferred from public to private for-profit 137 27.7
Always private for-profit 1 .2
No service or missing information 17 3.4
Total 494 100.0
Note.In municipalities in which more than one company provides water services, we defined the ownership status
of the municipality as the ownership of the company supplying the largest fraction of the population. Source: SPIDES,
ENOHSA.
provided services in the remaining one-third. Between 1991 and 1999,

about half of the public water companies servicing 28 percent of the
countrys municipalities and covering almost 60 percent of the countrys
population were transferred to private for-profit control (see table 1).
The remaining municipalities continued receiving water services from
either public companies or nonprofit cooperatives.1
A. Historical Context
The privatization of public water systems in Argentina represented a
small part of a massive program that transferred almost all SOEs to
private hands during the 1990s, which, in itself, was a part of a larger
program of structural reform intended to reverse decades of economic
decline. In the late 1980s Argentina was experiencing growing inflation
driven in large part by printing money to finance huge fiscal deficits.
The deficit averaged approximately 9 percent of gross domestic product
during the decade. While federal and provincial overspending gener-
ated the lions share of these deficits, a nontrivial portion was due to
significant SOE losses. By the end of the decade the ruling Radical
government was unable to balance the budget. Further deficit spending
could not be financed through printing money or issuing new debt. In
1989 the country entered a period of hyperinflation that led the Radical
government to resign six months before the official end of its
administration.
The newly appointed Peronist government immediately launched an
ambitious structural reform program designed to reduce the budget
deficit, control inflation, and put the country back on a positive growth
path. The program consisted of financial and trade liberalization, a
monetary currency board, the decentralization of health and educa-
1
The only exception is a small mining town in Jujuy, where a private mining company
provided water service throughout the period of analysis.
Fig. 2.Percentage of municipalities with privatized water systems
TABLE 2
Discrete-Time Hazard Estimate of the Probability of Being Privatized
Mean
(Standard
Deviation) Model 1 Model 2
(1) (2) (3)
Time-varying covariates:
Federal government operates services .018 15.975*** 16.035***
(p1) (.134) (2.719) (2.727)
Local government by Radical party .139 3.198*** 3.204***
(p1) (.346) (1.067) (1.067)
Local government by Peronist party .719 .042 .054
(p1) (.449) (.401) (.402)
D log GDP per capitat1 .047 4.294 4.259
(.135) (3.567) (3.561)
D unemployment ratet1 .006 6.692 6.805
(.029) (5.696) (5.711)
D income inequalityt1 .005 .483 .139
(.014) (7.483) (7.503)
D child mortality ratet1 .266 .034
(2.994) (.043)
Fixed pretreatment characteristics as of
1991:
GDP per capita 60.601 .022*** .022***
(30.388) (.007) (.008)
Unemployment rate .045 12.871** 12.790**
(.023) (5.384) (5.383)
Income inequality .452 3.591 3.469
(.021) (5.820) (5.805)
Child mortality rate 6.208 .009
(3.683) (.036)
Population is 5,00025,000 (p1) .419 .227 .225
(.493) (.471) (.480)
Population is 25,00050,000 (p1) .202 .106 .110
(.402) (.535) (.540)
Population is 50,000100,000 (p1) .114 .261 .256
(.318) (.605) (.610)
Population is 100,000250,000 (p1) .079 .663 .668
(.269) (.612) (.615)
Population is more than 250,000 (p1) .066 1.159* 1.151*
(.249) (.631) (.640)
Proportion of families with UBN .246 13.660** 13.328**
(.151) (6.067) (6.226)
Proportion of families living in over- .097 13.560* 13.444*
crowded housing (.059) (7.150) (7.200)
Proportion of families living in poor .060 6.980** 6.987**
housing (.049) (3.472) (3.451)
Proportion of families living below .036 5.221 4.917
subsistence (.022) (7.418) (7.449)
Proportion of houses with no toilet .095 10.143** 9.798**
(.117) (4.429) (4.563)
No sewerage connection (p1) .280 .182 .171
(.449) (.323) (.328)
Proportion of household heads with .025 27.242** 27.182**
more than high school education (.012) (10.971) (11.003)
96 journal of political economy
A. Identification and Estimation Methods
Our objective is to identify the average effect of privatization on child
mortality rates in the municipalities in which the water supply system
has been privatized (i.e., the average impact of treatment on the
treated). Specifically, we are interested in comparing mortality when
water services are privately provided to the counterfactual, that is, mor-
tality when services are publicly provided in the treatment areas at the
same point in time. Since the counterfactual is never observed, we must
estimate it. In principle, we would like to randomly assign private and
public ownership across municipalities and compare the average out-
comes of the two groups. In the absence of a controlled randomized
trial, we are forced to turn to nonexperimental methods that mimic it
under reasonable conditions.
A major concern is that the municipalities that chose to privatize could
be different from the municipalities that chose not to privatize and that
these differences may be correlated with mortality. For example, poorer
urban areas in which mortality rates were higher may have been the
ones that privatized. In this case, the correlation between privatization
and mortality would be confounded with the wealth effect. In principle,
many of the types of (unobservable) characteristics that may confound
identification are those that vary across municipalities but are fixed over
time. A common method of controlling for time-invariant unobserved
heterogeneity is to use panel data and estimate difference-in-differences
models.
Therefore, without the benefit of a controlled randomized trial, we
turn to a difference-in-differences approach, which compares the
change in outcomes in the treatment group before and after the in-
tervention to the change in outcomes in the control group. By com-
paring changes, we control for observed and unobserved time-invariant
municipality characteristics that might be correlated with the privati-
zation decision as well as with mortality. The change in the control group
is an estimate of the true counterfactual, that is, what would have hap-
pened to the treatment group if there had been no intervention. An-
other way to state this is that the change in outcomes in treatment areas
controls for fixed characteristics and the change in outcomes in the
control areas controls for time-varying factors that are common to both
control and treatment areas.
Formally, the difference-in-differences model can be specified as a
two-way fixed-effect linear regression model:
yit p adIit bx it l t mi eit, (1)
where yit is the mortality rate in municipality i in year t, dIit is an indicator

variable that takes on the value one if municipality is water services are
TABLE 3
Impact of Privatization of Water Services on Child Mortality
Kernel
Matching
on
Using Observations on Common
100
Common
Full Sample Support
Supporta
(1) (2) (3) (4) (5) (6) (7)
Private water services (p1) .334 .320 .283 .540 .541 .525 .604
(.169)** (.170)* (.170)* (.177)*** (.178)*** (.178)*** (.168)***
[.157]** [.163]** [.162]* [.191]*** [.198]*** [.195]***
{.195}* {.203} {.194} {.261}** {.274}** {.266}**
%D in mortality rate 5.3 5.1 4.5 8.6 8.6 8.4 9.7
Other covariates:
Real GDP per capita .007 .009 .005 .006
(.005) (.006) (.006) (.006)
[.006] [.006] [.007] [.007]
{.007} {.007} {.007} {.008}
Unemployment rate .555 .636 .778 .836
(1.757) (1.758) (1.797) (1.802)
[2.161] [2.166] [2.249] [2.263]
{2.862} {2.846} {2.635} {2.635}
Income inequality 5.171 5.085 2.932 3.052
(2.868)* (2.880)* (2.907) (2.926)
[3.468] [3.445] [3.314] [3.289]
{3.696} {3.691} {3.833} {3.838}
Public spending per capita .028 .035 .068 .070
(.038) (.038) (.039)* (.039)*
[.055] [.055] [.059] [.059]
{.054} {.055} {.049} {.050}
Local government by Radical party (p1) .482 .166
(.267)* (.284)
[.281]* [.301]
{.288}* {.365}
Local government by Peronist party .202 .168
(p1) (.191) (.193)
[.202] [.230]
{.254} {.309}
R2 .1227 .1256 .1272 .1390 .1415 .1420
101
Observations 4,732 4,597 4,597 3,970 3,870 3,870 3,970

Note.Each column reports the estimated coefficients of a separate regression model in which the dependent variable is the child mortality rate, whose mean was 6.25 per thousand
in 1990. Standard errors are in parentheses. Standard errors clustered at the municipality level are in brackets. Standard errors clustered at the province-year level are in braces. All the
regressions include year and municipality fixed effects. The sample includes the municipalities with always-public, privatized, and nonprofit cooperative water companies (see table 1).
a
Standard errors for the kernel matching estimate are bootstrapped standard errors using 100 replications.
* Statistically different from zero at the .1 level of significance.
** Statistically different from zero at the .05 level of significance.
*** Statistically different from zero at the .01 level of significance.
TABLE 4
Impact of Privatization on Child Mortality by Cause of Death
1990 Mean Estimated Impact %D in

Mortality Rate Coefficients Mortality Rate
Infectious and parasitic diseases .565 .103 18.2
(.048)**
[.055]*
{.068}
Perinatal deaths 2.316 .266 11.5
(.105)**
[.107]**
{.123}**
All other causes in aggregate 2.565 .082 3.2
(.114)
[.101]
{.109}
All other causes disaggregated:
Accidents .399 .004
(.057)
Congenital anomalies .711 .022
(.056)
Skin and soft-tissue diseases .000 .000
(.001)
Blood and hematologic diseases .024 .002
(.008)
Nervous system disorders .163 .025
(.026)
Cardiovascular diseases .236 .006
(.030)
Gastrointestinal tract disorders .051 .001
(.010)
Genital and urinary diseases .020 .006
(.007)
Osteoarticular and connective .003 .001
tissue diseases (.001)
Respiratory diseases .511 .038
(.051)
Immunodeficiencies, endocrine, .376 .035
and nutrition system diseases (.033)
Mental disorders .002 .001
(.001)
Tumors .068 .006
(.015)
Note.Each cell reports the estimated coefficient on the private water services dummy from a different difference-
in-differences regression. Standard errors are in parentheses. Standard errors clustered at the municipality level are in
brackets. Standard errors clustered at the province-year level are in braces. All the regressions include year and mu-
nicipality fixed effects. All the regressions use the 3,870 observations on the common support and the socioeconomic
and political covariates included in the regression in col. 6 of table 3.
* Statistically different from zero at the .1 level of significance.
TABLE 5
Impact of Privatization on Child Mortality by Poverty Level
1990 Mean Estimated Impact %D in

Mortality Rate Coefficients Mortality Rate
Nonpoor municipalities 5.07 .114
(.233)
[.165]
{.159}
Poor municipalities 6.97 1.004 14.4
(.279)***
[.297]***
{.278}***
Extremely poor municipalities 9.11 2.415 26.5
(.544)***
[1.051]**
{.605}***
Note.Municipalities are divided into poverty groups using the governments index of UBN using data from the
1991 census. Nonpoor municipalities are defined as those in which less than 25 percent of households have UBN. Poor
municipalities are defined as those in which 2550 percent of households have UBN. Extremely poor municipalities
are defined as those in which more than 50 percent of households have UBN. The reported coefficients are the
interaction of the private water services dummy and UBN (recoded in a set of dummy variables in the three categories:
below 25 percent, between 25 and 50 percent, and above 50 percent) in a difference-in-differences regression using
only the 3,870 observations on the common support. The regression includes year and municipality fixed effects and
the socioeconomic and political covariates used in the regression reported in col. 6 of table 3. Standard errors are in
parentheses. Standard errors clustered at the municipality level are in brackets. Standard errors clustered at the province-
year level are in braces.
in child mortality in municipalities with high levels of poverty (UBN

greater than 50 percent). This result is consistent with the predictions
of our causal model. The effect of privatization on child mortality should
be stronger for the groups that are more vulnerable to water-related
diseases.
IV. Pathways
In Section III, we provide evidence that child mortality fell faster in
areas that privatized water companies than in areas that did not privatize.
In this section, we show that firms that privatized made choices that
affected the pathways by which child health could be improved. There
are a number of potential pathways by which the privatization of water
systems might have induced the reduction in child mortality. First, pri-
vatization may have expanded the water supply and sewerage network
providing access to service to households that were not previously con-
nected to water and sewerage. Second, there may have been improve-
ments in service quality in terms of reduced spillage of water and sewage,
faster repair rates, fewer shortages, cleaner water, and better water pres-
sure and sewage treatment. All these quality enhancements improve the
epidemiological environment. In this section, we present evidence that
privatization affected these pathways.
water for life 109
TABLE 6
Comparison of OSN (Public) vs. Aguas Argentinas (Private) Performance, 198099
OSNa (before Aguas Argentinasb D after

Privatization) (after Privatization) Privatization (%)
Water production (1) (mil-
lions of m3 per day) 3.56 3.89 9.3
Spilled water (2) (millions of
m3 per day) 1.49c 1.27 14.8
Water supply (12) (millions
of m3 per day) 2.07c 2.62 26.6
Sewage drainage volume (mil-
lions of m3 per day) 2.18 2.45 12.4
Water network extension (km
of network) 10,148 13,287 30.9
Sewerage network extension
(km of network) 6,875 8,312 20.9
Average delay in attending
repair requests (days) 180d 32e 82.2
Water leakages repaired per
year 42,000c 96,383 129.5
Sewerage blockages repaired
per year 100,000c 148,500 48.5
Percentage of clients with
appropriate water pressure 17c 54f 217.6
Water turbidness (turbidness
units) 7.5 2.3 70
Usage fee indexg 100 84 16
Employees 9,300 4,000 57
a
Average for the period 198092.
b
Average for the period 199499.
c
1993 only.
d
1992 only.
e
Average excludes 1994.
f
1996 only.
g
Corresponds to the K tariff factor. Source: Universidad Argentina de la EmpresaCentro de Estudios Economicos
de la Regulacion.
suburban areas of Greater Buenos Aires. Since 98 percent of households

in the city of Buenos Aires were already connected to water services
before privatization, most of the expansion in access necessarily had to
occur among lower-income households in the suburban areas. Indeed,
table 7 shows that middle- and low-income households accounted for
84.6 percent of the new connections.
B. Access to Water Services

While the data for Buenos Aires show that the privatization improved
service quality and expanded access to water services, we are unable to
similarly assess the impact of privatization for the rest of the country.
We are, however, able to say something about the effect of privatization
on access to water services. Even though increased access may not be
the only mechanism through which privatization can affect child mor-
Fig. 3.Logarithm of population connected to OSNAguas Argentinas network
water for life 111
TABLE 7
Network Expansion by Income Group in Greater Buenos Aires,
19932000
New
Income Level Connections Percentage
High and upper-middle income 90,200 15.4
Lower-middle income 282,250 48.3
Low income 211,800 36.3
Total 584,250 100.0
Source.Subsecretara de Recursos Hdricos, from Abdala and Spiller (1999).
tality, it is probably one of the most important causal channels. Indeed,

acquiring water services for the first time is likely to imply a more im-
portant change in access to safe water relative to service improvements
to households with existing water and sewerage connections.
We evaluate the impact of privatization on access to water services
using data from the 1991 census and the 1997 Encuesta de Desarrollo
Social (EDS). The EDS was a stratified random survey of about 30,000
households from urban municipalities with more than 5,000 inhabitants,
and it asked questions about household connections to water services
identical to those asked in the census.
To identify the effect of privatization on access to water, we exploit
the fact that by 1997 a number of municipalities had already privatized
their water services (fig. 2). Using the data from municipalities in the
EDS survey, we calculate the difference-in-differences estimate of the
impact of privatization on the proportion of households that had access
to the water network. The difference-in-differences estimator compares
the change in the proportion of households connected to water services
in municipalities that privatized to the change in the proportion con-
nected in municipalities that did not privatize water services. For this
exercise, a municipality is in the privatized group if the privatization of
water services occurred between 1990 and 1996.
The results, reported in table 8, show a significantly larger increase
in the proportion of households connected to water services in the
municipalities that privatized than in those that did not. The estimated
impact is even higher when we exclude the capital city, where 98 percent
of households were already connected to water service before privati-
zation. In this sample, the results suggest that the number of households
connected to the water network increased by 4.2 percentage points as
a result of privatization.
This estimate, however, most likely underestimates the impact of pri-
vatization on access for two reasons. First, it includes the impact of
privatization only through 1997. In Cordoba, for example, water services
were privatized in that year, and coverage increased by more than 10
TABLE 8
Difference-in-Differences Estimates of the Impact of Privatization on the
Proportion of Households Connected to the Water Network, 199197
All Excluding
Municipalities Buenos Aires
Municipalities That Were Not
Privatized before 1997
Proportion of households connected in
public
1991 (p91 ) .866 .866
public
1997 (p97 ) .898 .898
Difference 1997 1991 (p97
public
p91
public
) .032 .032
Municipalities That Were Privatized
before 1997
private
1991 (p91 ) .730 .640
private
1997 (p97 ) .780 .714
Difference 1997 1991 (p97 private
p91
private
) .050 .074
Difference-in-differences (p97private
p91
private
)
(p97
public
p91 public
) .018 .042
Z-test for difference-in-differences
estimatea 2.83*** 5.78***
Note.The preintervention connection rates are higher in control areas than in treatment areas in this table.
However, this analysis takes into account privatization only through 1996, and therefore, the control group includes
the set of municipalities that privatized later. When all the privatized municipalities are included in the treatment
group, the preprivatization connection rates are the same in (eventually) treated and control groups. Specifically, in
1991, 74 percent of households were connected in eventually privatized areas and 70 percent in never privatized areas.
a
The statistic of contrast is
private
(p97 p91
private
) (p97
public
p91
public
)
zp ,
[p97
private
(1 p97
private private
)/n97 ] [p97
public
(1 p97
public public
)/n97 ]
where pt is the proportion of households with access to water connection in year t in a municipality in which water has
been privatized (private) or has not been privatized (public), and n is the number of observations. Note that there is
no sample variability when we estimate p for 1991 since these statistics are estimated from census data.
percentage points in the first three years of concession. Second, the

EDS grossly undersampled poor areas,15 and access expanded most in
poor areas in which fewer households were connected at the baseline.
For example, table 7 showed that connections increased the most among
the poor in Greater Buenos Aires, whereas Artana et al. (1999) report
that after privatization in Corrientes, one of the poorest provinces in
the country, the number of connections to the water network in the
15
Specifically, when we split the sample into three groupsnonpoor municipalities in
which less than 25 percent of households have UBN in the 1991 census, poor municipalities
in which between 25 and 50 percent have UBN, and extremely poor municipalities in
which more than 50 percent have UBNwe found that the EDS does not include any
extremely poor municipalities and includes only a few poor municipalities.
Diferencas em diferencas
O metodo de EF requer microdados repetidos (indivduos, firmas).

Porem muitas vezes, a variacao de interesse muda num nvel mais
agregado (estado ou coorte).

O metodo de diferencas-em-diferencas (DD) e uma versao de EF com

dados agregados. A ideia e usar variacoes em algumas regioes e nao
em outras para estimar o efeito desejado.

O metodo de diferencas-em-diferencas (DD) e uma versao de EF com

dados agregados. A ideia e usar variacoes em algumas regioes e nao
em outras para estimar o efeito desejado.
Exemplo, efeito de salario mnimo em emprego (Card e Krueger

1994). Eles usam a variacao no salario mnimo em NJ e nao em PA,
comparando a mudanca no emprego de rstaurantes em NJ entre
Fev-Nov em NJ comparado com restaurantes em PA.
Assumamos que na ausencia de programa, o resultado e deterkinado

por um componente que nao varia no tempo e outro que varia, mas
que e comum para os estados:
E [Y0ist |s, t] = s + t

E [Y0ist |s, t] = s + t
Assumindo que E [Y1ist Y0ist |s, t] e constante e igual a , podemos

escrever:
Yist = s + t + Dst + ist

E [Y0ist |s, t] = s + t
Assumindo que E [Y1ist Y0ist |s, t] e constante e igual a , podemos

escrever:
Qual e o pressuposto de identificacao? Que a tendencia em ambos

estado seria a mesma na ausencia do tratamento.
5.2. DIFFERENCES-IN-DIFFERENCES 123
Table 5.2.1: Average employment per store before and after the New Jersey minimum wage increase
PA NJ Dierence, NJ-PA
Variable (i) (ii) (iii)
1. FTE employment before, 23.33 20.44 -2.89
all available observations (1.35) (0.51) (1.44)
2. FTE employment after, 21.17 21.03 -0.14
all available observations (0.94) (0.52) (1.07)
3. Change in mean FTE -2.16 0.59 2.76
employment (1.25) (0.54) (1.36)
Notes: Adapted from Card and Krueger (1994), Table 3. The
table reports average full-time equivalent (FTE) employment at
restaurants in Pennsylvania and New Jersey before and after a
minimum wage increase in New Jersey. The sample consists of
all stores with data on employment. Employment at six closed
stores is set to zero. Employment at four temporarily closed stores
is treated as missing. Standard errors are reported in parentheses
Table 5.2.1 (based on Table 3 in Card and Krueger, 1994) shows average employment at fast food
restaurants in New Jersey and Pennsylvania before and after the change in the New Jersey minimum wage.
There are four cells in the rst two rows and columns, while the margins show state dierences in each
period, the changes over time in each state, and the dierence-in-dierences. Employment in Pennsylvania
restaurants is somewhat higher than in New Jersey in February but falls by November. Employment in New
Jersey, in contrast, increases slightly. These two changes produce a positive dierence-in-dierences, the
opposite of what we might expect if a higher minimum wage pushes businesses up the labor demand curve.
How convincing is this evidence against the standard labor-demand story? The key identifying assump-
tion here is that employment trends would be the same in both states in the absence of treatment. Treatment
induces a deviation from this common trend, as illustrated in gure 5.2.1. Although the treatment and con-
trol states can dier, this dierence in captured by the state xed eect, which plays the same role as the
unobserved individual eect in (5.1.3).7
The common trends assumption can be investigated using data on multiple periods. In an update of their
original minimum wage study, Card and Krueger (2000) obtained administrative payroll data for restaurants
in New Jersey and Pennsylvania for a number of years. These data are shown here in Figure 5.2.2, similar
to Figure 2 in their follow-up study. The vertical lines indicate the dates when their original surveys were
conducted, and the third vertical line denotes the increase in the federal minimum wage to $4.75 in October
1996, which aected Pennsylvania but not New Jersey. These data give us an opportunity to look at a new
minimum wage "experiment".
Like the original Card and Krueger survey, the administrative data show a slight decline in employment
from February to November 1992 in Pennsylvania, and little change in New Jersey over the same period.
However, the data also reveal fairly substantial year-to-year employment variation in other periods. These
swings often seem to dier substantially in the two states. In particular, while employment levels in
7 The common trends assumption can be applied to transformed data, for example,
E(log y0ist js; t) = s + t:
Note, however, that if there is a common trend in logs, there will not be one in levels and vice versa. Athey and Imbens
(2006) introduce a semi-parametric DD estimator that allows for common trends after an unknown transformation, which they
propose to use the data to estimate. Poterba, Venti and Wise (1995) and Meyer, Viscusi, and Durbin (1995) discuss DD-type
models for quantiles.
124 CHAPTER 5. FIXED EFFECTS, DD, AND PANEL DATA
employment
rate
employment trend in
treatment state employment trend in
control state
treatment
effect
counterfactual
employment trend in
treatment state
before after time
Figure 5.2.1: Causal eects in the dierences-in-dierences model
New Jersey and Pennsylvania were similar at the end of 1991, employment in Pennsylvania fell relative to
employment in New Jersey over the next three years (especially in the 14-county group), mostly before the
1996 change in Federal minimum. So Pennsylvania may not provide a very good measure of counterfactual
employment rates in New Jersey in the absence of a policy change, and vice versa.
A more encouraging example comes from Pischke (2007), who looks at the eect of school term length
on student performance using variation generated by a sharp policy change in Germany. Until the 1960s,
children in all German states except Bavaria started school in the Spring. Beginning in the 1966-67 school
year, the Spring-starters moved to start school in the Fall. The transition to a Fall start required two short
school years for aected cohorts, 24 weeks long instead of 37. Students in these cohorts eectively had
their time in school compressed relative to cohorts on either side and relative to students in Bavaria, which
already had a Fall start.
Figure 5.2.3 plots the likelihood of grade repetition for the 1962-73 cohorts of 2nd graders in Bavaria and
aected states (there are no repetition data for 1963-65). Repetition rates in Bavaria were reasonably at
from 1966 onwards at around 2.5%. Repetition rates are higher in the short-school-year states, at around
4 - 4.5% in 1962 and 1966, before the change in term length. But repetition rates jump up by about a
percentage point for the two aected cohorts in these states, a bit more so for the second cohort than the rst,
before falling back to the baseline level. This graph provides strong visual evidence of treatment and control
states with a common underlying trend, and a treatment eect that induces a sharp but transitory deviation
from this trend. A shorter school year seems to have increased repetition rates for aected cohorts.
5.2.1 Regression DD
As with the xed eects model, we can use regression to estimate equations like (5.2.2). Let N Js be a
dummy for restaurants in New Jersey and dt be a time-dummy that switches on for observations obtained
in November (i.e., after the minimum wage change). Then
yist = + N Js + dt + (N Js dt ) + "ist (5.2.3)
is the same as (5.2.2) where N Js dt =dst . In the language of Section 3.1.4, this model includes two main
eects for state and year and an interaction term that marks observations from New Jersey in November.
This is a saturated model since the conditional mean function E(yist js; t) takes on four possible values and
there are four parameters. The link between the parameters in the regression equation, (5.2.3), and those
1406 THE AMERICAN ECONOMIC REVIEW DECEMBER 2000
/ N J -,-PA; 7 c o u n t i e s ....... P A ; 1 4 c o u n t i e q
Figure Note:
5.2.2:Vertical lines indicate
Employment in dates
NewofJersey
originaland
Card-Kmeger survey fast-food
Pennsylvania and the October 1996 federal
restaurants, minimum-wage
October increase.
1991 to September
Source: Authors' calculations based on BLS ES-202 data.
1997 (from Card and Krueger 2000). Vertical lines indicate dates of the original Card and Krueger (1994)
survey and the October 1996 federal minimum-wage increase.

sections of fast-food restaurants for the period ment growth in Pennsylvania exceeded that in
from 1991 to 1997. We used these cross- New Jersey, on whole the figure provides little
sectional samples to calculate total employment evidence that Pennsylvania's employment
for New Jersey, for the 7 counties of Pennsyl- growth exceeded New Jersey's in the few years
vania used in our original study, and for the following the minimum-wage increase.
broader set of 14 eastern Pennsylvania counties
in each month. Figure 2 summarizes the time- A. The Effect of the 1996 Federal Minimum-
series patterns of aggregate employment from Wage Increase
these files. For each of the three geographic
regions, the figure shows aggregate monthly On October 1, 1996, the federal minimum
employment in the fast-food industry relative to wage increased from $4.25 per hour to $4.75
their respective February 1992 levels. per hour. This increase was binding in Pennsyl-
The figure reveals a pattern that is consistent vania, but not in New Jersey, where the state's
with the longitudinal estimates. In particular, $5.05 minimum wage already exceeded the new
between February and November of 1992-the federal standard. Consequently, the same com-
main months our survey was conducted-fast- parison can be conducted in reverse, with New
food employment grew by 3 percent in New Jersey now serving as a "control group" for
Jersey, while it fell by 1 percent in the 7 Penn- Pennsylvania's experience. This reverse com-
sylvania counties and fell by 3 percent in the 14 parison is particularly useful because any long-
Pennsylvania counties. Although it is possible run economic trends that might have biased
to find some pairs of months surrounding the employment growth in favor of New Jersey
minimum-wage increase over which employ- during the previous minimum-wage hike will
0.060
Grade Repetition Rates Grade 2

0.050 0.060
0.040 0.050
repeating
repeating
0.030 0.040
Fraction
Fraction
0.030 0.020
0.020
1962 1963 1964

1964 1965
1965 1966
1966 1967
1967 1968
1968 1969
1969 1970 1971 1972 1973
School
School year
year ending
ending
SSY States
SSY
repetition
States rates Bavaria
Bavaria
(control)
(control)
repetition rates
SSY States
Affected
repetition
cohorts
rates Bavaria (control) repetition rates
Figure 5.2.3: Average rates of grade repetition in second grade for treatment and control schools in Germany
(from Pischke 2007). The data span a period before and after a change in term length for students outside
of Bavaria.
Econometria Aplicada
Dados em Painel, Efeito-Fixo, e Dif-em-Dif
Professor Claudio Ferraz
PUC-Rio
Aula 11
Professor Claudio Ferraz (1o semestre 2009)Econometria AplicadaDados em Painel, Efeito-Fixo, e Dif-em-Dif Aula 11 1 / 10
Dados em painel Efeitos fixos
Assumindo que a variavel nao-observavel e fixa no tempo, podemos

utilizar o modelo de efeitos fixos:

Ele pode ser estimado de tres formas:


I Colocando uma dummy para cada observacao i e cada observacao t.
Isso pode ser muito intensivo computacionalmente quando i e grande.


I Utilizar desvios em relacao a media.


I Tirando a primeir diferenca.


I Tirando a primeir diferenca.
Com dois perodos todos estimadores sao iguais. Com t > 2, desvio
de medias e mais eficiente. Tomar cuidado com erro padrao quando
utilizar diferencas.
O metodo de EF requer microdados repetidos (indivduos, firmas) e

analisa variacao no nvel do indivduo (ex. trabalhador sindicalizado
ou nao). Porem muitas vezes, a variacao de interesse varia num nvel
mais agregado (estado ou coorte).

O metodo de diferencas-em-diferencas (DD) e uma versao de EF

quando o nvel do tratamentoe agregado. A ideia e usar variacoes
que ocorrem em algumas regioes e nao em outras para estimar o
efeito desejado.

O metodo de diferencas-em-diferencas (DD) e uma versao de EF

quando o nvel do tratamentoe agregado. A ideia e usar variacoes
que ocorrem em algumas regioes e nao em outras para estimar o
efeito desejado.
Exemplo, efeito de salario mnimo em emprego (Card e Krueger

1994). Eles usam a variacao no salario mnimo em NJ e nao em PA,
comparando a mudanca no emprego de rstaurantes em NJ entre
Fev-Nov em NJ comparado com restaurantes em PA.
Assumamos que na ausencia de programa, o resultado e determinado

E [Y0ist |s, t] = s + t

E [Y0ist |s, t] = s + t
Podemos escrever:

E [Y0ist |s, t] = s + t
Podemos escrever:
Podemos mostrar que:
E [Yist |s = PA, t = Nov ] E [Yist |s = PA, t = Feb] = Nov Feb

E [Yist |s = NJ, t = Nov ] E [Yist |s = NJ, t = Feb] = Nov Feb +
Assim, podemos mostrar que:
(E [Yist |s = NJ, t = Nov ] E [Yist |s = NJ, t = Feb])

(E [Yist |s = PA, t = Nov ] E [Yist |s = PA, t = Feb]) =




O estimador de Dif-em-Dif pode ser implementado atraves de uma

regressao. Seja NJs uma dummy de restaurantes em NJ e dt uma
dummy igual a um para o perdio pos-tratamento, podemos escrever
o modelo como:
Yist = + NJs + dt + (NJs dt ) + ist
O estimador de DD pode ser generalizado adicionando estados como

controle e perodos. Com varios perodos e estados, o modelo geral
inclui uma dummy para cada estado e perodo.

Alem disso, pode ser implementado com variaveis contnuas e com a

inclusao de outros controles.

Alem disso, pode ser implementado com variaveis contnuas e com a

inclusao de outros controles.
Mesmo em casos onde o tratamento varia no nvel do estado,

podemos usar microdados. As vezes, isto pode aumentar a nossa
precisao se estimarmos:
0
Yist = s + t + (Ds dt ) + Xist + ist
Quando temos dados de varios anos (ou coortes), podemos testar se

unidades que nao deveriam ser afetadas pelo tratamento foram
afetadas. Polticas futuras nao deveria influenciar Y no perodo t:
m
X
Yist = s + t + Ds,t
=0
q
X
0
+ + Ds,t+ + Xist + ist
=1
Quando temos dados de varios anos (ou coortes), podemos testar se

unidades que nao deveriam ser afetadas pelo tratamento foram
afetadas. Polticas futuras nao deveria influenciar Y no perodo t:
m
X
Yist = s + t + Ds,t
=0
q
X
0
+ + Ds,t+ + Xist + ist
=1
Com varios perodos, um teste de robustez adicional e incluir

tendencias que variam para cada unidade de tratamento (ex. estados):
0
Yist = 0s + 1s t + t + (Ds dt ) + Xist + ist
model includes both leads and lags, as in equation (5.2.6). The estimated leads and lags, running from two
years ahead to 4 years behind, are plotted in Figure 5.2.4, a reproduction of Figure 3 from Autor (2003). The
estimates show no eects in the two years before the courts adopted an exception, with sharply increasing
eects on temporary employment in the rst few years after the adoption, which then appear to atten out
with a permanently higher rate of temporary employment in aected states. This pattern seems consistent
with a causal interpretation of Autors results.
50
40
Vertical bands represent 1.96 times the standard error of each point estimate
30
20
Log points
10
0
2 Years Prior 1 Year Prior Year of Adoption 1 Year After 2 Years After 3 Years After 4 or More Years
After
-10
-20
Time passage relative to year of adoption of implied contract exception
Figure 5.2.4: Estimated impact of state courtsadoption of an implied-contract exception to the employment-
at-will doctrine on use of temporary workers (from Autor 2003). The dependent variable is the log of state
temporary help employment in 1979 - 1995. Estimates are from a model that allows for eects before, during,
and after adoption.
An alternative check on the DD identication strategy adds state-specic time trends to the regressors
in Xist . In other words, we estimate
yist = 0s + 1st + t + dst + X0ist + "ist ; (5.2.7)
where 0s is a state-specic intercept as before and 1s is a state-specic trend coe cient multiplying the
time-trend variable, t. This allows treatment and control states to follow dierent trends in a limited but
potentially revealing way. Its heartening to nd that the estimated eects of interest are unchanged by
the inclusion of these trends, and discouraging otherwise. Note, however, that we need at least 3 periods
to estimate a model with state-specic trends. Moreover, in practice, 3 periods is typically inadequate to
pin down both the trends and the treatment eect. As a rule, DD estimation with state-specic trends is
likely to be more robust and convincing when the pre-treatment data establish a clear trend that can be
extrapolated into the post-treatment period.
In a study of the eect of labor regulation on businesses in Indian states, Besley and Burgess (2004)use
state trends as a robustness check. Dierent states change regulatory regimes at dierent times, giving rise
to a DD research design. As in Card (1992), the unit of observation in Besley and Burgess (2004) is a
state-year average. Table 5.2.3 (based on Table IV in their paper) reproduces the key results.
Table 5.2.3: Eect of labor regulation

on the performance of rms in Indian states
(1) (2) (3) (4)
Labor regulation (lagged) -0.186 -0.185 -0.104 0.0002
(.0641)
(.0507) (.039) (.02)
Log development 0.240 0.184 0.241
expenditure per capita (.1277) (.1187) (.1057)
Log installed electricity 0.089 0.082 0.023
capacity per capita (.0605) (.0543) (.0333)
Log state population 0.720 0.310 -1.419
(.96) (1.1923) (2.3262)
Congress majority -0.0009 0.020
(.01) (.0096)
Hard left majority -0.050 -0.007
(.0168) (.0091)
Janata majority 0.008 -0.020
(.0235) (.0333)
Regional majority 0.006 0.026
(.0086) (.0234)
State-specic trends NO NO NO YES
Adjusted R-squared 0.93 0.93 0.94 0.95
Notes: Adapted from Besley and Burgess (2004), Table IV. The table reports
regression-DD estimates of the eects of labor regulation on productivity. The
dependent variable is log manufacturing output per capita. All models include
state and year eects. Robust standard errors clustered at the state level are
reported in parentheses. State amendments to the Industrial Disputes Act are
coded 1=pro-worker, 0 = neutral, -1 = pro-employer and then cumulated over
the period to generate the labor regulation measure. Log of installed electrical
capacity is measured in kilowatts, and log development expenditure is real per
capita state spending on social and economic services. Congress, hard left, Janata,
and regional majority are counts of the number of years for which these political
groupings held a majority of the seats in the state legislatures. The data are for
the sixteen main states for the period 1958-1992. There are 552 observations.
The estimates in column 1, from a regression-DD model without state-specic trends, suggest that labor
regulation leads to lower output per capita. The models used to construct the estimates in columns 2 and 3
add time-varying state-specic covariates like government expenditure per capita and state population. This
is in the spirit of Cards (1992) addition of state-level adult employment rates as a control in the minimum
wage study. The addition of controls aects the Besley and Burgess estimates little. But the addition
of state-specic trends kills the labor-regulation eect, as can be seen in column 4. Apparently, labor
regulation in India increases in states where output is declining anyway. Control for this trend therefore
drives the estimated regulation eect to zero.
Picking Controls
Weve labeled the two dimensions in the DD set-up states and time because this is the archetypical
DD example in applied econometrics. But the DD idea is much more general. Instead of states, the
subscript s might denote demographic groups, some of which are aected by a policy and others are not.
For example, Kugler, Jimeno, and Hernanz (2005) look at the eects of age-specic employment protection
policies in Spain. Likewise, instead of time, we might group data by cohort or other types of characteristics.
Table 5.2.3: Eect of labor regulation

on the performance of rms in Indian states
(1) (2) (3) (4)
Labor regulation (lagged) -0.186 -0.185 -0.104 0.0002
(.0641)
(.0507) (.039) (.02)
Log development 0.240 0.184 0.241
expenditure per capita (.1277) (.1187) (.1057)
Log installed electricity 0.089 0.082 0.023
capacity per capita (.0605) (.0543) (.0333)
Log state population 0.720 0.310 -1.419
(.96) (1.1923) (2.3262)
Congress majority -0.0009 0.020
(.01) (.0096)
Hard left majority -0.050 -0.007
(.0168) (.0091)
Janata majority 0.008 -0.020
(.0235) (.0333)
Regional majority 0.006 0.026
(.0086) (.0234)
State-specic trends NO NO NO YES
Adjusted R-squared 0.93 0.93 0.94 0.95
Notes: Adapted from Besley and Burgess (2004), Table IV. The table reports
regression-DD estimates of the eects of labor regulation on productivity. The
dependent variable is log manufacturing output per capita. All models include
state and year eects. Robust standard errors clustered at the state level are
reported in parentheses. State amendments to the Industrial Disputes Act are
coded 1=pro-worker, 0 = neutral, -1 = pro-employer and then cumulated over
the period to generate the labor regulation measure. Log of installed electrical
capacity is measured in kilowatts, and log development expenditure is real per
capita state spending on social and economic services. Congress, hard left, Janata,
and regional majority are counts of the number of years for which these political
groupings held a majority of the seats in the state legislatures. The data are for
the sixteen main states for the period 1958-1992. There are 552 observations.
The estimates in column 1, from a regression-DD model without state-specic trends, suggest that labor
regulation leads to lower output per capita. The models used to construct the estimates in columns 2 and 3
add time-varying state-specic covariates like government expenditure per capita and state population. This
is in the spirit of Cards (1992) addition of state-level adult employment rates as a control in the minimum
wage study. The addition of controls aects the Besley and Burgess estimates little. But the addition
of state-specic trends kills the labor-regulation eect, as can be seen in column 4. Apparently, labor
regulation in India increases in states where output is declining anyway. Control for this trend therefore
drives the estimated regulation eect to zero.
Picking Controls
Weve labeled the two dimensions in the DD set-up states and time because this is the archetypical
DD example in applied econometrics. But the DD idea is much more general. Instead of states, the
subscript s might denote demographic groups, some of which are aected by a policy and others are not.
For example, Kugler, Jimeno, and Hernanz (2005) look at the eects of age-specic employment protection
policies in Spain. Likewise, instead of time, we might group data by cohort or other types of characteristics.
TABLE 2
Discrete-Time Hazard Estimate of the Probability of Being Privatized
Mean
(Standard
Deviation) Model 1 Model 2
(1) (2) (3)
Time-varying covariates:
Federal government operates services .018 15.975*** 16.035***
(p1) (.134) (2.719) (2.727)
Local government by Radical party .139 3.198*** 3.204***
(p1) (.346) (1.067) (1.067)
Local government by Peronist party .719 .042 .054
(p1) (.449) (.401) (.402)
D log GDP per capitat1 .047 4.294 4.259
(.135) (3.567) (3.561)
D unemployment ratet1 .006 6.692 6.805
(.029) (5.696) (5.711)
D income inequalityt1 .005 .483 .139
(.014) (7.483) (7.503)
D child mortality ratet1 .266 .034
(2.994) (.043)
Fixed pretreatment characteristics as of
1991:
GDP per capita 60.601 .022*** .022***
(30.388) (.007) (.008)
Unemployment rate .045 12.871** 12.790**
(.023) (5.384) (5.383)
Income inequality .452 3.591 3.469
(.021) (5.820) (5.805)
Child mortality rate 6.208 .009
(3.683) (.036)
Population is 5,00025,000 (p1) .419 .227 .225
(.493) (.471) (.480)
Population is 25,00050,000 (p1) .202 .106 .110
(.402) (.535) (.540)
Population is 50,000100,000 (p1) .114 .261 .256
(.318) (.605) (.610)
Population is 100,000250,000 (p1) .079 .663 .668
(.269) (.612) (.615)
Population is more than 250,000 (p1) .066 1.159* 1.151*
(.249) (.631) (.640)
Proportion of families with UBN .246 13.660** 13.328**
(.151) (6.067) (6.226)
Proportion of families living in over- .097 13.560* 13.444*
crowded housing (.059) (7.150) (7.200)
Proportion of families living in poor .060 6.980** 6.987**
housing (.049) (3.472) (3.451)
Proportion of families living below .036 5.221 4.917
subsistence (.022) (7.418) (7.449)
Proportion of houses with no toilet .095 10.143** 9.798**
(.117) (4.429) (4.563)
No sewerage connection (p1) .280 .182 .171
(.449) (.323) (.328)
Proportion of household heads with .025 27.242** 27.182**
more than high school education (.012) (10.971) (11.003)
Uma das preocupacoes quando estimamos algo por DD e que a

composicao do grupo de tratamento e controle mude. Suponha que o
tratamento causa migracao.

Estimador de VI pode ser usado para solucionar o problema. Um

exemplo e a utilizacao de lugar de nascimento como instrumento para
a residencia atual.

Estimador de VI pode ser usado para solucionar o problema. Um

exemplo e a utilizacao de lugar de nascimento como instrumento para
a residencia atual.
Em alguns casos, alem da dimensao localizacao e tempo, podemos

explorar outra variacao no tratamento. Suponha que uma poltica so
afeta familias com criancas ate 15 anos. Podemos estimar um modelo
de diferencas triplas:
0
Yiast = st + at + as + Dast + Xiast + iast
Diferencas em diferencas com coortes
Uma aplicacao do modelo de DD e com dados em cross-section

explorando o fato que um programa afetou pessoas de diferentes
idades de forma diferente.
Diferencas em diferencas com coortes
Uma aplicacao do modelo de DD e com dados em cross-section

explorando o fato que um programa afetou pessoas de diferentes
idades de forma diferente.
Duflo (2001) estima o seguinte modelo:
Yijk = c1 + 1j + 1k + (Pj Ti )1 + (Cj Ti )1 + ijk
onde Ti e uma dummy se indivduo pertence a coorte afetada, Pj

mede a intensidade do programa, 1j e um efeito fixo de distrito e
1k um efeito fixo de coorte.
I Em alguns casos, grupos de tratamento e controle no seguem

tendncias paralelas => Did-Dif nos dar um estimador viesado
I Em outras situaes temos somente uma unidade tratada e

muitas unidades potenciais no grupo de controle
I Ideia bsica: combinao de unidades no tratadas pode nos

dar um grupo de comparao melhor do que unidades
individuais
2 / 13
Origem do mtodo: efeito de imigrao no mercado
de trabalho
I Em 20 de abril de 1980, Fidel Castro anuncia que Cubanos que

quisessem deixar a ilha em direo aos EUA podiam partir do
porto de Mariel
I Entre maio e setembro de 1980 125.000 cubanos (Marielitos)
deixaram a ilha e chegaram em Miami
3 / 13
de trabalho

porto de Mariel
3 / 13
Qual o efeito de uma grande onda de imigrao no
salrio dos residentes?
I Tema muito atual e controverso (imigrao da Sria, Brexit,
Trump)
I Card (1990) argumenta que aumento de 7% da fora de trabalho

local em Miami no teve efeitos sobre salrios e desemprego de
trabalhadores locais no-qualificados
I Implicitamente usa controle sinttico ao comparar o mercado de

trabalho de Miami com outras 4 cidades que tinham composio
de emprego e tendncias parecidas (Atlanta, Houston, Los
Angeles, Tampa-St. Petersburg). Mas no de forma eficiente
(mesmos pesos)
4 / 13
de trabalho

porto de Mariel
5 / 13
de trabalho

porto de Mariel
5 / 13
Comparao dos grupos
T H E MARIEL BOATIAFT
Table 4 . Unemployment Rates of Individuals Age 16-61 in Miami and
Four Comparison Cities, 1979-85.
(Standard Errors in Parentheses)
Blacks 8.3 5.6 9.6 16.0 18.4 14.2 7.8

(1.7) (1.3) (1.8) (2.3) (2.5) (2.3) (2.3)
Cubans 5.3 7.2 10.1 10.8 13.1 7.7 5.5
(1.2) (1.3) (1.5) (1.5) (1.ti) (1.4) (1.7)
Hispanics 6.3 7.7 11.8 9.1 7.5 12.1 3.7
(2.3) (2.2) (3.0) (2.5) (2.1) (2.4) (1.9)
Whites
Blacks
Hispanics 6.3 8.7 8.3 12.1 11.8 9.8 9.3

(0.6) (0.6) (0.6) ('3.7) (0.7) ('3.6) (0.8)
'Vote: Entries represent means of unemploylnent indicator variable for individuals age 16-61 in Miami and
four comparison cities: Atlanta, Houston, Los Angeles, and Tampa-St. Petersburg. Samples are based on
individuals in the labor force. See notes to Table 3 for definitions of groups and data sources.
2-4 points lower than those in the unemployment rates in Tables 3 and 4,
comparison cities from 1979 to 1981, which combine workers of all ages and
equalled or exceeded those in the compar- education levels. do not directlv address
ison cities from 1982 to 1984. The 1985 the question of whether the ~ a i i e immi-
l 6 / 13
T H E MAKIEL BOATLIFT 253
Efeito dos Marielitos
Table 6. Comparison of Wages, Unemployment Rates, and Employment Rates for Blacks in
Miami and Comparison Cities.
All Bncks Low-Educnt2otz Blacks

Dfferetzce ztz
Dfference 112
Dfference 111 Emp IC'tzemp ,

Dfference in Entp Il'nemp ,
Log I.17nges, ~ b f z n m-~ Compar~sotz

Log IVnges, lV11nm~- Compnrlson
1bf2amz - Co?r~par~son K- 1V11nmz - C o m p a ~ ~ s o n E~~ -

Year Actual Adizlited PoB. Rnte C'tzemB Rnte Actztnl Adrwtcd PoB Rnte C'nen~b Rnte
1979 -.I5 -.I2 .OO - 2.0 -.I3 -.I5
.03 - .8
i.03) i.03) (.03) (1.9) i.05) (.05)
(.04) (3.8)
1980 -.1G -.I2 .05 -7.1 - .07 - .07
.03 - 8.2
i.03) (.03) i.03) (1.6) i.05) i.05)
i.04) (3.5)
1981 -.I1 -.lo .02 - 3.0 - .05 -.I1
.04 - 7.7
i.03) i.03) i.03) (2.0) i.05) (.05)
i.04) (4.2)
1982 -.24 - .20 - ,015 3.3 -.I7 - .20
- .04 .6
,i.O3) i.03) i.03) (2.4) i.05) (.05)
i.04) (4.7)
1983 -.21 -.l5 - .02 1 -.I3 -.I1
.04 - 3.3
i.03) (.03) i.03) (2.7) i.06) i.05)
i.04) (4.7)
1984 -.lo - .05 - .04 2.1 - .04 - .03
.05 1
(.03) (.03) i.03) (2.4) i.06) i.05)
(.04) (4.7)
1985 -.05 -.01 - .06 - 5.5 .18 .09 .OO -4.7
(.04) i.04) i.04) (2.6) i.07) ~ 0 7 ) i.06) (5.6)
Notes: Low-education blacks are those with less than 12 years of completed education. Adjusted differences in
log \\.ages between blacks in Xliami and cotnparison cities are obtained from a linear regression model that
includes education, potential experience, and other control variables; see text. Wages are deflated by the
Consumer Price Index (1980= 100). "E1np.-Pop. Rate" refers to the employ~nent:populationratio. "Lnemp.
Rate" refers to the unemployment rate among those in the labor force.
population ratios and unemployment rates.Y ble 4, the series of unemployment rate dif-
Among all blacks, there is some evidence ferentials indicates a sharp downturn in la-
of a relative decline in the employment- bor market opportunities for blacks in 1982.
to-population ratio in Miami between 1979 Given the lag between the arrival of the 7 / 13

Dfferetzce ztz
Dfference 112



1979 -.I5 -.I2 .OO - 2.0 -.I3 -.I5
.03 - .8
i.03) i.03) (.03) (1.9) i.05) (.05)
(.04) (3.8)
1980 -.1G -.I2 .05 -7.1 - .07 - .07
.03 - 8.2
i.03) (.03) i.03) (1.6) i.05) i.05)
i.04) (3.5)
1981 -.I1 -.lo .02 - 3.0 - .05 -.I1
.04 - 7.7
i.03) i.03) i.03) (2.0) i.05) (.05)
i.04) (4.2)
1982 -.24 - .20 - ,015 3.3 -.I7 - .20
- .04 .6
,i.O3) i.03) i.03) (2.4) i.05) (.05)
i.04) (4.7)
1983 -.21 -.l5 - .02 1 -.I3 -.I1
.04 - 3.3
i.03) (.03) i.03) (2.7) i.06) i.05)
i.04) (4.7)
1984 -.lo - .05 - .04 2.1 - .04 - .03
.05 1
(.03) (.03) i.03) (2.4) i.06) i.05)
(.04) (4.7)
1985 -.05 -.01 - .06 - 5.5 .18 .09 .OO -4.7
(.04) i.04) i.04) (2.6) i.07) ~ 0 7 ) i.06) (5.6)

Dfferetzce ztz
Dfference 112



1979 -.I5 -.I2 .OO - 2.0 -.I3 -.I5
.03 - .8
i.03) i.03) (.03) (1.9) i.05) (.05)
(.04) (3.8)
1980 -.1G -.I2 .05 -7.1 - .07 - .07
.03 - 8.2
i.03) (.03) i.03) (1.6) i.05) i.05)
i.04) (3.5)
1981 -.I1 -.lo .02 - 3.0 - .05 -.I1
.04 - 7.7
i.03) i.03) i.03) (2.0) i.05) (.05)
i.04) (4.2)
1982 -.24 - .20 - ,015 3.3 -.I7 - .20
- .04 .6
,i.O3) i.03) i.03) (2.4) i.05) (.05)
i.04) (4.7)
1983 -.21 -.l5 - .02 1 -.I3 -.I1
.04 - 3.3
i.03) (.03) i.03) (2.7) i.06) i.05)
i.04) (4.7)
1984 -.lo - .05 - .04 2.1 - .04 - .03
.05 1
(.03) (.03) i.03) (2.4) i.06) i.05)
(.04) (4.7)
1985 -.05 -.01 - .06 - 5.5 .18 .09 .OO -4.7
(.04) i.04) i.04) (2.6) i.07) ~ 0 7 ) i.06) (5.6)
I Usado pela primeira vez por Abadie Gardeazabal (2003) para

medir o impacto de terrorismo no Pais Basco sobre crescimento
usando outras regies da Espanha como grupo de Controle
I Nenhuma outra regio da Espanha seguia uma tendncia

parecida ao Pais Basco
I Eles usam uma mdia ponderada de outras regies como grupo

de controle
10 / 13
The Basque Country is Dierent from the Rest of Spain
Waldinger (Warwick) 32 / 55
Detalhes
I Seja YitN o resultado observado para regio i no perodo t na

ausncia da interveno para unidades i = 1, ..., J + 1
I Seja YitI o resultado observado para regio i no perodo t que

afetada pela interveno
I Assumimos que antes da interveno YitN = YitI e que o efeito da

interveno : it = YitI YitN
I Temos informaes em T0 perodos pr-interveno de forma

que 1 T0 < T
11 / 13
Detalhes
I Vamos supor um modelo de fatores dado por:
YitN = t + t Zi + t i + it
I Consideremos um vetor de pesos W = (w2 , ..., wJ+1 )0 tal que

wj 0 e w2 + ... + wJ=1 = 1
I Podemos usar um estimador para o efeito da interveno dado

por:
J=1
X
1t = Y1t wj Yjt
j=2
12 / 13
The Synthetic Control Method
They have J available control regions (the 16 Spanish regions other

than the Basque Country).
They want to assign weights W = (w1 , ..., wJ )0 a (J x 1) to each
region. (wj 0 & wj = 1; this ensures that there is no
extrapolation outside the support of the growth predictors for the
control regions).
The weights are chosen so that the synthetic Basque country most
closely resembles the actual one before terrorism.
The Synthetic Control Method - Details
Let X1 be a (K x 1) vector of pre-terrorism of K economic growth

predictors (i.e. the values in the previous table: investment ratio,
population density, ...) in the Basque Country.
Let X0 be a (K x J) matrix which contains the values of the same
variables for the J possible control regions.
Let V be a diagonal matrix with nonnegative components reecting
the relative importance of the dierent growth predictors.
The vector of weights W* is then chosen to minimize:
(X1 X0 W)V(X1 X0 W)
They choose the matrix V such that the real per capita GDP path for
the Basque Country during the 1960s (pre terrorism) is best
reproduced by the resulting synthetic Basque Country.
The Synthetic Control Method - Details
The optimal weights they get are: Catalonia: 0.8508, Madrid: 0.1492,
and all other regions: 0.
Alternatively they could have just chosen the weights to reproduce
only the pre-terrorism growth path for the Basque country (and not
the growth predictors as well. In that case they would have minimized:
(Z1 Z0 W)(Z1 Z0 W)
Z1 is the (10 x 1) vector of pre-terrorism (1960-1969) GDP values for the

Basque Country.
Z0 is the (10 x J) vector of pre-terrorism (1960-1969) GDP values for the J
potential control regions.
The Synthetic Basque Country Looks Similar
Constructing the Counterfactual Using the Weights
Y1 is a (T x 1) vector whose elements are the values of real per

capital GDP values for T years in the Basque country.
Y0 is a (T x J) matrix whose elements are the values of real per
capital GDP values for T years in the control regions.
They then constructed the counterfactual GDP (in the absence of
terrorism) as: Y*1 =Y0 W*
Growth in the Basque Country with and without Terrorism
Terrorist Activity and Estimated GDP Gap
Synthetic Control Methods for Comparative Case
Studies: Estimating the Effect of Californias
Tobacco Control Program
Alberto A BADIE, Alexis D IAMOND, and Jens H AINMUELLER
Building on an idea in Abadie and Gardeazabal (2003), this article investigates the application of synthetic control methods to comparative
case studies. We discuss the advantages of these methods and apply them to study the effects of Proposition 99, a large-scale tobacco
control program that California implemented in 1988. We demonstrate that, following Proposition 99, tobacco consumption fell markedly
in California relative to a comparable synthetic control region. We estimate that by the year 2000 annual per-capita cigarette sales in
California were about 26 packs lower than what they would have been in the absence of Proposition 99. Using new inferential methods
proposed in this article, we demonstrate the significance of our estimates. Given that many policy interventions and events of interest in
social sciences take place at an aggregate level (countries, regions, cities, etc.) and affect a small number of aggregate units, the potential
applicability of synthetic control methods to comparative case studies is very large, especially in situations where traditional regression
methods are not appropriate.
KEY WORDS: Observational studies; Proposition 99; Tobacco control legislation; Treatment effects.
1. INTRODUCTION Comparing the evolution of an aggregate outcome (e.g.,

state-level crime rate) between a unit affected by the event or
Social scientists are often interested in the effects of events
intervention of interest and a set of unaffected units requires
or policy interventions that take place at an aggregate level and
only aggregate data, which are often available. However, when
affect aggregate entities, such as firms, schools, or geographic
data are not available at the same level of aggregation as the
or administrative areas (countries, regions, cities, etc.). To es-
outcome of interest, information on a sample of disaggregated
timate the effects of these events or interventions, researchers
units can sometimes be used to estimate the aggregate outcomes
often use comparative case studies. In comparative case stud-
of interest (like in Card 1990 and Card and Krueger 1994).
ies, researchers estimate the evolution of aggregate outcomes
Given the widespread availability of aggregate/macro data
(such as mortality rates, average income, crime rates, etc.) for a
(e.g., at the school, city, or region level), and the fact that many
unit affected by a particular occurrence of the event or interven-
policy interventions and events of interest in the social sciences
tion of interest and compare it to the evolution of the same ag-
take place at an aggregate level, comparative case study re-
gregates estimated for some control group of unaffected units.
search has broad potential. However, comparative case study
Card (1990) studies the impact of the 1980 Mariel Boatlift, a
research is limited in the social sciences by two problems that
large and sudden Cuban migratory influx in Miami, using other
affect its empirical implementation. First, in comparative case
cities in the southern United States as a comparison group. In a
studies there is typically some degree of ambiguity about how
well-known study of the effects of minimum wages on employ-
comparison units are chosen. Researchers often select compar-
ment, Card and Krueger (1994) compare the evolution of em-
ison groups on the basis of subjective measures of affinity be-
ployment in fast food restaurants in New Jersey and its neigh-
tween affected and unaffected units. Second, comparative case
boring state Pennsylvania around the time of an increase in New
studies typically employ data on a sample of disaggregated
Jerseys minimum wage. Abadie and Gardeazabal (2003) esti-
units and inferential techniques that measure only uncertainty
mate the effects of the terrorist conflict in the Basque Country
about the aggregate values of the data in the population. Un-
on the Basque economy using other Spanish regions as a com-
certainty about the values of aggregate variables can be elimi-
parison group.
nated completely if aggregate data are available. However, the
availability of aggregate data does not imply that the effect of
Alberto Abadie is Professor, John F. Kennedy School of Government, the event or intervention of interest can be estimated without
Harvard University, 79 John F. Kennedy Street, Cambridge, MA 02138 error. Even if aggregate data are employed, there remains un-
(E-mail: alberto_abadie@harvard.edu). Alexis Diamond is Evaluation Officer,
The International Finance Corporation, 2121 Pennsylvania Avenue, Washing- certainty about the ability of the control group to reproduce the
ton, DC 20433 (E-mail: adiamond@ifc.org). Jens Hainmueller is Assistant Pro- counterfactual outcome trajectory that the affected units would
fessor, Department of Political Science, Massachusetts Institute of Technology, have experienced in the absence of the intervention or event of
77 Massachusetts Avenue, Cambridge, MA 02139 (E-mail: jhainm@mit.edu).
In addition, during the course of this study all authors were affiliated with Har-
interest. This type of uncertainty is not reflected by the stan-
vards Institute for Quantitative Social Science (IQSS). We thank Joshua An- dard errors constructed with traditional inferential techniques
grist, Jake Bowers, Javier Gardeazabal, Bryan Graham, Dan Hopkins, Rustam for comparative case studies.
Ibragimov, Guido Imbens, Michael Lovenheim, Paul Rosenbaum, Don Rubin, This article addresses current methodological shortcomings
Jas Sekhon, Gary Solon, Chris Winship, and seminar participants at Harvard,
Syracuse, Uppsala, the NBER Labor Meetings, and the 2006 APSA Meetings of case study analysis. We advocate the use of data-driven pro-
in Philadelphia for helpful comments. Funding for this research was gener- cedures to construct suitable comparison groups, as in Abadie
ously provided by NSF grant SES-0350645 (Abadie). The views expressed in
this article are not necessarily those of IFC. This article was awarded the 2007
Gosnell Prize for the best work in methods presented at any political science 2010 American Statistical Association
conference during the preceding year. Journal of the American Statistical Association
Companion software developed by the authors (Synth package for MAT- June 2010, Vol. 105, No. 490, Applications and Case Studies
LAB, R, and Stata) is available at the authors webpages. DOI: 10.1198/jasa.2009.ap08746
493
Abadie, Diamond, and Hainmueller: Synthetic Control Methods for Comparative Case Studies 499
programs in the 19892000 period and they are excluded from

the donor pool. We also discard all states that raised their state
cigarette taxes by 50 cents or more over the 1989 to 2000 pe-
riod (Alaska, Hawaii, Maryland, Michigan, New Jersey, New
York, Washington). Notice that, even if smaller tax increases
substantially reduced smoking in any of the control states that
gets assigned a positive weight in the synthetic control, this
should if anything attenuate the treatment effect estimate that
we obtain for California. Finally, we also exclude the District
of Columbia from our sample. Our donor pool includes the
remaining 38 states. Our results are robust, however, to the in-
clusion of the discarded states.
Our outcome variable of interest is annual per capita ciga-
rette consumption at the state level, measured in our dataset as
per capita cigarette sales in packs. We obtained these data from
Orzechowski and Walker (2005) where they are constructed us-
ing information on state-level tax revenues on cigarettes sales.
This is the most widely used indicator in the tobacco research
literature, available for a much longer time period than survey-
Figure 1. Trends in per-capita cigarette sales: California vs. the rest
based measures of smoking prevalence. A disadvantage of tax-
of the United States.
revenue-based data relative to survey data on smoking preva-
lence is that the former are affected by cigarette smuggling
across tax jurisdictions. We discuss this issue later in this sec- United States relative to California. Following the laws pas-
tion. We include in X1 and X0 the values of predictors of sage, cigarette consumption in California continued to decline.
smoking prevalence for California and the 38 potential con- To evaluate the effect of Proposition 99 on cigarette smoking
trols, respectively. Our predictors of smoking prevalence are: in California, the central question is how cigarette consumption
average retail price of cigarettes, per capita state personal in- would have evolved in California after 1988 in the absence of
come (logged), the percentage of the population age 1524, and Proposition 99. The synthetic control method provides a sys-
per capita beer consumption. These variables are averaged over tematic way to estimate this counterfactual.
the 19801988 period and augmented by adding three years of As explained above, we construct the synthetic California as
lagged smoking consumption (1975, 1980, and 1988). Appen- the convex combination of states in the donor pool that most
dix A provides data sources. closely resembled California in terms of pre-Proposition 99 val-
Using the techniques described in Section 2, we construct ues of smoking prevalence predictors. The results are displayed
a synthetic California that mirrors the values of the predictors in Table 1, which compares the pretreatment characteristics of
of cigarette consumption in California before the passage of the actual California with that of the synthetic California, as
Proposition 99. We estimate the effect of Proposition 99 on per well as with the population-weighted average of the 38 states
capita cigarette consumption as the difference in cigarette con- in the donor pool. We see that the average of states that did not
sumption levels between California and its synthetic versions implement a large-scale tobacco-control program in 19892000
in the years after Proposition 99 was passed. We then perform a does not seem to provide a suitable control group for Califor-
series of placebo studies that confirm that our estimated effects nia. In particular, prior to the passage of Proposition 99 average
for California are unusually large relative to the distribution of beer consumption and cigarette retail prices were lower in the
the estimate that we obtain when we apply the same analysis to average of the 38 control states than in California. Moreover,
the states in the donor pool. prior to the passage of Proposition 99 average cigarette sales
per capita were substantially higher on average in the 38 con-
3.3 Results
Figure 1 plots the trends in per capita cigarette consumption Table 1. Cigarette sales predictor means
in California and the rest of the United States. As this figure
suggests, the rest of the United States may not provide a suit- California
Average of
able comparison group for California to study the effects of
Variables Real Synthetic 38 control states
Proposition 99 on per capita smoking. Even before the passage
of Proposition 99 the time series of cigarette consumption in Ln(GDP per capita) 10.08 9.86 9.86
California and in the rest of the United States differed notably. Percent aged 1524 17.40 17.40 17.29
Levels of cigarette consumption were similar in California and Retail price 89.42 89.41 87.27
the rest of the United States in the early 1970s. Trends began to Beer consumption per capita 24.28 24.20 23.75
Cigarette sales per capita 1988 90.10 91.62 114.20
diverge in the late 1970s, when Californias cigarette consump-
tion peaked and began to decline while consumption in the rest
of the United States was still rising. Cigarette sales declined in
the 1980s, but with larger decreases in California than in the rest NOTE: All variables except lagged cigarette sales are averaged for the 19801988 period
(beer consumption is averaged 19841988). GDP per capita is measured in 1997 dollars,
of the United States. In 1988, the year Proposition 99 passed, retail prices are measured in cents, beer consumption is measured in gallons, and cigarette
cigarette consumption was about 27% higher in the rest of the sales are measured in packs.
500 Journal of the American Statistical Association, June 2010
trol states than in California. In contrast, the synthetic Califor-

nia accurately reproduces the values that smoking prevalence
and smoking prevalence predictor variables had in California
prior to the passage of Proposition 99.
Table 1 highlights an important feature of synthetic control
estimators. Similar to matching estimators, the synthetic con-
trol method forces the researcher to demonstrate the affinity be-
tween the region exposed to the intervention of interest and its
synthetic counterpart, that is, the weighted average of regions
chosen from the donor pool. As a result, the synthetic control
method safeguards against estimation of extreme counterfactu-
als, that is, those counterfactuals that fall far outside the convex
hull of the data (King and Zheng 2006). As explained in Sec-
tion 2.3, we chose V among all positive definite and diagonal
matrices to minimize the mean squared prediction error of per
capita cigarette sales in California during the pre-Proposition 99
period. The resulting value of the diagonal element of V asso-
ciated to the log GDP per capita variable is very small, which
indicates that, given the other variables in Table 1, log GDP Figure 2. Trends in per-capita cigarette sales: California vs. syn-
per capita does not have substantial power predicting the per thetic California.
capita cigarette consumption in California before the passage
of Proposition 99. This explains the discrepancy between Cali-
that, in contrast to per capita sales in other U.S. states (shown
fornia and its synthetic version in terms of log GDP per capita.
in Figure 1), per capita sales in the synthetic California very
Table 2 displays the weights of each control state in the syn-
thetic California. The weights reported in Table 2 indicate that closely track the trajectory of this variable in California for the
smoking trends in California prior to the passage of Proposi- entire pre-Proposition 99 period. Combined with the high de-
tion 99 is best reproduced by a combination of Colorado, Con- gree of balance on all smoking predictors (Table 1), this sug-
necticut, Montana, Nevada, and Utah. All other states in the gests that the synthetic California provides a sensible approxi-
donor pool are assigned zero W-weights. mation to the number of cigarette packs per capita that would
Figure 2 displays per capita cigarette sales for California and have been sold in California in 19892000 in the absence of
its synthetic counterpart during the period 19702000. Notice Proposition 99.
Our estimate of the effect of Proposition 99 on cigarette con-
sumption in California is the difference between per capita ciga-
Table 2. State weights in the synthetic California
rette sales in California and in its synthetic version after the pas-
State Weight State Weight sage of Proposition 99. Immediately after the laws passage, the
two lines begin to diverge noticeably. While cigarette consump-
Alabama 0 Montana 0.199
tion in the synthetic California continued on its moderate down-
Alaska Nebraska 0
Arizona Nevada 0.234 ward trend, the real California experienced a sharp decline. The
Arkansas 0 New Hampshire 0 discrepancy between the two lines suggests a large negative ef-
Colorado 0.164 New Jersey fect of Proposition 99 on per capita cigarette sales. Figure 3
Connecticut 0.069 New Mexico 0 plots the yearly estimates of the impacts of Proposition 99, that
Delaware 0 New York is, the yearly gaps in per capita cigarette consumption between
District of Columbia North Carolina 0 California and its synthetic counterpart. Figure 3 suggests that
Florida North Dakota 0 Proposition 99 had a large effect on per capita cigarette sales,
Georgia 0 Ohio 0 and that this effect increased in time. The magnitude of the es-
Hawaii Oklahoma 0
timated impact of Proposition 99 in Figure 3 is substantial. Our
Idaho 0 Oregon
Illinois 0 Pennsylvania 0
results suggest that for the entire 19892000 period cigarette
Indiana 0 Rhode Island 0 consumption was reduced by an average of almost 20 packs per
Iowa 0 South Carolina 0 capita, a decline of approximately 25%.
Kansas 0 South Dakota 0 In order to assess the robustness of our results, we included
Kentucky 0 Tennessee 0 additional predictors of smoking prevalence among the vari-
Louisiana 0 Texas 0 ables used to construct the synthetic control. Our results stayed
Maine 0 Utah 0.334 virtually unaffected regardless of which and how many predic-
Maryland Vermont 0 tor variables we included. The list of predictors used for robust-
Massachusetts Virginia 0 ness checks included state-level measures of unemployment,
Michigan Washington
income inequality, poverty, welfare transfers, crime rates, drug
Minnesota 0 West Virginia 0
Mississippi 0 Wisconsin 0
related arrest rates, cigarette taxes, population density, and nu-
Missouri 0 Wyoming 0 merous variables to capture the demographic, racial, and social
structure of states.
502 Journal of the American Statistical Association, June 2010
Figure 4. Per-capita cigarette sales gaps in California and placebo Figure 5. Per-capita cigarette sales gaps in California and placebo
gaps in all 38 control states. gaps in 34 control states (discards states with pre-Proposition 99
MSPE twenty times higher than Californias).
provide a good fit for per capita cigarette consumption prior

to Proposition 99 for the majority of the states in the donor Figure 6 is based on a lower cutoff, excluding all states that
pool. However, Figure 4 indicates also that per capita cigarette had a pre-Proposition 99 MSPE of more than five times the
sales during the 19701988 period cannot be well reproduced MSPE of California. Twenty-nine control states plus California
for some states by a convex combination of per capita ciga- remain in the figure. The California gap line is now clearly the
rette sales in other states. The state with worst fit in the pre- most unusual line for almost the entire post-treatment period.
Proposition 99 period is New Hampshire, with a MSPE of 3437. In Figure 7 we lower the cutoff even further and focus
The large MSPE for New Hampshire does not come as a sur- exclusively on those states that we can fit almost as well
prise. Among all the states in the donor pool, New Hampshire as California in the period 19701988, that is, those states
is the state with the highest per capita cigarette sales for every with pre-Proposition 99 MSPE not higher than twice the pre-
Proposition 99 MSPE for California. Evaluated against the dis-
year prior to the passage of Proposition 99. Therefore, there is
tribution of the gaps for the 19 remaining control states in Fig-
no combination of states in our sample that can reproduce the
ure 7, the gap for California appears highly unusual. The nega-
time series of per capita cigarette sales in New Hampshire prior
tive effect in California is now by far the lowest of all. Because
to 1988. Similar problems arise for other states with extreme
this figure includes 19 control states, the probability of estimat-
values of per capita cigarette sales during the pre-Proposition 99
period.
If the synthetic California had failed to fit per capita ciga-
rette sales for the real California in the years before the pas-
sage of Proposition 99, we would have interpreted that much
of the post-1988 gap between the real and the synthetic Cal-
ifornia was also artificially created by lack of fit, rather than
by the effect of Proposition 99. Similarly, placebo runs with
poor fit prior to the passage of Proposition 99 do not provide
information to measure the relative rarity of estimating a large
post-Proposition 99 gap for a state that was well fitted prior
to Proposition 99. For this reason, we provide several different
versions of Figure 4, each version excluding states beyond a
certain level of pre-Proposition 99 MSPE.
Figure 5 excludes states that had a pre-Proposition 99 MSPE
of more than 20 times the MSPE of California. This is a very
lenient cutoff, discarding only four states with extreme values
of pre-Proposition 99 MSPE for which the synthetic method
would be clearly ill-advised. In this figure there remain a few
lines that still deviate substantially from the zero gap line in the
pre-Proposition 99 period. Among the 35 states remaining in Figure 6. Per-capita cigarette sales gaps in California and placebo
the figure, the California gap line is now about the most unusual gaps in 29 control states (discards states with pre-Proposition 99
line, especially from the mid-1990s onward. MSPE five times higher than Californias).
Revisitando os Marielitos
I Borjas (2016)- outros grupos de comparao
I Peri e Yasenov (2016) - controle sinttico
13 / 13
The Wage Impact of the Marielitos: A Reappraisal
George J. Borjas
Harvard University
July 2016

32
Figure 1. Number of Cuban immigrants, by year of migration, 1955-2010

120,000
110,000
100,000
90,000
80,000
Number of immigrants
70,000
60,000
50,000
40,000
30,000
20,000
10,000
0
1950 1960 1970 1980 1990 2000 2010
Year of migra1on

Notes: The specific year of migration (through 1999) is first reported in the 2000 census. The counts are
adjusted for mortality and out-migration by using information on the number of arrivals provided by the
1970 through 1990 censuses; see the text for details. The 2000-2008 counts are drawn from the pooled 2009-
2011 American Community Surveys (ACS), while the 2009-2010 counts are drawn from the 2012 ACS.

43
Table 4. Rates of employment and wage growth before Mariel

Employment Employment growth: Wage growth: high
Rank Metropolitan area growth: all workers high school dropouts school dropouts
1 San Diego, CA 0.194 0.067 -0.127
2 Greensboro-Winston Salem, NC 0.182 -0.063 -0.336
3 Kansas City, MO/KS 0.179 0.052 -0.178
4 Anaheim-Santa Ana- Garden Grove, CA 0.162 0.257 0.068
5 Rochester, NY 0.153 -0.172 0.019
6 Miami-Hialeah, FL 0.153 0.086 0.012
7 Nassau-Suffolk, NY 0.151 0.056 -0.047
8 San Jose, CA 0.137 0.130 0.138
9 Albany-Schenectady-Troy, NY 0.130 0.065 0.065
10 Boston, MA 0.121 -0.100 -0.024
11 Milwaukee, WI 0.121 -0.006 0.056
12 Indianapolis, IN 0.115 0.071 -0.048
13 Seattle-Everett, WA 0.110 -0.079 -0.051
14 Norfolk-Virginia Beach-Newport News, VA 0.103 0.052 0.110
15 Philadelphia, PA/NJ 0.102 -0.033 0.012
16 Newark, NJ 0.092 -0.116 -0.124
17 Tampa-St. Petersburg-Clearwater, FL 0.083 0.068 0.120
18 Denver-Boulder-Longmont, CO 0.082 -0.139 0.027
19 Houston-Brazoria, TX 0.078 0.090 0.003
20 Sacramento, CA 0.078 0.152 -0.034
21 Dallas-Fort Worth, TX 0.076 0.062 -0.050
22 Portland-Vancouver, OR/WA 0.071 -0.074 -0.020
23 Riverside-San Bernardino, CA 0.071 -0.017 0.254
24 Atlanta, GA 0.069 -0.087 0.014
25 Cincinnati-Hamilton, OH/KY/IN 0.063 0.038 -0.032
26 Washington, DC/MD/VA 0.061 0.028 0.090
27 Detroit, MI 0.060 -0.099 0.010
28 Fort Worth-Arlington, TX 0.058 -0.006 -0.059
29 Los Angeles-Long Beach, CA 0.056 0.075 -0.128
30 Columbus, OH 0.048 -0.324 -0.004
31 Buffalo-Niagara Falls, NY 0.039 0.040 -0.144
32 Chicago-Gary-Lake IL 0.025 -0.082 -0.017
33 St. Louis, MO/IL 0.019 -0.060 -0.060
34 Bergen-Passaic, NJ 0.015 -0.051 0.007
35 Baltimore, MD 0.012 -0.108 -0.001
36 Minneapolis-St. Paul, MN 0.007 -0.050 -0.012
37 Cleveland, OH 0.001 -0.071 -0.014
38 New York, NY 0.000 -0.146 0.074
39 Pittsburg, PA -0.013 -0.111 0.123
40 Birmingham, AL -0.020 -0.172 -0.124
41 San Francisco-Oakland-Vallejo, CA -0.027 -0.200 -0.105
42 Gary-Hammond-East Chicago, IN -0.029 0.119 0.042
43 New Orleans, LA -0.046 -0.313 -0.017
44 Akron, OH -0.110 -0.351 -0.005

Notes: The rate of employment growth is the log ratio of average employment in 1979-1980 to average employment in
1977-1978, calculated from the 1977-1980 survey years of the March CPS. The rate of wage growth is the difference in
the (age-adjusted) log weekly wage between 1978-1979 and 1976-1977.

33
Figure 2. Log wage of high school dropouts, 1972-2003

(95 percent confidence band)

5.8
Outside Miami
5.6
Log weekly wage
5.4
5.2
Miami
5
4.8
1972 1976 1980 1984 1988 1992 1996 2000 2004
Year

Notes: The log weekly wage is a 3-year moving average of the average log wage of high school dropouts in
each geographic area. The data are drawn from the March CPS files.

34
Figure 3. Trends in the wage of low-skill workers in the March CPS, 1977-1992

A. Log weekly wage of high school dropouts
5.9
Synthetic placebo
5.8
Employment placebo
5.7
5.6
Log weekly wage

Card placebo
5.5
Low-skill placebo
5.4
5.3
5.2
Miami
5.1
5
1976 1978 1980 1982 1984 1986 1988 1990 1992
Year
B. Log wage of high school dropouts relative to college graduates

-0.2
Employment placebo
-0.4 Synthetic placebo
Log wage gap
-0.6
Card placebo
-0.8
Low-skill placebo
-1
Miami
-1.2
1976 1978 1980 1982 1984 1986 1988 1990 1992
Year
C. Log wage of high school dropouts relative to high school graduates

0.1
Synthetic placebo
0
Employment placebo
-0.1
Log wage gap
-0.2
-0.3
Card placebo
-0.4
Low-skill placebo
Miami
-0.5
-0.6
1976 1978 1980 1982 1984 1986 1988 1990 1992
Year

Notes: The figures use a 3-year moving average of the average log wage of high school dropouts, high school
graduates, and college graduates in each specific geographic area.

The Labor Market Eects of a Refugee Wave:
Synthetic Control Method meets the Mariel Boatlift
Giovanni Peri (University of California, Davis)
Vasil Yasenov (University of California, Davis)
May 10, 2016
Abstract
We apply the synthetic control method to re-examine the wage and employment
eect of the Mariel Boatlift, a large inflow of Cuban refugees to Miami in 1980. This
method improves on previous studies by choosing a control group for Miami so as to
best match its labor market features in the eight years before the Boatlift. Given the
presence of significant measurement error for average city wages we emphasize the
importance of using the May-ORG CPS sample rather than the March-CPS. The
first includes a more reliable measure of weekly wages, has larger sample size and

We thank David Card, Colin Cameron, Thomas Lemieux, David Green, Michael Clemens, Pat Kline,
Doug Miller, Joan Monras, Enrico Moretti, David Roodman, Shu Shen and participants in seminars at
NBER winter meetings, University of British Columbia and Georgetown University for useful comments.
We are grateful for the criticism contained in Borjas (2015b) to the first draft of this paper, as it has
helped our current revision. We have not received any financial support for this paper.

Giovanni Peri: Department of Economics, UC Davis, One Shields Avenue, Davis CA, 95616; email:
gperi@ucdavis.edu.

Vasil Yasenov: Department of Economics, UC Davis, One Shields Avenue, Davis CA, 95616; email:
viyasenov@ucdavis.edu.
1
Figure 3:
Log Wage Measures, High School dropouts Miami and Synthetic Control, 1973-1991
May + ORG CPS, Preferred Sample
Note: Each Panel shows the outcome variable for Miami (solid line) and Synthetic control (dashed line) in the period 1972-1991. The outcome shown and the sample used are noted in the title of
each panel (A through D). Preferred sample means: non-Cubans, high school dropouts, not self-employed, in the labor force. The age group varies upon panel. The vertical line is drawn between the
data points of 1979 and 1980 and it identifies the interval in which the Mariel Boatlift took place. The cities with positive weight in the synthetic control are as follows. Panel A: New Orleans, LA
43.2%, New York City, NY, 30.1%, Baltimore, MD 24.9%; Panel B: Cincinnati, OH, 35.5%, New York City, NY, 26.5%, Tampa Bay-St. Petersburg, FL 23.2%, Sacramento, CA 14.7%; Panel C: New Orleans,
LA 43.9%, New York City, NY, 29.9%, Baltimore, MD 24.8%; Panel D: Tampa Bay-St. Petersburg, FL 45.5%, Sacramento, CA 44%, Los Angeles, CA 7.5%, Greensboro, SC 3.1%.
41
Figure 4:
Regression Adjusted Log Wage Measures, High school dropouts, Miami and Synthetic Control, 1972-1991
Regression Adjusted, May + ORG CPS, Preferred Sample
Note: Each Panel shows the outcome variable for Miami (solid line) and Synthetic control (dashed line) in the period 1972-1991. The outcome shown and the sample used are noted in the title of
each panel (A through D). Preferred sample means: non-Cubans, high school dropouts, not self-employed in the labor force. The cities with positive weight in the synthetic control are as follows.
Panel A: Tampa-St. Petersburg, FL 52%, Birmingham, AL 40.5%, New York City, NY 7.5%; Panel B: Tampa-St. Petersburg, FL 45.4%, New York City, NY 24.6%, Sacramento, CA 20.8%, Albany-
Schenectady-Troy, NY 9.2%; Panel C: Tampa-St. Petersburg, FL 54%, Birmingham, AL 46%; Panel D: Greensboro, NC 48.3%, Norfold-Portsmouth, VA 30.5%, Cincinnati, OH 21.2%.
42
Figure 5:
Additional Labor Market Outcomes: Log Weekly Wages at different percentiles and Dropouts Unemployment
Miami and Synthetic Control, 1973-1991, May + ORG CPS, Preferred Sample
Note: Each Panel shows the outcome variable for Miami (solid line) and Synthetic control (dashed line) in the period 1972-1991. The outcome showed for each Panel and the sample used are noted
in the title of each panel. Preferred sample means: non-Cubans, not self-employed individuals, in the labor force, age 19-65 for Panels A, B and C and non-Cubans, high school dropouts not self-
employed, in the labor force, age 19-65 for Panel D. The cities with positive weight in the synthetic control are as follows. Panel A: Birmingham, AL 60.6%, Rochester, NY 28.6%, Nassau-Suffolk, NY
10.4%; Panel B: San Diego, CA 57.7%, Birmingham, AL 28.4%, Nassau-Suffolk, NY 13.8%; Panel C: Tampa-St Petersburg, FL 67.3%, Nassau-Suffolk, NY 32.7%; Panel D: New Orleans, LA 48.4%, New
York City, NY 30.9%, Albany-Schenectady-Troy, NY 19.5% Cincinnati, OH 1.1%
43
Combining DD and IV
Sometimes combining DD and IV methods can be quite useful.

In a recent paper (Waldinger, 2010), I have done that to estimate the
eect of faculty quality on the outcomes of PhD students.
Estimating the eect of faculty quality on PhD student outcomes is
challenging because of:
1 Selection of good students into good universities.
2 Omitted variables aecting both faculty quality and student outcomes.
3 Measurement error in faculty quality.
I address these issues by using the dismissal of scientists in Nazi
Germany as an exogenous shock to faculty quality.
The dismissal aected some departments very strongly, while other
departments were not aected.
Historical Background
Germany was the leading country for scientic research at the

beginning of the 20th century.
Immediately after gaining power in 1933 the new Nazi government
dismissed all Jewish and politically unreliablescholars from the
German universities.
Dismissed Professors Across German Universities
Dismissed Professors Across German Universities II
Eect of Dismissals on Department Size
Eect of Dismissals on Faculty Quality
Panel Data on PhD graduates from German Universities
I obtained a panel dataset of all mathematics PhD students

graduating from all German universities between 1923 and 1938 and
use the dismissal as exogenous variation in faculty quality.
The empirical strategy essentially compares changes in outcomes of
PhD students in aected department before and after 1933 to
changes in outcomes in unaected departments.
I investigate the following outcomes:
1 Whether former PhD student publishes dissertation in a top journal.
2 Whether former PhD student ever becomes full professor.
3 # of lifetime citations.
4 Positive lifetime citations.
Reduced Form Graphical Analysis - Publishing Dissertation
Reduced Form Graphical Analysis - Full Professor
Reduced Form Graphical Analysis - Lifetime Citations
Reduced Form Estimates
The reduced form of the dismissal eect is essentially a DD estimator.
Outcomeidt = 1 + 2 (Dismissal induced Reduction in Faculty Quality)dt

+ 3 (Dismissal induced increase in Student/Faculty Ratio)dt
+ 4 Femaleidt + 5 Foreignidt + 6 CohortFEt + 5 DepFEd + idt
Dismissal induced Reduction in Faculty Quality is 0 until 1933 and

equal to the dismissal induced fall in faculty quality after 1933 (and
remains 0 in departments without dismissals).
Dismissal induced increase in Student/Faculty Ratio is also 0 until
1933 but equal to the dismissal induced increase in student/faculty
ratio after 1933
! essentially a dierences-in-dierences estimator but with dierent
treatment intensities.
Reduced Form Estimates
Common Robustness Check for Parallel Trend Assumption
Only Look at Pre-Period Data and Move Placebo Treatment some Years Back
Here I move a placebo treatment to 1930.
Use Dismissal as IV
OLS model to the eect of university quality on PhD student

outcomes:
Outcomeidt = 1 + 2 (Avg. Faculty Quality)dt 1
+ 3 (Student/Faculty Ratio)dt 1
University quality and student/faculty ratio are endogenous ! use
dismissal as IV.
2 Endogenous Variables ! 2 First Stage Regressions:
1 Avg. Faculty Qualityidt = 1
+ 2 (Dismissal induced Reduction in Faculty Quality)dt
2 Student/Faculty Ratioidt = 1
+ 2 (Dismissal induced Reduction in Faculty Quality)dt
First Stages
To test for weak instruments one cannot simply look at the rst stage
F-statistics because here we have 2 endogenous regressors and 2 IVs.
! use Cragg-Donald EV statistic here critical value is 7.03.
OLS and IV

EconometriaAplic - Aula6 (Diferencas em Diferencas) Ferraz PUC

Diunggah oleh

Informasi Dokumen

Judul Asli

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

EconometriaAplic - Aula6 (Diferencas em Diferencas) Ferraz PUC

Diunggah oleh

Hak Cipta:

Format Tersedia

Econometria Aplicada

Usando variacao temporal

Suponha que estamos interessados em estimar o efeito de um

Usando variacao temporal

Suponha que estamos interessados em estimar o efeito de um

Na impossibilidade de obter variacao aleatorizada em Di e sabendo

Usando variacao temporal

Suponha que estamos interessados em estimar o efeito de um

Na impossibilidade de obter variacao aleatorizada em Di e sabendo

Na pratica, encontrar uma variavel instrumental valida nao e uma

Usando variacao temporal

Suponha que estamos interessados em estimar o efeito de um

Na impossibilidade de obter variacao aleatorizada em Di e sabendo

Na pratica, encontrar uma variavel instrumental valida nao e uma

Nestas aulas iremos explorar variacao temporal, ou entre coortes, para

Efeitos fixos individuais

Comecamos com um exemplo. Trabalhadores que tem seus salarios

Efeitos fixos individuais

Comecamos com um exemplo. Trabalhadores que tem seus salarios

Seja Yit Log da renda e Dit um indicador que e igual a 1 se a pessoal

Efeitos fixos individuais

Comecamos com um exemplo. Trabalhadores que tem seus salarios

Seja Yit Log da renda e Dit um indicador que e igual a 1 se a pessoal

Temos um vetor de variaveis Ai (habilidade) que nao observamos,

E [Y0it |Ai , Xit , t, Dit ] = E [Y0it |Ait , Xit , t]

Efeitos fixos individuais

Podemos escrever o resultado de Y0it como:

E [Y0it |Ai , Xit , t] = + t + A0i + Xit0

Efeitos fixos individuais

Podemos escrever o resultado de Y0it como:

E [Y0it |Ai , Xit , t] = + t + A0i + Xit0

Por ultimo, iremos assumir que o efeito causal e aditivo e constante:

E [Y1it |Ai , Xit , t] = E [Y0it |Ai , Xit , t] +

Efeitos fixos individuais

Podemos escrever o resultado de Y0it como:

E [Y0it |Ai , Xit , t] = + t + A0i + Xit0

Por ultimo, iremos assumir que o efeito causal e aditivo e constante:

E [Y1it |Ai , Xit , t] = E [Y0it |Ai , Xit , t] +

Estes supostos implicam num modelo dado por:

E [Yit |Ai , Xit , t, Dit ] = + t + Dit + A0i + Xit0

Efeitos fixos individuais

Podemos escrever o resultado de Y0it como:

E [Y0it |Ai , Xit , t] = + t + A0i + Xit0

Por ultimo, iremos assumir que o efeito causal e aditivo e constante:

E [Y1it |Ai , Xit , t] = E [Y0it |Ai , Xit , t] +

Estes supostos implicam num modelo dado por:

E [Yit |Ai , Xit , t, Dit ] = + t + Dit + A0i + Xit0

E uma equacao estimada na forma:

Yit = i + t + Dit + Xit0 + it

Efeitos fixos individuais

Efeitos fixos individuais

Uma alternativa, que nos da um coeficiente identico, e utilizar desvios

(Yit Yi ) = (t t ) + (Dit Di ) + (Xit Xi )0 + (it i )

Efeitos fixos individuais

Uma alternativa, que nos da um coeficiente identico, e utilizar desvios

(Yit Yi ) = (t t ) + (Dit Di ) + (Xit Xi )0 + (it i )

Uma terceira alternativa para a eliminacao do efeito-fixo i e tirar a

Yit = t + Dit + Xit + it

Efeitos fixos individuais

Uma alternativa, que nos da um coeficiente identico, e utilizar desvios

(Yit Yi ) = (t t ) + (Dit Di ) + (Xit Xi )0 + (it i )

Uma terceira alternativa para a eliminacao do efeito-fixo i e tirar a

Yit = t + Dit + Xit + it