Health Care Manag Sci (2011) 14:115124

DOI 10.1007/s10729-010-9147-2

A model for allocating CDCs HIV prevention resources

in the United States
Arielle Lasry & Stephanie L. Sansom &
Katherine A. Hicks & Vladislav Uzunangelov

Received: 30 December 2009 / Accepted: 29 November 2010 / Published online: 24 December 2010
# Springer Science+Business Media, LLC (outside the USA) 2010

Abstract The Division of HIV/AIDS Prevention (DHAP)

at the Centers for Disease Control and Prevention has an
annual budget of approximately $325 million for funding
HIV prevention programs in the U.S. The purpose of this
paper is to thoroughly describe the methods used to develop
a national HIV resource allocation model intended to
inform DHAP on allocation strategies that might improve
the overall effectiveness of HIV prevention efforts. The
HIV prevention resource allocation problem consists of
choosing how to apportion prevention resources among
interventions and populations so that HIV incidence is
minimized, given a budget constraint. We developed an
epidemic model that projects HIV infections over time
given a specific allocation scenario. The epidemic model is
then embedded in a nonlinear mathematical optimization
program to determine the allocation scenario that minimizes
HIV incidence over a 5-year horizon. In our model, we
consider the general U.S. population and specific at-risk
populations. The at-risk populations include 15 subgroups
structured by gender, race/ethnicity and HIV transmission
risk group. HIV transmission risk groups include high-risk
heterosexuals, men who have sex with men and injection
drug users. We consider HIV screening interventions and
interventions to reduce HIV-related risk behaviors. The
output of the model is the optimal funding scenario
A. Lasry (*) : S. L. Sansom
Division of HIV/AIDS Prevention,
Centers for Disease Control and Prevention,
1600 Clifton Road, Mailstop E-48,
Atlanta, GA 30333, USA
e-mail: alasry@cdc.gov
K. A. Hicks : V. Uzunangelov
RTI International,
Research Triangle Park, NC, USA

indicating the amounts to be allocated to all combinations

of populations and interventions. For illustrative purposes
only, we provide a sample application of the model. In this
example, the optimal allocation scenario is compared to the
current baseline funding scenario to highlight how the
current allocation of funds could be improved. In the
baseline allocation, 29% of the annual budget is aimed at
the general population, while the model recommends
targeting 100% of the budget to the at-risk populations
with no allocation targeted to the general population.
Within the allocation to behavioral interventions the model
recommends an increase in targeting diagnosed positives.
Also, the model allocation suggests a greater focus on
MSM and IDUs with a 72% of the annual budget allocated
to them, while the baseline allocation for MSM and IDUs
totals 37%. Incorporating future epidemic trends in the
decision-making process informs the selection of populations and interventions that should be targeted. Improving
the use of funds by targeting the interventions and
population subgroups at greatest risk may lead to improved
HIV outcomes. These models can also direct research by
pointing to areas where the development of cost-effective
interventions can have the most impact on the epidemic.
Keywords HIV/AIDS . Resource allocation .
Optimization model

1 Background
The Division of HIV/AIDS Prevention (DHAP) at the
Centers for Disease Control and Prevention (CDC) has an
annual budget of approximately $325 million for funding
HIV prevention programs domestically; this budget has
been stable over the past several years. Approximately two-

116

thirds of DHAPs extramural program budget is allocated to

state and large metropolitan health departments, who in turn
reallocate the funds to HIV programs within their jurisdictions. The remaining one third is allocated to
community-based organizations that provide HIV prevention services. Overall, in the United States, the annual
number of HIV infections has been stable at approximately
56,000 since the year 2000 [1].
We created an HIV resource allocation model that can
generate the mathematically optimal allocation of
DHAPs budget for funding HIV prevention programs
domestically. The model assists DHAP in estimating the
number of HIV infections prevented in the United States
with its current allocation of prevention dollars and
evaluate whether improving the allocation of its HIV
prevention funds can further reduce the number of new
HIV infections.
The purpose of this paper is to thoroughly describe the
methods used to develop a national HIV resource allocation
model intended to inform DHAP on allocation strategies
that might improve the overall effectiveness of HIV
prevention efforts. This paper is arranged as follows: we
begin with an overview of the structure of the model. Then,
we describe the underlying epidemic projection model,
describe how HIV interventions are embedded in the model
and present the optimization model. Finally, we provide an
illustrative example of the models application before a
brief discussion.

2 Methods
The HIV prevention resource allocation problem consists of
choosing how to apportion prevention resources among
interventions and populations so that HIV incidence is
minimized, given a budget constraint. We developed an
epidemic model that projects HIV infections over time
given a specific allocation scenario. The epidemic model is
embedded in a nonlinear mathematical optimization program to determine the allocation scenario that minimizes
HIV incidence over a 5-year horizon.
We consider two types of interventions: interventions
to reduce HIV-related risk behaviors, and HIV screening
interventions. Both intervention types are targeted to the
U.S. population aged 13 to 64 years that is at risk for
HIV transmission, according to gender, race/ethnicity and
HIV transmission risk group, and more broadly to the
general U.S. population aged 13 to 64. We consider
screening as a prevention intervention; this approach is
supported by two meta-analyses and a recent primary
study that have demonstrated that those aware of their
HIV seropositivity tend to engage in safer sexual
behavior with partners [24].

Health Care Manag Sci (2011) 14:115124

2.1 Epidemic model

2.1.1 Epidemic model: a dynamic compartmental model
We developed an epidemic model to determine the outcome
of different allocation decisions. The epidemic model is a
compartmental model defined by a nonlinear system of
differential equations. The formulation of the set of
ordinary differential equations draws on the structure of a
Susceptible-Infected model with an effective contact rate.
Such compartmental models are commonly used for
defining an epidemic model and have been used extensively elsewhere [59].
We consider the U.S. population aged 1364 years that
is at risk for HIV transmission, broken down into 15
subpopulations defined by risk group, gender and race.
Risk groups are men who have sex with men (MSM),
injection drug users (IDU) and high-risk heterosexuals
(HRH). Race is defined as non-Hispanic black (Black),
Hispanic and other, where the other category includes
primarily whites (98%) as well as Asians, Pacific Islanders,
Alaskan Natives and American Indians. We assume that all
HIV infections in the U.S. are the result of transmission
within these at-risk populations.
Each of the 15 subpopulations is modeled using a threecompartment epidemic model defined by those susceptible
to HIV infection (S), those who are HIV-positive but
undiagnosed (U) and those diagnosed with HIV (D). Figure
1 is a graphical representation of the compartmental model
for any given subpopulation.
We assume that entry is made only into the susceptible compartments. The entry rate, denoted by , is the
rate at which 13-year-olds age into the population. The
exit rates for the susceptible, undiagnosed and diagnosed
compartments are noted by, S, U, and D, respectively.
The exit rate is defined as the sum of the mortality rate and
the age out rate. New infections, noted I, represent the
incident cases that result from unsafe contact with
undiagnosed and diagnosed persons in the same or other
subpopulations. The diagnosis rate, denoted by X, is a
function of the resources allocated to HIV screening
interventions.

D
X

Fig. 1 Schematic of the compartmental model for a subpopulation

Health Care Manag Sci (2011) 14:115124

117

adjusted to control for the positive effect of DHAPs

existing allocation of funds. Details are provided in
Appendix A.

2.1.2 Epidemic model: differential equations, no

interventions
The epidemic model is presented in two stages. First, in
Eqs. 13, the model is presented assuming that no funds are
invested in a subpopulation from any source and no
interventions or diagnoses are taking place. We then present
the complete model in Eqs. 68.
The compartmental epidemic model illustrated in Fig. 1
can be written as a nonlinear system of differential
equations as follows:
0

Si t Ii t bSi  Si t  dSi  Si t

Ui t Ii t  d Ui  Ui t
Di t dDi  Di t

Let i be an index over the subpopulation (i=1 to 15).

Si(t), Ui(t) and Di(t) represent the numbers of individuals
who are susceptible, undiagnosed HIV positive, and
diagnosed HIV positive, respectively, in subpopulation i at
time t.
Equation 4 defines the calculation for the number of new
infections in subpopulation i in time period t and Eq. 5
defines the entire subpopulation.
Ii t

15 
P
lUij Uj tSi t
j1

Nj t

lDij Dj tSi t

Nj t

Ni t Si t Ui t Di t

8i

In Eq. 4, index j represents the subpopulations that may

transmit HIV to subpopulation i. An effective contact is
defined as a contact between an infected and a susceptible
person that results in infection of the susceptible person.
The effective contact rates are represented in Eq. 4 by
lambdas (1). Effective contact rates are indexed by
subpopulations i and j, and by the state of diagnosis U or
D because those aware of their seropositivity tend to
practice safer behaviors.
The contact rates 1Uij and 1Dij are calculated based on
HIV incidence and prevalence data and assumptions about
mixing rates between subpopulations. This method averts
the need to untangle the complex behaviors that contribute
to transmission among and between populations. The
effective contact rates combine the impact of all factors
that contribute to infection transmission, such as number of
partnerships, number of exposures to HIV within partnerships, transmission probabilities and infection co-factors
together into one rate. The calculated rates have been

2.1.3 Epidemic model: differential equations including

intervention effects
We assume, for the purpose of implementing interventions,
that individuals who are susceptible to infection are not
distinguishable from those who are HIV-infected but
undiagnosed and are therefore targeted together. Behavior
change interventions are aimed either at those in compartments S and U or at those who are diagnosed. Screening is
only aimed at those in compartments S and U. Incorporating possible investment in these interventions, the epidemic
model defined in Eqs. 13 above can be rewritten as in Eqs.
68. Since the model is the same in each subpopulation, i is
suppressed from the notation.
(t) is a production function that determines the
proportion of people who will be diagnosed positive based
on an investment of xk in screening intervention k for a
given subpopulation at time t; they are removed from the
undiagnosed compartment and transferred to the
corresponding diagnosed compartment. We assume that all
undiagnosed individuals who get screened at time t move to
the diagnosed compartment in the next period. A description of the production function is provided in subsection 2.3.1.
A(S+U)(t) and AD(t) are numbers of S and U, and D,
respectively, who have temporarily ceased engaging in
risky behavior following a behavior change intervention in
time period t. A(S+U)(t) and AD(t) are defined by production
functions that determine the impact of an investment of yi
for implementing behavior change interventions with the S
and U or D compartments of subpopulation i. The effect of
behavior change interventions is not permanent; therefore
those removed from the subpopulations will be reinserted
following the duration of effect (dur) of the intervention,
provided they have survived and not aged-out of the group
during that time. We consider four periods for the duration
of effect: 3 months, 6 months, 12 months and beyond 12
months. These periods were chosen based on follow-up
times typically reported in the behavior intervention
literature. For simplicity of the notation, the system of
equations presented in Eqs. 68 only presents one period of
the duration of effect.
In Eqs. 68, the penultimate term serves to decrease the
size of compartment S by the number of individuals who
successfully completed a behavioral intervention, and the
last term serves to augment compartment S by the number
of individuals who revert back to their original behavior
following the duration of effect of the intervention. The
system of equations in Eqs. 68 can be solved numerically
by simulating the model over a time horizon. The baseline

118

Health Care Manag Sci (2011) 14:115124

horizon is 5 year and time periods are set as monthly

increments.

S 0 t It bS  St  dS  St 

St
StU t

 
 ASU t S

S tdurSU
tdurSU U tdurSU



 ASU t  durSU  i  d S

6
0

U t ItdU  U tX t  U t

U t
StU t

 
 ASU t S

U tdurSU

tdurSU U tdurSU

 ASU t  durSU  1  dU
7

D0 t dD  Dt  X t  U t  AD t AD t  durD  1  dD

2.2 Intervention programs

2.2.1 HIV screening
We considered screening interventions targeted both by risk
group (one each aimed at MSM, HRH and IDUs) and by the
risk-, race- and gender-specific subpopulations. In addition, to
resemble the context of screening in general health care
settings, we considered a broad screening intervention aimed
at the general U.S. population aged 13 to 64 regardless of
susceptibility to HIV infection, and three broad screening
interventions aimed at general U.S. populations aged 13 to 64
by race (black, white and Hispanic). In total, there are 22 target
groups for screening interventions considered in the model;
these are summarized in Table 1.
The cost of targeting interventions is in part driven by
the cost of recruiting those who meet the inclusion criteria.
Typically, the cost of interventions increases along with the
level of targeting, so that more broadly targeted interventions are less costly per participant than more narrowly
targeted ones.

We also considered broad behavior change interventions aimed at the general U.S. population aged 13 to 64
and aimed at the black, Hispanic and white general U.S.
populations aged 13 to 64. These interventions require
spending on a broad population of both at-risk and not
at-risk individuals, but the benefits only accrue to those
at risk. More narrowly targeted interventions are assumed
to have both higher per person costs and expected
effects. In total, there are 49 behavior change type
interventions considered in the model; these are summarized in Table 2.
2.3 Production functions
In the context of epidemic control, production functions
translate the investment in a prevention program into a
favorable epidemiologic outcome. In our model, funding
affects the size of the susceptible, undiagnosed and diagnosed
risk populations thereby reducing the number of new HIV
infection transmissions. A similar technique, known as
harvesting, is used in predator-prey type models [10].

2.2.2 Behavior change interventions

Behavior change interventions reduce HIV incidence by
reducing unsafe behaviors. This effect is achieved by
temporarily removing people from the model and placing
them at no risk of HIV transmission or acquisition for
the duration of their exclusion. We consider HIV
behavior change interventions targeted to 6 groups by
risk and state of diagnosis (e.g., diagnosed IDUs), to 9
groups targeted by risk and race (e.g., Hispanic MSM),
and to 30 groups targeted by risk-, race-, gender- and
state of diagnosis (susceptible or undiagnosed and
diagnosed).

Table 1 Targets for HIV screening programs in the model

Target level (Highest to lowest)

Number of target
groups for screening
interventions

General population
General population by race
At risk population by risk group

1
3
3

At risk population by risk group,

race and gender (e.g., Hispanic male IDUs).
Total

15
22

Health Care Manag Sci (2011) 14:115124

119

Table 2 Targets for behavioural change interventions in the model

Target level (Highest to lowest)

Number of target
groups for behavior
change interventions

General population
General population by race
At risk population by risk group and
state of diagnosis
At risk population by risk group and
by race
At risk population by risk group,
race, gender and state of diagnosis
Total

1
3
6
9
30
49

Ak t h

2.3.1 HIV screening

Investment in screening interventions results in the transfer
of a portion of the undiagnosed compartment to the
diagnosed where the effective contact rate is much smaller.
Xk t

cScreenNegk 

dichotomous choice, for instance, whether or not they

engaged in any unprotected sex since the last follow-up.
The binary answer is used to evaluate the programs
effectiveness. Therefore in our model, those who practice
safe behaviors following the intervention will be temporarily removed from their compartment. Their removal lasts
through the estimated duration of the interventions benefits. For the behavior change interventions considered, we
estimate the mean effect size at 3 months, 6 months, and
12 months and beyond 12 months. The effects can
increase or decrease from one time period to another,
and can permit a proportion of participants to remain out
of the model for the entire five-year time horizon.

xk t
P
Si t cScreenPosk  Ui t

i2k

i2k

9
Equation 9 is the production function representing the
linear effect of investing in screening for period t. The
invested amount xk is a decision variable, denoting the best
possible expenditure, that is determined by the optimization
model where index k represents the HIV screening
program. cScreenNegk and cScreenPosk are the screening
costs per negative and positive diagnosis, respectively. Xk
represents the proportion of the undiagnosed who will be
diagnosed positive by investing xk in screening population
i; they are removed from the undiagnosed compartment and
transferred to the diagnosed compartment within i. The
target and level of HIV screening program k, determines
which subpopulations i belong to k. For example, if k is a
screening program aimed at the HRH risk group, then 6
subpopulations belong to k: Black male HRH, Black female
HRH, Hispanic male HRH, Hispanic female HRH, Others
male HRH and Others female HRH. When HIV screening
program k aims at the most targeted level (i.e. the at risk
population defined by risk group, race and gender) then
only one subpopulation i belongs to k.
2.3.2 Behavior change interventions
Behavior change interventions are generally evaluated at a
follow-up period of 3, 6 and/or 12 months. At every followup evaluation, participants are asked about their behaviors
over the most recent period; typically they are given a

yk t
cBehChk
effectBehChk

10

Equation 10 is the production function representing the

effect of investing in behavior change for period t. The
invested amount yk is a decision variable that is determined
by the optimization model. cBehChk is the per person cost
of a behavioral change intervention. effectBehChk is the
outcome of a behavioral change intervention, specifically,
the proportion of those who have ceased their risky
behavior as a result of the intervention. Ak represents the
total number of people who will be positively affected by
the intervention; they are removed from the subpopulations
i belonging to k proportionately to the size of the
compartment within i. The target and level of HIV
screening program k, determines which subpopulations i
belong to k.
2.4 Optimization model
Consistent with other work on HIV prevention, we assume
that the objective is to minimize the number of new
infections given a total budget constraint (e.g., [11]). The
allocation model is expressed below as Eqs. 1117. The
objective function in Eq. 11 aims to minimize the total
number of new infections in all i subpopulations over the
entire time horizon T by varying the decision variables xk(t)
and yk(t). xk(t) represents the amount allocated toward
screening in every time period and subpopulation, while
yk(t) represents the amount allocated toward behavior
change interventions in every time period and subpopulation. As per Eq. 4, Ii(t) is a function of the size of
compartments Si(t), Ui(t) and Di(t) who in turn are a
function of xk(t) and yk(t), see Eqs. 610.
In Eq. 12, B(t) is the total budget available in period t.
The invested amounts are the decision variables of the
optimization model, which are constrained by a maximum
budget, in Eq. 12, and by non-negativity, in Eq. 17. The

120

Health Care Manag Sci (2011) 14:115124

model offers the option for portfolio rebalancing on an

annual basis or assuming a constant annual allocation over T.
Two main types of constraints are formulated: allocation
constraints and penetration rate (or reach) constraints. The
allocation constraints, expressed in Eqs. 15 and 16 can be used
to set an upper and lower bound on the amounts allocated to
an intervention in a subpopulation by time period. The reach
constraints, expressed in Eqs. 13 and 14, are used to set an
upper and lower bound on the proportion of a subpopulation
that can be reached by an intervention in a given time period.
RSi(t) is the proportion of subpopulation i reached by a
screening program in time period t. It is calculated based on
the sum of funds allocated to screening for all k interventions
that serve subpopulation i, see Eq. 9, to which we add a
background testing rate intended to reflect the rate of testing
in subpopulation i that would occur in the absence of any
DHAP funded testing initiatives. Setting a background testing
level enables us to include the impact of outside funding
sources onto testing and diagnoses when evaluating RSi(t).
RBt(t) is the proportion of subpopulation i reached by a
behavior change program in time period t. For each
subpopulation i, it is calculated as the proportion of Ak(t),
see Eq. 10, to the size of the respective compartment of
subpopulation i.

Minimizexk ; yk

T X
X

Ii xk t; yk t

11

t1

subject to:
P

xk t yk t

 Bt 8t

RSLBi

 RSUBi

RBLBi

 RBi t 

12

RSi t

RBUBi

8i; t
8i; t

13

two notions are different, they bear a similar impact on the

proportion of a population that can reach or be reached by
an intervention.
We are not aware of a closed-form analytic solution for
the epidemic model specified by Eqs. 68. Thus, the
epidemic model and resulting number of new infections
between time zero and time T are estimated using discretetime approximations to the continuous system in monthly
time intervals [12, 13].
The model was formulated in Microsoft Excel. In order to
handle the nonlinearities associated with the epidemic model,
the optimization is performed using the Large Scale SQP
solver engine from Frontline Systems [14]. We validated the
results from the solver engine by restricting the problem size
and creating a utility that allows us to solve optimally using
a discrete search over the feasible region.
A univariate sensitivity analysis feature was designed to
gauge the stability of the model results. The sensitivity
analysis feature automates the evaluation of up to ten
variables to be modified, in sequence, according to five user
specified values. We established 12 benchmark criteria and
create upper and lower bounds based on a 10% increase or
decrease of the proportion of the budget allocated according
to the optimal model results. For example, one criterion
relates the proportion of the total budget allocated to
behavioral interventions, if the optimal model suggests this
proportion is 54%, then the upper and lower bounds are set
to 64% and 44%, respectively. The output of the sensitivity
analyses will highlight the scenarios where one or more of
the benchmark criteria have been violated and provides the
related details. When results of a scenario lie within the
bounds for all benchmark criteria, then the scenario is
sufficiently comparable to the optimal scenario to suggest
robustness, otherwise, the results are sensitive to the
variation and decreasing the uncertainty of the parameter
value is warranted.

14
3 Sample application

ASLBk

xk t

 ASUBk

8k; t

15

ABLBk

 yk t

 ABUBk

8k; t

16

xk t; yk t 

0 8k; t

17

Our definition of reach includes two main concepts. The

first is an individuals predisposition to seek an intervention, ability to access an intervention or be identified and
recruited for an intervention. The second is an intervention
providers structural and logistical capacity to recruit and
serve a population, within a fixed timeframe. Though these

In this section we present an example of the model applied

to DHAPs extramural program budget and offer illustrative
results. All data input parameters used for the example are
summarized in Table 3. For parameter values that do not
cite a reference, the parameter value is a reasonable
assumption and does not represent real data or actual
estimates.
DHAPs annual extramural budget for funding HIV
prevention programs is estimated at approximately $325
million. We estimate the current allocation of DHAPs
budget for each of the screening and behavioral intervention targets in our model and refer to it as the baseline
allocation. The optimal allocation strategy suggested by the

Health Care Manag Sci (2011) 14:115124

121

Table 3 Data assumptions for illustrative example

Parameter

Estimated value

Total annual budget (Bt)

Total number of persons at-risk of HIV transmission. Sum over all i subpopulations of Ni(0).
Total number of HIV infected and undiagnosed persons. Sum over all i subpopulations of Ui(0).
Total number of the HIV infected and diagnosed persons. Sum over all i subpopulations of Di(0).
Total number of new HIV infections in U.S, 2006. Sum over all i subpopulations of Ii(0).
Entry rate into the susceptible compartment, , average of all i subpopulations
Exit rate from the susceptible and undiagnosed compartments, S=U, average of all i subpopulations
Exit rate from the diagnosed compartments, D, average of all i subpopulations

$325,000,000
20,000,000
230,000 [22]
870,000 [22]
56,300 [1]
0.023
0.014
0.045

Cost of screening per negative, cScreenNegk, where k targets the general U.S. adult population (US$ 2006).
Cost of screening per positive, cScreenPosk, where k targets the general U.S. adult population (US$ 2006).
Cost of screening per negative, cScreenNegk, where k targets the at-risk populations (US$ 2006).
Cost of screening per positive, cScreenPosk, where k targets the at-risk populations (US$ 2006).
Per person cost of behavioral change interventions, cBehChk, where k targets HIV positives (US$ 2006).
Per person cost of behavioral change interventions, cBehChk, where k targets the at-risk populations (US$ 2006).
Effect size of behavioral change interventions at 3, 6 and 12 months of follow-up, effectBehChk, where k targets
HIV positives.
Effect size of behavioral change interventions at 3, 6 and 12 months of follow-up, effectBehChk, where k
targets the at-risk
populations.
Default minimum and maximum allocation to a behavioral change intervention (ABLBk; ABUBk).

$18 [23]
$82 [23]
$52 [23, 24]
$126 [23, 24]
$613
$340
40%

Default minimum and maximum allocation to a screening intervention (ASLBk; ASUBk).

Default minimum reach level for a behavioral change or a screening intervention (RBLBi, and RSLBi).
Median maximum reach level for behavioral change interventions (RBUBi)
Median maximum reach level for screening interventions (RSUBi)
Time horizon (T)

$0; $100,000,000
0%
33%
60%
5 years

model is compared to the baseline allocation and these

results are summarized in Table 4.
The current and model allocations of the annual budget
are comparable with respect to the split between screening
and behavioral interventions; however, within the allocation
to behavioral interventions the model suggests a considerably greater focus on diagnosed positives. The baseline
allocation allocates 29% of the annual budget to the general
population, and the remaining funds to the at-risk populations,
while results of the model indicate a marked preference for
allocation to the at-risk populations. The model allocation
suggests a greater focus on MSM and IDUs with a 72% of the
annual budget allocated to them, while the baseline allocation

43%

$0; $100,000,000

for MSM and IDUs totals 35%. In the context of our

illustrative example, these results imply that reallocating
funds currently aimed at the general population to target
MSM and IDUs for screening and behavioral interventions for
positives is likely to reduce HIV incidence.
Contrasting the existing allocation of the current budget
against the optimal allocation suggested by the model
provides guidelines for decision-makers as to how the
current budget could be reallocated to improve epidemic
outcomes without increasing the overall budget requirements. Table 4 is presented as an example only and is not
intended as an actual recommendation for improving the
allocation of HIV prevention resources.

Table 4 Illustrative results: allocation of annual budget

% of total
allocation by

Intervention

Risk group

Race

Screening Behavioral
Behavioral interventions General
HRH MSM IDUs Blacks Hispanics Other races
interventions for diagnosed positives population
Current Allocation 49%
Model Allocation 45%

51%
55%

6%
42%

29%
0%

36%
28%

24%
50%

11%
22%

47%
48%

24%
21%

30%
31%

122

4 Discussion
Current annual lifetime treatment costs for an HIV-infected
individual are estimated at $367,164 (US$ 2009), assuming
a life expectancy of 32 years and treatment for 24 years
[15]. Therefore, even a modest reduction in HIV incidence
translates to significant savings. Incorporating epidemic
projections in the decision-making process informs the
selection of populations and interventions that should be
targeted. Using the results of such models as guidelines to
better target funds to interventions and population subgroups is likely to reduce HIV incidence. These models can
also direct research by pointing to areas where the
development of cost-effective interventions can have the
most impact on the epidemic. The model supports what-if
analysis capabilities, which can be used to help decisionmakers understand the impact of deviations from the
optimal funding scenario.
Our model has several limitations. The structure of the
epidemic model does not include geographic or age stratifications. However, the data required to further stratify the
model by geography and age are difficult to obtain. Further,
limiting the size of the epidemic model is necessary to
maintain the performance of the nonlinear optimization
program, which is driven by the epidemic model.
American Indians, Alaska Natives, Asians and Pacific
Islanders are not considered explicitly and model results do
not address these populations. However, combined these
populations comprise 5% of the U.S. population aged 13
64 and 1.4% of AIDS case reports. We recommend that
resource allocation to these minority populations be
addressed through an analysis that focuses on these
minority groups separately.
Behavior change interventions are targeted to particular
subpopulations but they do not address the type of behavior
that is evaluated. For example, when an intervention is
targeted to female IDUs, there is no distinction between
reduction in needle sharing and increase in condom use. We
suggest that the results of our national-level model be used
to highlight which subpopulations benefit most from
behavior change interventions, and to guide future analyses
into the specific types of behavior change that have the
most effect on HIV acquisition and transmission.
We do not report results relating the number of new
infections projected under the baseline and the model
scenarios because the example provided is illustrative and
any improvements demonstrated by the model should not
be considered factual. However, we expect to publish a
manuscript focusing on actual model results in the near
term.
We expect to improve this model in several ways.
Currently, the positive effect of HIV diagnosis, that is the
movement from the undiagnosed to the diagnosed com-

Health Care Manag Sci (2011) 14:115124

partment, is maintained throughout the 5-year time horizon.

While cross-sectional data support the sustained behavioral
improvement resulting from a positive HIV diagnosis for
up to 10 years [2, 4, 16, 17]; we intend to explore allowing
the positive effect of HIV diagnosis to wane in time
horizons beyond 10 years. Also, HIV transmission rates are
a function of viral load and stage of disease. For example,
an individual at primary infection is far more infectious
than at the clinically asymptomatic stage of disease [1820].
We expect to expand this model, in future versions, to
account for these variations, which would also indicate the
extent to which early diagnosis benefits incidence.
In general, such models should not be expected to dictate
healthcare resource allocation, rather their results should be
used to support the decision-making process for allocating
resources, guide discussion and provide clarity to policy
debates. Leadership at CDCs Division of HIV/AIDS
Prevention state that the national HIV resource allocation
model provides information that will be used, in conjunction with program and other data, to continue to improve
the impact of HIV prevention efforts in the United States
and address the priorities identified in the National HIV/
AIDS Strategy[21], particularly, reducing the number of
people who become infected with HIV. The model provides
useful information that, consistent with the National HIV/
AIDS Strategy, can be used to direct resources to those
strategies and populations that are likely to have the
greatest impact on HIV incidence.
Financial disclosure The authors received no financial support for
this work.
Disclaimer The findings and conclusions in this paper are those of
the authors and do not necessarily represent the views of the Centers
for Disease Control and Prevention.

Appendix A
Identifying the effective contact rates
We identify incidence estimates in each subpopulation i at
the start of the horizon and denote it by Ii(0) [25]. We then
define IiD(0) and IiU(0), the incidence resulting from unsafe
contact with those diagnosed and the incidence resulting
from unsafe contact with those that are undiagnosed,
respectively. Over all subpopulations, 21% of those who
are HIV infected are currently undiagnosed [26] yet they
contribute 48% of the annual incidenceresulting in a
transmission rate that is more than three times higher than
that of those who have been diagnosed with HIV [16].
Investment in HIV screening leads to diagnoses thereby
moving people from the undiagnosed compartment into the
less risky diagnosed compartment.

Health Care Manag Sci (2011) 14:115124

For every combination of ij subpopulations, we estimate,

pij, the proportion of new infections in subpopulation i that
results from unsafe contact with subpopulation j. The
matrix of pij values is multiplied by the IiU(0) vector
resulting in IijU(0), the number of new infections in the
susceptible subpopulation i that result from unsafe contact
with the undiagnosed subpopulation j. Also, the matrix of
pij values is multiplied by the IiD(0) vector resulting in
IijD(0), the number of new infections in susceptible
subpopulation i that result from unsafe contact with
diagnosed subpopulation j.
Then, knowing the size of each subpopulation compartment, we identify l Uij and l Dij as follows:
I ijU 1

lijU  Uj 0  Si 0
IijU 1  Nj 0
) lijU
Nj 0
Uj 0  Si 0

IijD 1

lijD  Dj 0  Si 0
IijD 1  Nj 0
) lijD
Nj 0
Dj 0  Si 0

We assume that Ii Eq. 1 can be approximated by Ii(0). In

the calculation of new infections, the denominator which
expresses the size of the partner subpopulation, Ij(t),
includes all those who have ceased engaging in risky
behavior following a behavior change intervention.
These contact rates combine the impact of all factors that
contribute to infection transmission such as number of
partnerships, number of acts, transmission probabilities and
infection co-factors together into one rate. Defining contact
rates from incidence averts the needs to untangle the factors
that contribute to an effective transmission.
l Uij and l Dij are fixed throughout the time horizon;
benefits from investing in prevention programs result from
reducing the size of the susceptible and infected compartments.
Our estimates of HIV incidence are based on HIV
surveillance data that reflect epidemic trends occurring
under current conditions, including the current expenditure
of prevention funds. Given that our contacts rates are
derived from incidence estimates, an adjustment is required
to remove the effects of the existing allocation from the
contact rates. The goal is to find contact rates that yield the
latest HIV incidence estimates under the existing allocation.
To do this, we define a linear optimization model where all
contact rates as set as decision variables, we then set the
amounts allocated to all interventions (i.e. decision variables referred to in Eq. 11) as per the existing allocation. The
objective is to minimize the squared sum of the difference
between the HIV incidence estimates in each subpopulation
and the number of new infections in each subpopulation
given the existing allocation of funds. The optimal solution
sets the number of new infections in each subpopulation
under the existing allocation of funds equal to the HIV

123

incidence estimates and defines the final contact rates for

the resource allocation model.

References
1. Hall HI et al (2008) Estimation of HIV Incidence in the United
States. J Am Med Assoc 300(5):520529
2. Marks G et al (2005) Meta-analysis of high-risk sexual behavior
in persons aware and unaware they are infected with HIV in the
United States: Implications for HIV prevention programs. J
Acquir Immune Defic Syndr 39:446453
3. Weinhardt LS et al (1999) Effects of HIV counseling and testing
on sexual risk behavior: a meta-analytic review of published
research, 19851997. Am J Public Health 89(9):13971405
4. Marks G et al (2009) Understanding differences in HIV sexual
transmission among Latino and black men who have sex with
men: The Brothers y Hermanos study. AIDS Behav 13(4):682
690
5. Bailey NTJ (1975) The Mathematical Theory of Infectious
Diseases and its Applications, 2nd edn. Charles Griffin and
Company, United Kingdom
6. Capasso V (1993) Mathematical Structures of Epidemic Systems.
Springer, Berlin
7. Edwards D, Shachter R, Owens DK (1998) A dynamic HIVtransmission model for evaluating the costs and benefits of
vaccine programs. Interfaces 28(3):144166
8. Tchetgen E, Kaplan EH, Friedland GH (2001) Public Health
Consequences of Screening Patients for Adherence to Highly
Active Antiretroviral Therapy. J Acquir Immune Defic Syndr
26:118129
9. Zaric GS, Brandeau ML, Barnett PG (2000) Methadone maintenance treatment and HIV prevention: A cost effectiveness
analysis. Manage Sci 25:10131031
10. Supriatna AK, Possingham HP (1998) Optimal Harvesting for a
Predator-Prey Metapopulation. Bull Math Biol 60(1):4965
11. Kaplan EH, Pollack HA (1998) Allocating HIV prevention
resources. Socio-Economic Planning Sciences 32:257263
12. Luenberger, D.G., Introduction to dynamic systems : theory,
models, and applications. 1979, New York: Wiley. 446.
13. Zaric GS et al (1998) The effect of protease inhibitors on the
spread of HIV and the development of drug resistance: a
simulation study. Simulation 71:262275
14. Frontline Systems (2008) Large-Scale SQP Solver Engine Version
8.0 for Microsoft Excel. Incline Village, NV
15. Schackman BR et al (2006) The lifetime cost of current human
immunodeficiency virus care in the United States. Med Care 44
(11):990997
16. Marks G, Crepaz N, Janssen RS (2006) Estimating sexual
transmission of HIV from persons aware and unaware that
they are infected with the virus in the USA. AIDS 20:1447
1450
17. Weinhardt L (2005) HIV diagnosis and risk behavior. In: Kalichman
S (ed) Positive prevention: reducing HIV transmission among people
living with HIV/AIDS. New York, Kluwer Academic/Plenum, pp
2963
18. Quinn TC (1997) Acute primary HIV infection. J Am Med Assoc
278(1):5862
19. Rapatski B, Suppe F, Yorke J (2005) HIV epidemics driven by late
disease stage transmission. J Acquir Immune Defic Syndr 38
(3):241253
20. Pilcher C et al (2004) Brief but efficient: acute HIV infection
and the sexual transmission of HIV. J Infect Dis 189(10):1785
1792

124
21. The White House Office of National AIDS Policy (July 2010)
National HIV/AIDS strategy for the United States. Washington,
DC, pp 60
22. Campsmith ML et al (2010) Undiagnosed HIV prevalence
among adults and adolescents in the United States at the
end of 2006. J Acquir Immune Defic Syndr 53(5):619
624
23. Farnham PG et al (2008) Comparing the costs of HIV screening
strategies and technologies in health-care settings. Public Health
Rep 123(Suppl 3):5162

Health Care Manag Sci (2011) 14:115124

24. Shrestha RK et al (2008) Cost-effectiveness of finding new HIV
diagnoses using rapid HIV testing in community-based organizations. Public Health Rep 123(Suppl 3):94100
25. Centers for Disease Control and Prevention (2008) Subpopulation
estimates from the HIV incidence surveillance system - United
States, 2006. Morbidity and mortality. Weekly Rep 57(36):985989
26. Campsmith M, Rhodes P, Hall I (2009) Estimated prevalence of
undiagnosed HIV infection: US, end of 2006, in 16th Conference
on Retroviruses and Opportunistic Infections (CROI 2009).
Montreal, QC. p. Abstract No.1036

Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.