ESSAYS IN APPLIED MICROECONOMICS

BY
NICHOLAS L. WILSON
B.A., REED COLLEGE, 1999
MPA/ID, HARVARD UNIVERSITY, 2001
A.M., BROWN UNIVERSITY, 2004

A DISSERTATION SUBMITTED IN PARTIAL FULFILLMENT OF THE
REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY
IN THE DEPARTMENT OF ECONOMICS AT BROWN UNIVERSITY

PROVIDENCE, RHODE ISLAND
MAY 2008

c Copyright 2008 by Nicholas L. Wilson
°

iii

VITA

Nicholas L. Wilson was born on August 8, 1977 in Kansas City, Missouri. He received his B.A.
in Economics from Reed College in 1999, his MPA in International Development from Harvard
University in 2001, and his A.M. in Economics from Brown University in 2004. He has taught
economics at Brown University, Park College, and Harvard University and has conducted field
research in Bangladesh and Zambia.

iv

ACKNOWLEDGMENTS
I would like to thank Andrew Foster, Kaivan Munshi, and David Weil for their advice and support. In addition, my research benefited from numerous discussions with Daniel Bennett, Jason
Goldman, Emily Hazan, and Alaka Holla. The chapters on HIV testing would not be possible without the cooperation of Bates Buckner at MEASURE Evaluation and many individuals in Zambia,
including Bristol Cheembo and Clement Mulenga at the Virology Lab of the University Teaching
Hospital, Dr. Albert Mwango at the Ministry of Health, and Nchimunya Nkombo at the Central
Statistical Office.
I would like to thank Paul Koussa for excellent computing support, Angelica Spertini for tremendous administrative and emotional support, and Svetla Vitanova, without whom I would not have
completed my Ph.D. Of course I would be remiss if I did not also thank my parents, Catherine and
Jeff, my sister, Rachel, and my brother, Jacob, all of whom have been extremely supportive during
my lengthy academic studies.
I dedicate this dissertation to two of my grandparents who passed away while I was writing it:
Marion R. Simons and Logan L. Wilson.
All errors that may remain in this dissertation are solely my responsibility.

v

Contents
Vita

iv

Acknowledgements

v

List of Figures

viii

List of Tables

ix

Introduction

1

1 A Model of Demand for HIV Testing: Descriptive and Theoretical Analysis
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3.1 Basic setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3.2 Demand for risky sexual behavior . . . . . . . . . . . . . . . . . . . . . . . .
1.3.3 Demand for HIV testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3.4 Change in risky sexual behavior . . . . . . . . . . . . . . . . . . . . . . . .
1.3.5 Effect of ART rationing . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.5 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.

3
3
6
10
10
11
12
15
16
17
19

2 A Model of Demand for HIV Testing: Structural Estimation and Results
2.1 Data and estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.2 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.1 Parameter estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.2 Predicted testing-age profiles . . . . . . . . . . . . . . . . . . . . . . . .
2.2.3 Predicted risky behavior demand function . . . . . . . . . . . . . . . . .
2.3 Policy scenarios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.1 No ART . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.2 Status quo ART policy . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.3 Expanding the supply of ART . . . . . . . . . . . . . . . . . . . . . . .
2.3.4 Eliminating the source of inefficiency . . . . . . . . . . . . . . . . . . . .
2.3.5 Eliminating the inefficiency while expanding the supply of ART . . . . .
2.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.5 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.6 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

25
25
25
26
28
28
29
30
30
31
33
33
35
36
37
39
40

vi

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

3 Identity, Parochial Institutions, and Career Decisions
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 The Institutional Setting . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.1 The Settling of the Midwest . . . . . . . . . . . . . . . . . . . . . .
3.2.2 Ethnic Labor Networks in the Midwest . . . . . . . . . . . . . . .
3.2.3 Ethnic Fractionalization Across the Midwest . . . . . . . . . . . .
3.2.4 Historical Fractionalization and Current Economic Conditions . . .
3.3 The Church and the Persistence of Identity . . . . . . . . . . . . . . . . .
3.4 Local Identity and Career Decisions . . . . . . . . . . . . . . . . . . . . .
3.5 Empirical Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5.1 Individual Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5.2 Fractionalization in the County of Birth and Occupational Choice
3.5.3 Fractionalization and Outcomes Related to Occupational Choice .
3.5.4 Alternative Explanations . . . . . . . . . . . . . . . . . . . . . . .
3.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.7 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.8 Appendix: Labor Market Equilibrium . . . . . . . . . . . . . . . . . . . .
3.8.1 Population and Production Technology . . . . . . . . . . . . . . .
3.8.2 The Central Planner’s Problem . . . . . . . . . . . . . . . . . . . .
3.8.3 Equilibrium Labor Allocation . . . . . . . . . . . . . . . . . . . . .

vii

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

46
46
52
52
54
56
58
61
68
70
70
72
75
76
78
79
83
83
84
86

List of Figures
1.1
1.2
1.3
1.4
1.5
1.6

HIV Prevalence, 2001 Zambia DHS . . . . . . . .
HIV Testing and Proximity to ART . . . . . . .
HIV Testing and Age, Females ZSBS . . . . . . .
HIV Testing and Age, Males ZSBS . . . . . . . .
Proximity to VCT and Age, 2005 ZSBS . . . . .
Demand for HIV Testing and for Risky Behavior

.
.
.
.
.
.

19
20
21
22
23
24

2.1
2.2

Predicted and Observed Testing-Age Profiles . . . . . . . . . . . . . . . . . . . . . .
Estimated Risky Behavior Demand Function . . . . . . . . . . . . . . . . . . . . . .

40
41

3.1
3.2
3.3
3.4
3.5
3.6

Early Growth in the Midwest . . . . . . . . . . . . . . . . . .
Ethnic Fractionalization in 1860 . . . . . . . . . . . . . . . . .
Religious Participation . . . . . . . . . . . . . . . . . . . . . .
Religious Participation by Denomination . . . . . . . . . . . .
The Central Planner’s Problem . . . . . . . . . . . . . . . . .
Historical Fractionalization and Current Occupational Choice

87
88
89
90
91
92

viii

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

List of Tables
2.1
2.2
2.3
2.4

Parameter Estimates . . . . . . . . . . . . . . . . . . . . .
Testing Rates Under Alternative Policy Scenarios . . . . .
Mean Risky Behavior Under Alternative Policy Scenarios
New Infections Under Alternative Policy Scenarios . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

42
43
44
45

3.1
3.2
3.3
3.4
3.5
3.6
3.7
3.8
3.9
3.10
3.11
3.12

Ethnic Distribution, 1860-1900 . . . . . . . . . . . . . . . . . . . . . .
Occupational Distribution, 1860-1900 . . . . . . . . . . . . . . . . . . .
Transportation Infrastructure and County Characteristics, 1860 . . . .
Fractionalization in 1860 and County Characteristics in 1990 . . . . .
Fractionalization in 1860 and Local Government Expenditure in 1990 .
Distribution of Denominations and Religious Participation, 1860-2000
Fractionalization in 1860 and Religious Participation in NLSY79 . . .
NLSY79 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . .
Fractionalization in the County of Birth and Occupational Choice . .
Fractionalization and Outcomes Related to Occupational Choice . . .
Fractionalization in the County of Residence and Occupational Choice
Fractionalization and Educational Attainment . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

93
94
95
96
97
98
99
100
101
102
103
104

ix

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

Introduction
This thesis examines several topics in applied microeconomics. The first chapter presents a model
of demand for HIV testing and derives the conditions under which HIV testing will reduce risky
behavior. This model captures the two main reasons why an individual would want to take a
HIV test: (i) to begin antiretroviral therapy (ART) if the individual is HIV positive, and (ii) to
increase the precision of belief about the marginal cost of risky behavior. This model shows that
in the absence of ART, the individuals that are most willing to take a HIV test are individuals
that have the least precise beliefs about their own HIV status (i.e., individuals in the middle of
the distribution of past risky behavior) and are expected to demonstrate the greatest reductions
in post-testing risky behavior. Introducing ART into the model means that individuals with the
highest subjective probability of being HIV positive should demonstrate the greatest increases in
testing demand between the pre-ART period and the ART era.
The second chapter examines empirical evidence on the predictions of the model using newly
assembled data from Zambia, one of the highest HIV prevalence countries in the world. I show that
testing behavior in Zambia prior to the introduction of ART is consistent with the prediction of the
model: the individuals most likely to test are the individuals expected to demonstrate the greatest
reductions in post-testing risky behavior. The observed testing response to the introduction of free
ART, however, is not consistent with the second main prediction of the model. Specifically, a large
group of individuals whose testing decisions appear to be particularly responsive to the introduction
of ART are in fact expected to demonstrate only small or modest reductions in post-testing risky
behavior. This finding is both a puzzle and - from the perspective of maximizing the prevention
impact of ART-induced testing - evidence of an inefficiency. I provide an explanation for this
puzzle and provide evidence rejecting alternative hypotheses. In addition, I estimate the structural
parameters of the model and use these estimates to simulate several counterfactual ART policies.
The simulation results indicate that the magnitude of the inefficiency is quite large: eliminating the
inefficiency while holding constant the quantity of ART supplied would increase the number of new
infections avoided due to ART-induced testing by more than ten-fold.
1

2
The third chapter examines ethnic (migrant) labor market networks in the American Midwest
when it was first being settled in the middle of the 19th century, the endogenous emergence of a
sense of local identity that helped support the functioning of these networks, and how this sense
of local identity may have persisted across multiple generations and affect economic outcomes in
the late 20th century. While fractionalization may have adversely may adversely affect the performance of secular institutions, ethnic competition in the labor market would at the same time
have strengthened within-group loyalty and parochial institutions. These values and their complementary institutions, notably the church, could have mutually reinforced each other over many
overlapping generations, long after the networks themselves had ceased to be salient. Counties with
greater ethnic fractionalization in 1860 are indeed associated with steadily increasing participation
in select religious denominations historically dominated by migrants all the way through the twentieth century. Complementing this result, individuals born in high fractionalization counties are
significantly less likely to select into geographically mobile professional occupations and, hence, to
migrate out of their county of birth, despite the fact that these counties are indistinguishable from
low fractionalization counties in terms of local public good provision and economic activity today.

Chapter 1

A Model of Demand for HIV
Testing: Descriptive and
Theoretical Analysis
1.1

Introduction

Policymakers believe that HIV testing is an important intervention in the HIV/AIDS pandemic.
Proponents of HIV testing presume that on average individuals that learn whether or not they are
infected will respond to this new information by reducing risky behavior. However, HIV testing
rates are low in much of the world and a vast majority of HIV positive individuals are not aware
that they are infected. For example, among the 85 percent of the world population infected with
HIV/AIDS that reside in Sub-Saharan Africa only 10 percent know their HIV status (WHO 2006a).
One of the main mechanisms by which policymakers hope to increase demand for HIV testing is
increased availability of subsidized antiretroviral therapy (Global HIV Prevention Working Group
2004). The availability of antiretroviral therapy (ART) gives individuals a direct incentive to take
a HIV test: if they take a HIV test and are HIV positive then they may begin therapy, reducing
morbidity and prolonging life. However, the incentive effects of ART may be heterogeneous and
many individuals may not choose to test in response to the introduction of ART. Moreover, those
that test because of ART may not demonstrate substantial behavior change subsequent to testing.
Chapters 1 and 2 examine the effect of ART availability on demand for HIV testing and the
effect of ART-induced testing on demand for risky behavior. In Chapter 1, I present a model of
demand for HIV testing and for risky behavior. Initially I assume that the conditions under which
individuals have an incentive to test and testing reduces risky behavior hold. In particular, I assume
that demand for risky behavior is concave in the prior probability of being HIV positive. Under these

3

4
conditions, I show that prior to the introduction of ART it is the riskiest individuals (i.e., individuals
with the highest prior probabilities of being HIV positive) that have the greatest incentive to test
in the setting examined in the current analysis. The riskiest individuals are also those individuals
whose testing decisions should be most responsive to the introduction of ART. I show that under
these same conditions the expected reduction in risky behavior associated with taking a HIV test
in this setting should be increasing in the prior probability an individual is HIV positive. Thus,
ART-induced testing has the potential to substantially reduce the spread of HIV/AIDS.
Using newly assembled data from before and after the introduction of ART in Zambia I examine
the empirical evidence on these predictions. Testing behavior among women and men prior to
the introduction of ART is consistent with the prediction that individuals with the highest prior
probabilities of being HIV positive are those that should be most likely to test. The change in
testing behavior among women after the introduction of ART is consistent with the prediction
that individuals with the highest prior probabilities of being HIV positive are those that should
demonstrate the greatest response to the availability of ART. However, the change in testing behavior
among men after the introduction of ART is not consistent with this prediction. Specifically, testing
decisions among older men appear to be particularly responsive to the introduction of ART despite
the fact that older men are one of the lowest HIV prevalence groups in Zambia. Moreover, testing
decisions among men in the middle of the age distribution are not particularly responsive to the
introduction of ART despite the fact that these men are one of the higher HIV prevalence groups. I
interpret this as evidence of a non-random rationing mechanism determining the allocation of ART
among HIV positive individuals (in favor of older males) and provide evidence rejecting alternative
hypotheses.
In Chapter 2, I estimate the structural parameters of the model and show that the conditions for
individuals to have an incentive to test and for testing to reduce risky behavior are indeed satisfied in
this setting. In addition, as part of the estimation process I recover the ART rationing rule. Using
the parameter estimates and the information on the rationing rule I simulate the effect of ART
availability on testing demand and the effects of ART-induced testing on demand for risky behavior
and on the spread of HIV/AIDS. I also examine the effects of eliminating the non-random rationing
mechanism and/or expanding the supply of ART. Simulation results show that under the existing
policy ART availability increased testing demand by approximately 2 percentage points and ARTinduced testing reduced the incidence of HIV by approximately 1 percent. Expanding the supply
of antiretroviral drugs without eliminating the non-random rationing mechanism would only have
small effects on testing rates and risky behavior. In contrast, eliminating the non-random rationing

5
mechanism while holding fixed the existing supply of antiretroviral drugs would increase testing
rates by approximately 25 percentage points and reduce the incidence of HIV by approximately 15
percent.
These chapters contributes to the small body of existing economic literature on the behavioral
effects of HIV testing and of ART availability.1 Philipson and Posner (1995) examine the effects of
public subsidies for HIV testing on risky behavior and the incidence of HIV. The authors present
a trade model with asymmetric information where HIV testing is a method of quality verification.
An individual that is possibly HIV positive may show a negative HIV test to their potential partner
to verify that they are in fact HIV negative and induce the partner to engage in unprotected sex.
Although the model presented in the current analysis shares this focus on uncertainty about HIV
status, I focus on uncertainty about one’s own status and not the status of one’s partner.
Boozer and Philipson (2000) examine the conditions under which HIV testing will reduce risky
behavior and estimate the effect of learning the test result on subsequent behavior. Unlike the other
literature in this area, the authors emphasize the role of prior beliefs about one’s own HIV status
in conditioning the effect of HIV testing on risky behavior. In particular, they argue that HIV
testing should only affect behavior among the subset of individuals for whom the result of the test
differs from the prior belief. The model presented in the current analysis is similar to the analytic
framework discussed in Boozer and Philipson (2000). An important difference lies in the empirical
analysis. Boozer and Philipson (2000) use a sample of individuals who all take a HIV test to estimate
the reduced form relationship between (i) the interaction of prior beliefs and the result of the test
and (ii) subsequent behavior. In contrast, I estimate the structural parameters of the risky behavior
demand function, allowing for simulations of the effects of alternative ART policies.
Thornton (2006) examines the effect of randomly assigned financial incentives (and randomly
placed counseling sites) on the propensity to return to receive the results of a HIV test among a
sample of individuals who were asked to provide blood samples for a HIV test. Using the financial
incentive and distance to the counseling site as exogenous sources of variation in the effective price of
learning one’s results, the author finds that demand is very sensitive to price. Moreover, the author
finds that an exogenous change in whether an individual learns the result of a HIV test has only
1 There is more literature on the behavioral effects of HIV testing than on the behavioral effects of ART availability
and I focus on the former in the discussion of the existing literature. Lakdawalla et al (2006) is one of the few
economics articles on the behavioral effects of ART availability. The authors argue that technological advances in the
efficacy of antiretroviral therapy have increased the incidence of HIV in the United States. By lowering the health cost
of acquiring HIV, improvements in ART have reduced the marginal cost of risky sexual behavior and hence increased
the equilibrium quantity demanded (yielding an associated increase in rates of HIV transmission). Although this
effect may be operative in the setting I examine in the current analysis, I focus on the effect of ART-induced testing
on demand for risky behavior.

6
a small effect on subsequent choices about risky sexual behavior. A fundamental insight shared by
the current analysis and Boozer and Philipson (2000) may explain the seemingly small effect of HIV
testing on risky sexual behavior in Thornton (2006): the effect of HIV testing on subsequent choices
about risky behavior depends on the individual’s prior belief about their HIV status. If individuals
are approximately evenly distributed between those that (i) believe they are HIV positive but are
in fact HIV negative and (ii) believe that they are HIV negative but are in fact HIV positive, then
HIV testing may have little effect on subsequent behavior.

1.2

Setting

Zambia has one of the highest rates of HIV prevalence in the world. According to the anonymous
HIV testing component of the 2001 Zambia Demographic and Health Survey (ZDHS), the HIV
prevalence rate among adults age 15-49 is 14.5 percent. Moreover, high HIV prevalence in Zambia
is not a recent development. Data from antenatal clinics in the sentinel surveillance program are
consistent with the hypothesis that HIV prevalence has remained relatively stable from the early
1990s onwards (UNAIDS/WHO 2006).
As in many Sub-Saharan African countries (i.e., high prevalence countries), the highest rates
of HIV prevalence are among individuals in the middle of the age distribution. Figure 1 presents
the results of a locally weighted smoothed regression of an indicator variable for HIV status on
age, estimated separately for females and for males in the 2001 ZDHS. As shown in Figure 1, HIV
prevalence is highest among individuals in the middle of the age distribution, prevalence is slightly
higher among women than among men, and the age of peak prevalence among females is less than
that among males (34 and 40 years, respectively). The average age difference in sexual partners may
explain this last fact. Data from the Zambia Sexual Behavior Survey (ZSBS) on the age of sexual
partners show that on average males are approximately six years older than their female sexual
partners.
Magruder (2006) provides an explanation for the inverted U-shape that characterizes the prevalenceage profiles of most high prevalence countries. Individuals in early adulthood engage in “marital
shopping” - the search for a potential marital partner - and experience higher than average turnover
in sexual partners. Moreover, the effect of this behavior on prevalence is magnified by the biology
of the HIV virus: infectiousness during the first month of having acquired HIV is much higher than
during the majority of the life course of the virus (Wawer et al 2005) and rapid turnover of sexual
partners during the period of having recently acquired HIV can lead to a sharp increase in incidence

7
and prevalence.
In an effort to address the high HIV prevalence rate, the government of Zambia began offering
free voluntary counseling and testing (VCT) in 1998. Although initially rates of HIV testing were
low, they are increasing. According to the ZSBS, the proportion of individuals voluntarily taking a
HIV test in the 12 months prior to the survey round increased from between 5.5 and 6 percent in
2000 and 2003 to over 10 percent for women and nearly 8 percent for men in 2005.
One possible explanation for the increase in testing rates is that free antiretroviral therapy (ART)
became available in Zambia around this time. The early 21st century witnessed a dramatic expansion
in the provision of antiretroviral therapy in Sub-Saharan Africa. Between 2000 and the end of 2005,
the number of individuals on ART in the region increased from fewer than 50,000 (Attawell and
Mundy 2003) to 810,000 (WHO 2006b). Zambia is a participant in this expansion and in May
2004 the government of Zambia began to provide ART free of charge at primary health care centers
(Stringer et al 2006).
The introduction of free antiretroviral therapy in Zambia may have increased testing demand by
providing those individuals that believed they may have been HIV positive with an added incentive
to test. However, other factors affecting testing rates could have changed during the period as well.
For example, the number of VCT sites increased between 2000-03 and 2005.
If the introduction of ART had a causal effect on testing demand, then we should expect to see an
inverse relationship between distance to the nearest ART site and the change in testing rates between
2000-03 and 2005. The Japanese International Cooperation Agency 2004 Health Facilities Census
(HFC) provides global positioning system (GPS) information for each of the ART sites open at the
time of the 2005 survey round of the ZSBS. In addition, the ZSBS includes the same information for
the precise location of the rural households in 2000-2005 survey rounds. For the urban households
in the ZSBS, I compute the GPS coordinates of the centroid of the Standard Enumeration Area
(SEA) of residence; urban SEAs are relatively small, so each individual resides relatively close to the
centroid. Using the GPS data from the HFC and the ZSBS I calculate the distance between each
household and each ART site, and for each household I construct a measure of the distance to the
nearest ART site.
Figure 2 presents the results of a locally weighted smoothed regression of an indicator variable
equal to one if the individual voluntarily took a HIV test in the 12 months prior to the survey
month on distance to the location of the nearest ART site, estimated separately for the pre-ART
and ART periods. In the pre-ART period, individuals closer to sites that eventually provided ART
were more likely to test. Nonetheless, the increase in testing rates associated with the ART period is

8
greatest for those individuals closest to ART sites. Of course an important omitted variable in this
analysis is the distance to the nearest VCT site; proximity to a VCT site affects the testing decision
and is positively correlated with proximity to an ART site. However, health clinics in which ART
was introduced in 2004 and 2005 tended to be sites that had previously provided VCT services and
the pre-ART regression line reflects this fact. (The expansion in VCT between the pre-ART and
ART periods occurred largely in areas that were not close to ART sites and hence may explain the
increase in testing rates among individuals further from ART sites in 2005.)
Although it appears that the introduction of subsidized ART caused an increase in aggregate
testing demand, there exists substantial heterogeneity in the change in testing demand associated
with this intervention. Figure 3 shows the testing-age profile for women prior to the availability of
ART and after the introduction of ART. Prior to the availability of ART, women in the middle of
the age distribution were most likely to take a HIV test. These women were also those with the
highest HIV prevalence. After ART was introduced in Zambia testing increased among women of
all ages, but the increase was largest for the women in the middle of the age distribution. These
differences by age are consistent with the idea that ART availability increased testing demand and
the incentive effect of ART on testing demand is greatest for those individuals with the highest
probability of being HIV positive.
For males, the pattern is noticeably different. Figure 4 shows the testing-age profile for men prior
to the availability of ART and after the introduction of ART. As with females, in the “pre” period
it is men in the middle of the age distribution that were most likely to take a HIV test. Likewise,
these men were also those with the highest HIV prevalence. In the “post” period, however, it is the
oldest males that are most likely to take a HIV test. It appears that the idea that individuals with
the highest probability of being HIV positive should have testing decisions that are most responsive
to ART availability cannot explain the observed patterns of testing demand among males.
Although much of the change between 2000-03 and 2005 in the testing-age profiles for females and
for males may be attributed to the introduction of ART, it may be that the explanation for the large
increase in testing rates among older males (and perhaps that among women in the middle of the
age distribution) lies elsewhere. One change that paralleled the introduction of ART in Zambia was
a continued expansion in the number of VCT sites. Nonetheless, it is unlikely that the expansion
in the number of testing sites caused the observed changes in the testing-age profiles for women
and for men. For the expansion in VCT sites to explain these changes, the expansion should have
decreased the distance to the nearest VCT site differentially by age. Moreover, the age differential
in the expansion intensity should have differed by gender.

9
The ideal test of this hypothesis would be to regress the change in the distance to the nearest
VCT site on measures of each individual’s age, age squared, and gender as well as interactions
between age (and age squared) and gender. Unfortunately, data on which of the VCT sites open in
2005 existed prior to the introduction of ART in Zambia do not exist. In lieu of such data, we may
examine the relationship between distance to the nearest VCT site and age in 2005. If the change
in the distance to the nearest VCT site is correlated with age and gender, then we should expect
to see evidence of spatial clustering by age/gender and possibly evidence of substantial differences
by age in distance to the nearest VCT site in 2005. The HFC includes GPS information for VCT
sites and for each respondent in the 2005 ZSBS I construct a measure of the distance to the nearest
VCT site. Figure 5 presents the results of a locally weighted smoothed regression of distance to the
nearest VCT site on age, estimated separately for females and males. There are small differences by
age and gender in the average distance to the nearest VCT site, but these differences do not readily
explain the changes in the testing-age profiles (e.g., older males and males in the middle of the age
distribution appear to reside at similar distances from VCT sites). Moreover, the overall flatness
of the distance-age profiles suggests that individuals are not spatially clustered by age. Thus, it
appears that the expansion in the number of VCT sites between 2003 and 2005 does not explain the
large increase in testing rates among older males.
The likely explanation for the large testing response among older males associated with the
introduction of ART is that there exists a non-random rationing mechanism for allocating ART.
Currently, the supply of antiretroviral drugs in Zambia is insufficient to ensure that all individuals
that are HIV positive and require ART are guaranteed coverage. If supply-side constraints mean
that priority for access to ART depends on an individual’s age and gender or factors correlated
with these characteristics, then we would expect to see the observed changes in testing-age profiles
deviate from those predicted by a simple model of testing demand where the incentive effect of ART
is increasing in the prior probability that the individual is HIV positive.
Social norms in Zambia are consistent with the notion that older males receive priority for access
to ART. Although there may exist variation across the more than seventy ethnolinguistic groups in
Zambia, in much of traditional Zambian society younger men are subordinate to older men (Bond
1982). The oldest resident male is general identified as the household head (Scudder 1962). Among
the Tonga ethnic group, one of the largest in Zambia, an elder is called mupati or “big person”
(Colson and Scudder 1981). Among the Bemba ethnic group, another of the largest ethnic groups
in Zambia, males are not considered to be fully adult members of the community until they are
married and have children. Older males, or bakalamba, control decision-making not just within their

10
kin group, but also more broadly within the community. They are figures of authority and largely
control village decision-making (Epstein 1981).
Although scant, direct evidence on the process determining the allocation of ART among HIV
positive individuals is also consistent with the idea that older males receive priority in this process.
Two of the main groups that are documented as receiving priority for access to ART are richer
individuals and individuals with political connections (e.g., civil servants) (Jones 2004, Bwalya 2006).
Despite the lack of direct evidence on age-based preferences, it is likely that older men represent a
disporportionate share of these two groups.
Thus, although it appears that ART availability in Zambia increased testing demand, the change
in demand was conditioned by a non-random ART allocation mechanism. To frame the analysis of
the effects of this mechanism on demand for HIV testing and on demand for risky sexual behavior,
the next section presents a simple model of demand for HIV testing and risky behavior. I then
estimate the parameters of the model and use these to quantify the magnitude of these effects.

1.3

Model

The changes in testing behavior associated with the introduction of ART indicate that there exists
a non-random rationing mechanism determining the allocation of ART among HIV positive individuals. In particular it appears that priority for access to ART is given to older males at the expense
of males in the middle of the age distribution. To examine this finding more rigorously, I present a
model of demand for HIV testing and for risky behavior. I match the model to the data and find
that the model fits pre-ART testing behavior quite well but fails to completely match the change
in testing behavior after the introduction of ART. Thus, I expand the model to capture the effect
of ART rationing and recover the rationing rule from the observed testing behavior. Then I use
the parameter estimates and the information on the rationing rule to simulate the effect of ART
availability on testing demand and demand for risky behavior under several ART policies.

1.3.1

Basic setup

In this model an individual chooses whether to take a HIV test and chooses how much risky behavior
in which to engage. There are two components to expected utility: the pleasure the individual derives
from engaging in risky behavior (e.g., unprotected sex) and the displeasure associated with acquiring
HIV. Taking a HIV test means that the individual learns his HIV status with certainty and allows
him to avoid ex post errors in forming beliefs about the marginal cost of risky behavior. In addition,

11
when ART is available he may be able to begin treatment if he tests HIV positive.
The individual is endowed with a prior probability of being HIV positive, p0 ∈ [0, 1]. Given this
probability, the individual forms expectations about the net benefit of testing and decides whether
to take a HIV test. After deciding whether or not to take a HIV test and learning the results if he
chooses to test (and beginning treatment if he is HIV positive and eligible), the individual chooses an
amount of risky behavior in which to engage. The amount of risky behavior in which the individual
chooses to engage is summarized by the probability that he is HIV positive after engaging in this
risky behavior if in fact he was HIV negative prior to this choice. Let p1 denote this probability.

1.3.2

Demand for risky sexual behavior

Figure 6 provides a graphical representation of the mechanics of the model. For the moment,
consider the upper half of this figure. On the x-axis is the exogenous prior probability the individual
is HIV positive, p0 . On the y-axis is the choice of risky behavior, p1 . The concave function, p1 (p0 ),
represents the risky behavior demand function if the individual does not take a HIV test. Although
the p1 (p0 ) depicted in Figure 6 is increasing and concave in p0 , this function need not be increasing
nor concave in p0 in practice and this model does not assume this to be the case. If it is increasing
and concave in p0 , then the model indicates a role for HIV testing in reducing risky behavior. I will
explain why this is the case momentarily.
Suppose the individual knows with certainty that he is HIV negative (i.e., p0 = 0). Then he will
choose an amount of risky behavior given by p1 (0). This is the y-intercept for the p1 (p0 ) function.
Now suppose that the individual knows with certainty that he is HIV positive (i.e., p0 = 1). Then
he will choose an amount of risky behavior given by p1 (1). Note that p1 (0) and p1 (1) are bounded
below and above by zero and one but they need not equal these values.
In this model, taking a HIV test and receiving the results means that the individual knows his
HIV status with certainty. Thus, if the individual takes a HIV test then he will choose one of two
possible choices of risky behavior: p1 (0) or p1 (1). The result of the test determines which choice he
will make. In expectation, the individual’s choice of risky behavior if he were to take a HIV test is
simply the linear combination of these two points: (1 − p0 )p1 (0) + p0 p1 (0). This is the linear function
in the upper half of Figure 6.
It should now be clear that the expected change in choice of risky behavior associated with taking
a HIV test is the difference between the two functions plotted in the upper half of Figure 6. That
is, the expected change in risky behavior associated with taking a HIV test is given by

12

(1 − p0 )p1 (0) + p0 p1 (1) − p1 (p0 ).

(1.1)

If demand for risky behavior (i.e., p1 (p0 )) is concave in p0 , then the sign of this expression is negative
and in expectation HIV testing will reduce risky behavior. Conversely, if p1 (p0 ) is convex, then the
sign of this expression if positive and in expectation HIV testing will increase risky behavior. To fix
ideas, I will restrict the following analysis to the case where p1 (p0 ) is concave and in expectation
HIV testing will reduce risky behavior.
Let p∗ denote the value of p0 that maximizes the expected reduction in risky behavior associated
with taking a HIV test. For p0 < p∗ , the expected reduction in risky behavior associated with taking
a HIV test is increasing in p0 . For p0 ≥ p∗ , the expected reduction in risky behavior associated with
taking a HIV test is decreasing in p0 .

1.3.3

Demand for HIV testing

Now consider the lower half of Figure 6. Again, on the x-axis is the exogenous prior probability
the individual is HIV positive, p0 . On the y-axis is expected utility. There are two components to
expected utility in this model: the utility derived from engaging in (risky) sex and the utility loss
associated with acquiring HIV. The utility derived from the choice of risky sexual behavior, p1 (p0 ),
is denoted V (p1 (p0 )). The utility loss associated with acquiring HIV is denoted θ. I assume that
utility is additively separable in choice of risky sexual behavior and health status.
The convex function in Figure 6, E[U (not test)], is the expected utility of not taking a HIV test.
Additive separability means that I may write the expected utility of not testing as

E[U (not test)] = V (p1 (p0 )) − [1 − (1 − p0 )(1 − p1 (p0 ))]θ

(1.2)

where [1 − (1 − p0 )(1 − p1 (p0 ))] is the probability that the individual is HIV positive conditional on
the prior probability of being HIV positive, p0 , and on the probability that he acquires HIV through
his choice of risky behavior, p1 (p0 ).
Although Figure 6 depicts the expected utility of not taking a HIV test as being decreasing and
convex in p0 this need not be the case in practice and this model does not assume this to be true.
If it is decreasing in p0 , then the increase in the expected utility cost of becoming HIV positive
due to a increase in p0 (and the subsequent choice to increase p1 ) outweighs the utility gain (if
any) from engaging in a greater amount of risky behavior, p1 (p0 ), that results from the increase in

13
p0 . The convexity of E[U (not test)] follows from the fact that the marginal cost of additional risky
behavior is decreasing in the prior probability of being HIV positive. As p0 increases, the expected
utility of not testing falls because the probability that the individual is HIV positive and incurs the
associated utility cost is higher. On the other hand, the loss in expected utility from an increase
in p0 is partly offset by the increase in choice of risky behavior (given the current assumption that
p1 (p0 ) is increasing in p0 ). However, the increase in choice of risky behavior, p1 , itself increases the
probability that the individual acquires HIV. The magnitude of this latter effect (although always
non-negative) is decreasing in p0 and hence E[U (not test)] is convex.
To understand the expected utility of testing in the absence of ART, first consider the case where
the individual takes a HIV test and finds that he is HIV negative. Because he knows that he is HIV
negative with certainty, he will choose to engage in an amount of risky behavior given by p1 (0) and
the only risk of acquiring HIV is because of this choice. Thus, conditional on testing HIV negative
his expected utility is

V (p1 (0)) − p1 (0)θ − c

(1.3)

where c is the utility cost of taking a HIV test. Now consider the case where the individual takes a
HIV test and finds that he is HIV positive. Because he knows that he is HIV positive with certainty,
he will choose to engage in an amount of risky behavior given by p1 (1) and is certain of incurring
the utility cost of acquiring HIV. Thus, conditional on testing HIV positive his expected utility is

V (p1 (1)) − θ − c.

(1.4)

The expected utility of testing in the absence of ART is the linear combination of these two conditional expectations, where the weights are the prior probabilities of being HIV negative and of being
HIV positive, respectively. That is, the expected utility of testing in the absence of ART, denoted
E[U (test)] in Figure 6, is

E[U (test)] = (1 − p0 )[V (p1 (0)) − p1 (0)θ − c] + p0 [V (p1 (1)) − θ − c].

(1.5)

Assuming V (p1 (0)) − p1 (0)θ > V (p1 (1)) − θ (i.e., the individual would prefer to engage in p1 (0) and
acquire HIV with probability p1 (0) than to engage in p1 (1) and acquire HIV with certainty), implies
that

∂E[U (test)]
∂p0

< 0.

With expressions for the expected utility of not testing and the expected utility of testing in the

14
absence of ART, it is possible to examine predictions about the relationship between the testing
decision and the prior probability of being HIV positive. In the absence of ART, the model predicts that demand for HIV testing should be greatest among individuals in the middle of the p0
distribution. As long as the cost of testing is sufficiently high, individuals with prior probabilities
of being HIV positive that are either very low or very high will not choose to take a HIV test and
the only individuals that test will be those in the middle of the p0 distribution. The prediction is
apparent in the lower half of Figure 6: the expected utility of testing is greater than that of not
testing only for individuals in the middle of the p0 distribution. The intuition underlying this result
is that the expected change in risky behavior associated with taking a HIV test is relatively small
for individuals with very high (or very low) values of p0 . It is individuals that are in the middle of
the p0 distribution that have the largest expected change in demand for risky behavior associated
with taking a HIV test and hence the greatest willingness to pay for the test.
The introduction of ART (without rationing) does not affect the risky behavior demand function
or the expected utility of not testing, but it does affect the expected utility of testing. Specifically,
the expected utility of testing increases by the prior probability that the individual is HIV positive
times the net utility benefit of beginning ART conditional on being HIV positive. Let α denote the
net utility of beginning ART conditional on being HIV positive. Then the expected utility of testing
when there is ART, denoted E[U (test)] with ART in Figure 6, is

E[U (test)] with ART = (1 − p0 )[V (p1 (0)) − p1 (0)θ − c] + p0 [V (p1 (1)) − θ − c + α].

(1.6)

Suppose the individual takes a HIV test and finds that he is HIV negative. Given the set-up of the
model, to receive the utility benefit of beginning ART the individual needs to be HIV positive before
engaging in his choice of risky behavior, p1 . Thus, conditional on testing HIV negative his expected
utility when there is ART is identical to that if there were no ART (i.e., V (p1 (0)) − p1 (0)θ − c).
Now suppose the individual takes a HIV test and finds that he is HIV positive. Then the individual
will receive the treatment benefit of testing with certainty and his expected utility conditional on
testing HIV positive when treatment is available is given by

V (p1 (1)) − θ − c + α.

(1.7)

The expected utility of testing when there is ART is the linear combination of these two conditional

15
expectations where the weights are the prior probabilities of being HIV negative and of being HIV
positive, respectively. Assuming V (p1 (0)) − p1 (0)θ > V (p1 (1)) − θ + α (i.e., the individual would
prefer to engage in p1 (0) and acquire HIV with probability p1 (0) than to engage in p1 (1) and acquire
HIV (and begin ART) with certainty), implies that

∂E[U (test)] with ART
∂p0

< 0.

The model predicts that when ART is available (and there is no rationing) demand for testing
should be lowest among individuals with the lowest prior probabilities of being HIV positive. For
an individual with a low value of p0 , the expected treatment benefit of testing is relatively low
and the expected change in demand for risky behavior associated with taking a HIV test is also
relatively low. For an individual with a high value of p0 , the expected treatment benefit of testing is
relatively high but the expected change in demand for risky behavior associated with taking a HIV
test is relatively low. Whether individuals with high values of p0 or individuals with values of p0 in
the middle of the p0 distribution have the greatest demand for HIV testing when ART is available
depends on the magnitude of the treatment benefit of testing relative to the benefit associated with
the expected change in demand for risky behavior if the individual were to take a HIV test.

1.3.4

Change in risky sexual behavior

Linking the upper and lower halves of Figure 6 yields several important insights about the effect of
HIV testing on risky sexual behavior. First, consider the state of the world in which ART is not
available. The individual chooses to take a HIV test if and only if E[U (test)] ≥ E[U (not test)]. That
is, the individual will test if and only p0 ∈ A. The associated expected change in risky behavior for
this individual is (1 − p0 )p1 (0) + p0 p1 (1) − p1 (p0 ).
Now suppose that ART is available. As in the absence of ART, if p0 ∈ A then the individual
will choose to take a HIV test. Moreover, if p0 ∈ A1 or p0 ∈ A2 then the individual will also test. If
individuals are uniformly distributed on the unit interval, the number of individuals in the interval
A1 is greater than those in the interval A2 and hence the number of individuals induced to test
with p0 > max{p0 | p0 ∈ A} is greater than those induced to test with p0 < min{p0 | p0 ∈ A}.
Because the introduction of ART increases the mean p0 among those individuals taking a HIV test
as compared to the “no ART” case, the per-person reduction in risky behavior associated with HIV
testing when ART is available is greater than that in the absence of ART.

16

1.3.5

Effect of ART rationing

We may use Figure 6 to examine the effects of ART rationing on demand for HIV testing and
for risky behavior. In particular, we may examine the effect a rationing mechanism that favors
individuals that happen to have relatively low prior probabilities of being HIV positive (e.g., males
age 40-59). As compared to the no-rationing case, such a rationing mechanism would serve to reduce
the expected treatment benefit of testing among individuals with relatively high values of p0 while
affecting the expected treatment benefit of testing among low-p0 individuals less. Because it reduces
the expected treatment benefit of testing more among individuals expected to demonstrate larger
reductions in post-testing risky behavior, a rationing mechanism of this sort reduces the prevention
impact of ART-induced testing.

17

1.4

References

Attawell, Kathy and Jackie Mundy. 2003. Provision of Antiretroviral Therapy in Resource-Limited
Settings: A Review of Experience up to August 2003. MIMEO.
Bond, George. 1982. Education and Social Stratification in Northern Zambia: The Case of the
Uyombe. Anthropology and Education Quarterly, 13(3): 251-267.
Boozer, Michael and Tomas Philipson. 2000. The Impact of Public Testing for Human Immunodeficiency Virus. The Journal of Human Resources, 35(3): 419-446.
Colson, Elizabeth and Thayer Scudder. 1981. “Old Age in Gwembe District, Zambia”. in Pamela
Amoss and Stevan Harrell, eds. 1981. Other Ways of Growing Old. Stanford, CA: Stanford University Press.
Epstein, A. L. 1981. Urbanization and Kinship. New York, NY: Academic Press.
Global HIV Prevention Working Group. 2004. HIV Prevention in the Era of Expanded Treatment
Access.
Jones, Peris. 2004. Who is Getting or Will Get Anti-Retroviral Treatment in Zambia? Equity,
Access, and Fairness in the Governance of “Scaling Up” HIV/AIDS Medication. MIMEO.
Lakdawalla, Darius; Neeraj Sood; and Dana Goldman. 2006. HIV Breakthroughs and Risky Sexual
Behavior. The Quarterly Journal of Economics, 121(3): 1063-1102.
Magruder, Jeremy. 2006. Marital Shopping and Epidemic AIDS. MIMEO.
Philipson, Tomas and Richard Posner. 1995. A Theoretical and Empirical Investigation on the
Effects of Public Health Subsidies for STD Testing. The Quarterly Journal of Economics, 110(2):
445-474.
Scudder, Thayer. 1962. The Ecology of the Gwembe Tonga. Manchester, England: Manchester
University Press.
Stringer, Jeffrey; Isaac Zulu; Jens Levy; Elizabeth Stringer; Albert Mwango; Benjamin Chi; Vilepe
Mtonga; Stewart Reid; Ronald Cantrell; Marc Bulterys; Michael Saag; Richard Larlink; Alwyn
Mwinga; Tedd Ellerbrock; and Moses Sinkala. 2006. Rapid Scale-Up of Antiretroviral Therapy at
Primary Care Sites in Zambia: Feasibility and Early Outcomes. Journal of the American Medical

18
Association, 296(7): 782-793.
Thornton, Rebecca. 2006. The Demand for and Impact of Learning HIV Status: Evidence from a
Field Experiment. MIMEO.
UNAIDS/WHO. 2006. Epidemiological Fact Sheets on HIV/AIDS and Sexually Transmitted Infections.
Wawer, Maria; Ronald Gray; Nelson Sewankambo; David Serwadda; Xianbin Li; Oliver Laeyendecker; Noah Kiwanuka; Godfrey Kigozi; Mohammed Kiddugavu; Thomas Lutalo; Fred Nalugoda;
Fred Wabwire- Mangen; Mary Meehan; and Thomas Quinn. 2005. Rates of HIV-1 Transmission
Per Coital Act, by Stage of HIV-1 Infection, in Rakai, Uganda. The Journal of Infectious Diseases,
191: 1403-9.
World Health Organization (WHO). 2006a. WHO News Release. August 2006.
World Health Organization (WHO). 2006b. Progress on Global Access to HIV Antiretroviral Therapy: A Report on “3 by 5” and Beyond.

.25

Proportion HIV positive
.05
.1
.15
.2

0
15

20

25
30

Age
Females

35

40

Males

45

50

55

60

1.5

Figure 1.1: HIV Prevalence, 2001 Zambia DHS

19

Appendix

0

50

Pre-ART

ART era

100
150
Kilometers to nearest ART site

200

Figure 1.2: HIV Testing and Proximity to ART

250

20

Proportion t ested in past 12 months
.05
.1
.15
.2

0

15

20
2000

25

30

Age

2003

35

40
2005

45

Figure 1.3: HIV Testing and Age, Females ZSBS

50

21

Proportion t ested in past 12 months
.02 .04
.06 .08
.1
.12

0

15

20

25
2000

30

35

Age

2003

40

45

2005

50

55

Figure 1.4: HIV Testing and Age, Males ZSBS

60

22

Proportion t ested in past 12 months
.02
.04
.06
.08
.1
.12

0

15

20

25

30

Age
Females

35

40

Males

45

50

55

Figure 1.5: Proximity to VCT and Age, 2005 ZSBS

60

23

25

Kilometers to nearest VCT site
5
10
15
20

0

24

Figure 1.6: Demand for HIV Testing and for Risky Behavior
p1(p0)
p1(1)

(1- p0)p1(0)+ p0p1(1)

p1(0)

V(p1(0)) - p1(0)θ

A2

A

A1

p*

1

p0

V(p1(0)) - p1(0)θ - c

V(p1(1)) - θ
V(p1(1)) - θ - c + α
V(p1(1)) - θ - c

E[U(not test)]
E[U(test)] with ART
E[U(test)]

p0

Chapter 2

A Model of Demand for HIV
Testing: Structural Estimation and
Results
2.1
2.1.1

Data and estimation
Data

I use data on the outcomes of individuals’ testing decisions (i.e., whether or not they chose to take a
HIV test) and data on HIV prevalence by demographic group to estimate the structural parameters
of the model. I do not observe choices about post-testing risky behavior. In addition, I do not
observe the results of the voluntary tests reported in the ZSBS.
The 2001 Zambia DHS (ZDHS) includes the results of anonymous HIV testing linked to information on the age, gender, and province of residence of each individual tested. The sample of
individuals for whom a HIV test was requested was defined to be all men and all women in one-half
of households selected for the men’s survey module in the ZDHS. In total, approximately 5,000 individuals were asked to provide a blood sample. Among women approached to provide a sample, 15.4
percent refused and 5.3 percent were either absent or the result is missing. Among men approached
to provide to a sample, 14.8 percent refused and 12.1 percent were either absent or the result is
missing. The resulting sample consists of testing results for 3,949 observations.
Because the sample size for any given age-gender-province interaction is small, individuals are
25

26
grouped into five-year age ranges (e.g., 15-19 years old, 20-24 years old, etc.). The ZDHS includes
women age 15-49 and men age 15-59 and there are nine provinces in Zambia, so this procedure yields
144 groups with HIV prevalence ranging from zero percent to 35.3 percent. For each group I treat
the proportion that were found to be HIV positive as the initial probability of being HIV positive,
p0 .1 I assume that the age-profile of HIV prevalence is static for a given gender-province interaction
over the period covered by the repeated cross-sectional data used in this analysis, the Zambia Sexual
Behavior Survey (ZSBS).
The data on whether an individual voluntarily chose to take a HIV test come from the ZSBS. I
use the three most recent rounds of the ZSBS: 2000, 2003, and 2005. The sample sizes for each of
these survey rounds are 3316, 4471, and 4208 individuals, respectively. To match the testing data to
the prevalence data the testing decisions of individuals in the ZSBS are aggregated by demographic
group defined as the interaction of five-year age group, gender, and province of residence. Moreover,
for each demographic group I aggregate the two pre-ART survey years (i.e., 2000 and 2003). I then
calculate the proportion of individuals that report taking a HIV test in the 12 months prior to the
survey date, denoted testi . This procedure yields 288 observations.

2.1.2

Estimation

For a group of individuals with the homogeneous belief they are HIV positive equal to p0i , the
proportion predicted to choose to take a HIV test in the pre-ART period is given by

di =
test

1
1 + e−(E[U (test)]i −E[U (not test)]i )/δ

(2.1)

where δ is a smoothing parameter that is to be estimated along with the other parameters of the
model. The empirical specification for E[U (test)]) includes as additive terms an indicator variable
equal to one if the observation is a male demographic group, denoted male, and an indicator variable,
1 Although the available Zambian data do not include information on the self-reported likelihood of being HIV
positive, there are data from Malawi that do include this information. The Malawi Diffusion and Ideational Change
Project (MDICP) includes information on respondents’ subjective beliefs of being HIV positive as recorded on a scale
of “no likelihood”, “low”, “medium”, and “high”. A priori, it is reasonable to think that beliefs about HIV status
in Zambia should be similar to those in Malawi. Malawi is adjacent to Zambia, landlocked (as is Zambia), and has a
HIV prevalence rate of approximately 14 percent (roughly equal to that in Zambia.) Over the age ranges available in
the 2001 ZDHS, the belief-age profiles for females and males in the 2001 survey round of the MDICP appear to be
similar to the prevalence-age profiles in the 2001 ZDHS: both for females and for males, it is individuals in the middle
of the age distribution that tend to report the greatest likelihood of being HIV positive. One difference between the
belief-age profiles in the 2001 survey round of the MDICP and the prevalence-age profiles in the 2001 ZDHS is that
young women age 15-19 in the MDICP report beliefs of being HIV positive that are as high as those reported by
women in the middle of the age distribution. However, in the 2001 survey round of the MDICP there are only 24
women between the ages of 15 and 19 as compared to 124 in the five-year age group with the second fewest number of
observations. Moreover, I fail to reject the hypothesis that the mean belief among women ages 15-19 is significantly
different from that among women ages 20-24.

27
denoted post, equal to one if the the observation comes from the 2005 survey round of the ZSBS.
In the ART era, the predicted proportion testing in group i is generated by Equation (8) as
well, but with an expanded expression for E[U (test)] with ART substituted for E[U (test)]. As
discussed previously, the changes in the testing-age profiles discussed in Section 2 differ from those
predicted by the theoretical model. The model predicts that the individuals who demonstrate the
greatest testing response to the introduction of ART should have prior probabilities of being HIV
positive that are at least as large as those among individuals that were more likely to test prior
to the introduction of ART. In contrast, the observed testing response among males do not fit this
prediction. One explanation for this puzzle is that there exists a non-random rationing mechanism
for allocating ART. In this case, we may expect that the probability that an individual is favored in
this process is correlated with his demographic characteristics.
To capture this possibility, I estimate the probability that the individual in group i will receive
0

ART conditional on p0i , denoted p0i , by including measures of each individual’s age, age squared,
gender, and interaction terms in the empirical specification for the expected utility of testing when
ART is available. Specifically, I include the righthand side of following expression as an additive
term in the empirical specification for E[U (test)] with ART (i.e., Equation 6)

0

p0i =

1
1 + eXi β

(2.2)

where Xi is a vector whose elements include age, age squared, gender, and interaction terms. In
addition, I examine the relative importance of this rationing mechanism by estimating weights for
the effects of the prior probability of being HIV positive (i.e., p0 ) and of the probability of meeting
0

the non-random rationing criteria (i.e., p0 ). These weights are denoted (1 − σ) and σ, respectively.
Thus, the expanded expression for E[U (test)] with ART that I use in the empirical specification is

E[U (test)] with ART = (1 − p0 )[V (p1 (0)) − p1 (0)θ − c] + p0 [V (p1 (1)) − θ − c]
0

+ (1 − σ)αp0 + σαp0 .

(2.3)

Because the ultimate product of the empirical analysis is a set of policy simulations, I need to
parameterize the model. In particular, I need to specify a functional form for the utility derived
from choice of risky behavior, V (p1 ). I choose a functional form for V (p1 ) that is sufficiently flexible
to allow a variety of shapes for the risky behavior demand function, p1 (p0 ). Specifically, the utility
derived from choice of risky behavior is given by

28

V (p1 ) =

(p1 − eb )2
.
ep1 − 1

(2.4)

The flexibility of this specification means that I am able to let the data confirm that the risky
behavior demand function is in fact increasing and concave in p0 , as was assumed in the discussion
of the model in Section 3.
To estimate the structural parameters of the model, I minimize a simple weighted sum of squared
residuals
288
X
popi
i=1

pop

di )2
(testi − test

(2.5)

where the weights are given by the proportion of the total adult population (pop) that is in p0 group
i (popi ).

2.2
2.2.1

Results
Parameter estimates

Table 1 provides the parameter estimates that are the solution to the weighted non-linear least
squares minimization problem. Column (1) lists the estimates for the baseline model with no additional controls. The baseline model includes the variables indicated by the theoretical model, as
well as an indicator variable for the “post” period and an indicator variable for male. Substituting these parameters into Equation (2) and examining the first and second derivatives verifies that
the expected utility of not testing is decreasing and convex in the prior probability of being HIV
positive. In addition, the utility cost of testing is positive (i.e., c > 0), utility is decreasing in the
probability that the individual eventually acquires HIV (i.e., θ > 0), males appear to be slightly
more likely to test than females, and there is evidence of a secular increase in testing rates unrelated
to the introduction of ART. However, neither of these two latter effects are significantly different
from zero.
The parameter estimates indicate that individual behavior is broadly consistent with a model
in which ART is not allocated to all HIV positive individuals, but is instead rationed based on
the individual’s age and gender (or factors associated with these characteristics). In the baseline
specification, the estimated σ equals 0.97. In other words, the probability that an individual responds
to the availability of ART by testing is increasing in the probability that he is HIV positive but is

29
also correlated with the individual’s demographic characteristics conditional on p0 . Of course the
fact that σ < 1 indicates that individual behavior is consistent with a model where the probability
that an individual is HIV positive also affects the testing decision irrespective of these demographic
characteristics. Nonetheless, this effect appears to be small and it appears that there exists a nonrandom rationing mechanism that does affect testing behavior.
If province-level unobserved heterogeneity in testing demand is correlated with variables in the
model, then the estimates presented in Column (1) are biased. For example, individuals in highprevalence provinces may face lower costs of testing than individuals in low-prevalence provinces
(perhaps because policymakers may place more VCT sites in high-prevalence areas). To address this
concern (as well as other concerns) about time-invariant factors, Column (2) presents parameter
estimates from a model including province dummies as well as the two controls included in the
baseline model.
As shown in Column (2), the basic results of the model do not change after it is re-estimated with
province dummies. The expected utility of not testing is still decreasing and convex in the prior
probability of being HIV positive, θ and c still have the expected signs, and individual behavior
remains consistent with a model in which there exists substantial rationing of treatment based on
age.
As previously mentioned, the number of VCT sites increased between the pre- and post periods.
If the intensity of this expansion differed across provinces and this difference was correlated with
variables in the model, then we may be concerned that the estimates in Columns (1) and (2) are
biased. To address this possibility, I interact the “post” dummy with the province dummies and
include these in the expanded empirical specification. Column (3) presents the results of this exercise.
Again, the basic results do not change.

2.2.2

Predicted testing-age profiles

Comparing the predicted and observed testing rates is a simple method of assessing the fit of the
empirical specification discussed above. Figure 7 presents predicted and observed testing-age profiles
for females and for males in the pre-ART period and in the ART era. The predicted testing-age
profile in a given period in Figure 7 is the result of a locally weighted smoothed regression of the
predicted testing rate for group i on age. The observed testing-age profile in a given period in Figure
7 is the result of a locally weighted smoothed regression of an indicator variable equal to one if an
individual tests on age. For a given gender there are three panels in Figure 7, one panel for each of

30
the three empirical specifications discussed above.
The empirical specification appears to generate predicted values that are relatively close to the
observed HIV testing rates. The predicted testing rates tend to be within one percentage point or
less of the observed testing rate. The mean residuals are largest for the youngest females in the
ART era and for the youngest males in the pre-ART period. In general, the empirical specification
generates predicted testing-age profiles in the pre-ART period that take the inverted U-shape that
characterizes the observed testing-age profiles and predicted testing-age profiles in the ART era that
take the distinct shapes that characterize the observed testing-age profiles for females and males.

2.2.3

Predicted risky behavior demand function

Figure 8 presents the predicted risky behavior demand function, p1 (p0 ), from the empirical specification that includes the full set of controls. Although the empirical specification does not impose
that p1 (p0 ) be increasing or concave in p0 , the estimated risky behavior demand function is in fact
increasing and concave. The estimated p1 (p0 ) function implies that if an individual knew with certainty that they were HIV negative then they would choose an amount of risky sexual behavior in
which to engage that would imply that they would acquire HIV with probability 0.140. On the other
hand, if an individual believed with certainty that they were HIV positive then they would choose
an amount of risky sexual behavior in which to engage that would imply that they would acquire
HIV with probability 0.574 if they were in fact not HIV positive. The prior probability of being HIV
positive that maximizes the expected reduction in risky behavior associated with taking a HIV test
is 0.460. However, even at its maximum the expected reduction in risky behavior associated with
taking a HIV test in terms of the change in the probability the reference individual acquires HIV 0.047 - is relatively small.

2.3

Policy scenarios

This section uses the parameter estimates from the preceeding section to examine the behavioral and
welfare implications of five alternative ART policies: (i) no ART, (ii) the status quo ART policy, (iii)
expanding the supply of ART, (iv) eliminating the non-random rationing mechanism while holding
constant the quantity of ART supplied, and (v) simultaneously eliminating the non-random rationing
mechanism while expanding the supply of ART. I employ three metrics to examine the performance
of each of these policies. The first metric is the predicted demand for HIV testing (i.e., the HIV
testing rate). The second metric is the mean predicted choice of risky behavior. The third metric

31
is the simulated number of new infections. The first two metrics are relatively straightforward to
calculate. The simulated number of new infections, however, requires particular discussion.
I simulate the effect of testing on new infections as follows. Each individual in a given demographic group may be characterized along two dimensions: whether they take a HIV test or not and
whether they are HIV positive or not. In a given demographic group, whether an individual chooses
to take a HIV test or not (as well as the result of the test) affects the individual’s choice of risky
behavior. Only individuals that are HIV negative are at risk of acquiring a new infection.
The number of new infections among individuals taking a HIV test is given by
N
X

di (1 − p0i )pb1 (0)
ni test

(2.6)

i=1

where ni is the number of individuals in group i and N equals 144, the number of demographic
groups. For demographic group i, the product of the first three terms in this expression yields the
number of individuals in this group that are at-risk of acquiring HIV. Because these individuals are
HIV negative and the results of the HIV tests reveal this fact to each of the individuals, they choose
to engage in an amount of risky behavior summarized by pb1 (0); this is the probability that they
acquire HIV.
The number of new infections among individuals not taking a HIV test is given by
N
X

di )(1 − p0i )pb1 (p0i ).
ni (1 − test

(2.7)

i=1

Again, the product of the first three terms in this expression yields the number of individuals in
demographic group i that are at-risk of acquiring HIV. Because these individuals choose not to take
a HIV test, they engage in an amount of risky behavior summarized by pb1 (p0 ); this is the probability
that they acquire HIV. Total new infections are given by the sum of these two expressions.
This method of simulating the number of new infections has one major shortcoming. Namely, it
ignores new infections among partners of the reference individuals. Thus, it is important to recognize
that this method measures the expected number of new infections in a given population with error.

2.3.1

No ART

The first policy scenario I examine is a scenario where there is no ART. Applying the three metrics
to this case provides a baseline for understanding the impact of each of four other ART policies. I
calculate the predicted HIV testing rate in the “no ART” scenario by shutting down the treatment

32
benefit to testing (i.e., setting α equal to zero), setting post equal to one, and generating predicted
values using Equation (11). The first row of Table 2 presents the results of this calculation for each
of the three empirical specifications. Model (1) is the baseline empirical specification with controls
for “post” (i.e., the ART era) and for gender. Using the parameter estimates from this specification,
the predicted HIV testing rate in the state of the world where there is no ART is 6.1 percent. Model
(2) includes province dummies as well as the controls included in Model (1). As Column (2) of
Table 2 indicates, the predicted testing rate does not change substantially, consistent with the fact
that the parameter estimates did not change substantially. The simulation results indicate that the
predicted HIV testing rate is 5.6 percent. Model (3) includes an indicator variable for the “post”
period interacted with the province dummies as well as all of the controls in Model (2). Again,
neither the parameter estimates nor the predicted testing rate changes dramatically. Using the
parameter estimates from Model (3), I estimate that the predicted HIV testing rate in the absence
of ART would be 5.5 percent.
The second metric is the mean choice of risky behavior. The first row of Table 3 presents the mean
choice of risky behavior under the “no ART” scenario for each of the three empirical specifications.
The parameter estimates from Model (1) imply that the mean choice of risky behavior would be
0.1361. This implies that the representative individual would choose to engage in an amount of
risky behavior that would lead them to acquire HIV with probability 0.1361 if they were in fact HIV
negative at the time of this choice. Models (2) and (3) imply substantially higher mean choices of
risky behavior: 0.1651 and 0.1554, respectively. The riskier behavior implied by Models (2) and (3)
is partly due to the fact that fewer individuals are predicted to test (and hence learn with certainty
that they are HIV negative). In addition, the estimated risky behavior demand function given by the
parameter estimates associated with each of these empirical specifications predicts riskier behavior
conditional on a negative test result than does Model (1).
The third metric is the simulated number of new infections. Table 4 presents the results of this
simulation; the first row of this table represents the “no ART” scenario. According to the parameter
estimates from Model (1), the number of new infections in the state of the world where there is
no ART appears relatively large: 927,000 new infections. Of course the theoretical model allows
for one chance to take a HIV test and one chance to make a decision about future risky behavior.
Thus, it may make sense to think about the simulated number of new infections as representing the
number of new infections in the population over the course of the average lifetime of an individual
in the population. Given that on average an individual between the age of 15 and 59 in Zambia is
expected to live approximately 25 more years (WHO 2007), then the simulated annual incidence is

33
nearly 40,000. As a point of comparison, the annual incidence of new adult HIV cases in Zambia is
approximately 50,000. Consistent with the higher mean choice of risky behavior in Models (2) and
(3) than in Model (1), the simulated number of new infections in Models (2) and (3) - 1,000,000 and
940,000, respectively - are greater than that in Model (1).

2.3.2

Status quo ART policy

The second policy scenario I consider is the status quo ART policy. To calculate the three metrics
under this policy scenario, I set post equal to one and make no additional changes. The second row
in Table 2 presents the predicted proportion of individuals choosing to take a HIV test under this
scenario. Depending on the empirical specification, between 7.5 and 8.4 percent of individuals are
predicted to take a HIV test. This represents an increase in testing demand of between 2 and 2.3
percentage points (an increase of nearly 40 percent), regardless of the empirical specification.
Although the effect is not nearly as large, ART provision under the status quo ART policy is
also predicted to reduce mean choice of risky behavior. As presented in the second row of Table 3,
Models (1) through (3) imply that the mean choice of risky behavior would fall by between 0.07 and
0.30 percent (or thousandths of a percentage point).
The second row of Table 4 presents the the simulated number of new infections under the status
quo ART policy. Depending on the empirical specification, the number of new infections ranges from
916,000 to 989,000. These figures represent a decline in the incidence of HIV of between 0.96 and
1.19 percent. In general the effect of ART-induced testing under the status quo ART policy appears
to be modest at best. However, comparing these figures with the performance of the alternative
policies should yield a more complete understanding of the relative magnitude of this effect.

2.3.3

Expanding the supply of ART

The third policy scenario I examine is expanding the supply of ART while retaining the existing
source of inefficiency in the process determining the allocation of ART. If the quantity of antiretroviral drugs available in Zambia is not sufficient to meet the demand for ART, then treatment providers
may ration ART among the pool of eligible individuals without regards to each individual’s identity.
I consider the policy of expanding supply to ensure that all individuals that would be allocated
ART under the existing non-random rationing mechanism are able to begin treatment. Expanding
the supply of ART should yield an increase in the prevention impact of ART-induced testing as
compared to the status quo ART policy. Because expanding the supply of ART increases the proba-

34
bility that an individual will receive the payoff associated with taking a HIV test conditional on the
prior probability the individual is HIV positive (and on the probability that they would be allocated
ART under the non-random rationing mechanism), it should increase the proportion of individuals
testing. The cost of this policy is simply the increase in drug expenditures (and associated labor
costs) required to cover these additional patients.
It appears that the quantity of antiretroviral drugs in Zambia is not sufficient to meet the demand
for ART. I calculate the ratio of the number of individuals on ART at the time of the 2005 round
of the ZSBS (i.e., 30,000) to the number of individuals that are predicted to test in the 2005 round
of the ZSBS and to be eligible for ART. Let q denote this ratio. Then

q = PN

i=1 ((1

30, 000

(2.8)

0

di ni
−σ
b)pc
bpc
0i + σ
0i )test

where N equals 144, the number of demographic groups. Consistent with the existence of excess
demand for ART, I find that this ratio is substantially less than one (q = 0.59). If there are
individuals that did not take a HIV test in the 12 month interval but were still eligible for ART and
able to demonstrate their status to the health authorities (e.g., from the results of a test taken prior
to the beginning of the 12 month interval), then the q given by Equation (15) is going to understate
the extent of rationing of ART using a random lottery. On the other hand, not all HIV positive
individuals are recommended to begin ART and so the q given by Equation (15) may overstate the
extent of rationing of ART using a random lottery.
To calculate the expected level of testing associated with ART availability under the expanded
supply scenario, first I substitute the parameter estimates from Table 1 (except the estimated α)
into the predicted testing equation and set post equal to one. Then I substitute

1
b
0.59 α

for α in the

predicted testing equation (where α
b is the estimated α). Because the effect of rationing without
regards to individuals characteristics is not separately identified from the benefit of beginning ART
(i.e., α), eliminating the rationing of ART using a random lottery is equivalent to increasing α.
The third set of rows of Table 2 present the results of this exercise. Depending on the empirical
specification, I predict that expanding the supply of ART while retaining the existing non-random
ART rationing mechanism should increase testing rates to between 9.2 and 10.3 percent. This
represents an increase in testing rates as compared to the status quo ART policy of between 1.7 and
1.9 percentage points.
The third set of rows in Table 3 presents the mean choice of risky behavior under the “expand
supply” scenario. Although the mean choice of risky behavior in this scenario is smaller than that

35
under either the “no ART” or the “status quo ART” scenarios, the reduction in mean risky behavior
is still small in an absolute sense. As compared to the “no ART” scenario, mean risky behavior is
between 0.15 and 0.60 percent smaller under the “expand supply” scenario.
The simulation results fail to provide support for the idea that the continued roll-out of ART
(in the absence of other interventions designed to eliminate the non-random allocation process) will
induce substantial reductions in the incidence of new HIV infections. As shown in the third set of
rows in Table 4, the number of new infections avoided under this policy as compared to the state of
the world without ART is only between 1.81 and 2.27 percent. In relative terms, however, this policy
is much more effective than the status quo ART policy: simply expanding the supply of ART to
eliminate the random lottery increases the number of new infections avoided by at least 88 percent.

2.3.4

Eliminating the source of inefficiency

The fourth policy scenario is to eliminate the existing source of inefficiency in the ART allocation
process while holding constant the quantity of ART supplied. Under this policy, all individuals
that are HIV positive are eligible for ART and age and gender no longer affect the ART allocation
decision conditional on HIV status. One way to implement this policy would be to mandate that
the enrollment-age profile at ART sites in each province match the province’s prevalence-age profile.
Because the existing supply of ART is held constant, eliminating the inefficiency would require no
additional expenditure on antiretroviral drugs. The tradeoff or cost of this policy is that individuals
that are currently favored in the ART allocation process would face a lower probability of beginning
ART conditional on being HIV positive. Moreover, because the quantity of ART supplied is held
constant while the number of individuals eligible for ART increases, the probability that an HIV
positive individual that takes a HIV test is allocated ART in the random lottery decreases.
Eliminating the source of inefficiency in the process determining the allocation of ART has the
potential to yield a large increase in the number of new infections avoided. The model provided
in this paper indicates that the expected reduction in risky behavior associated with taking a HIV
test is increasing in the initial probability of being HIV positive over the range relevant in the
current setting. However, empirical evidence indicates that current ART policy effectively favors
populations (i.e., older males) with prior probabilities of being HIV positive that are much lower
than the maximum prior probability observed in the Zambia DHS. Thus, eliminating the source of
the inefficiency will increase the number of new infections avoided due to ART-induced testing in
part because the composition of the pool of individuals choosing to take a HIV test will shift toward

36
individuals expected to demonstrate relatively large reductions in post-testing risky behavior. In
addition, there are many more individuals in groups that are predicted to voluntarily test at high
rates in the absence of the source of inefficiency (e.g., males in the middle of the age distribution)
than there are in the group (i.e., older males) that currently tests at a high rate. Hence, eliminating
the source of inefficiency will increase the number of new infections avoided due to ART-induced
testing in part because the number of individuals testing after the inefficiency is addressed would be
much greater than the number that test under the status quo ART policy.
I operationalize this policy experiment by generating predicted testing values using the parameter
estimates from Table 1 (except the estimated α), setting post = 1, σ = 0, and recalibrating q. I
recalibrate q to reflect the fact that if the policy eliminates the non-random rationing mechanism
then the number of individuals that respond to ART by testing will be greater. The new value
of q is 0.3. Because the quantity of ART supplied is held constant in this policy, the increase in
testing demand means a decrease in the chance of being selected in the random lottery. Therefore, I
substitute 0.3b
α for α in the predicted testing equation. Again, because the effect of a random lottery
determining the allocation of ART among HIV positive individuals is not separately identified from
the benefit of beginning ART (i.e., α), the effect of decreasing the probability an individual is selected
in the lottery is equivalent to decreasing the benefit of beginning ART.
Regardless of the metric under consideration, the simulated results of this policy provide clear
evidence of the cost of the non-random rationing mechanism identified in this research. Predicted
testing demand is between 33.8 and 36.2 percent, a more than three-fold increase over the that under
the status quo ART policy. The percent reduction in mean risky behavior as compared to the “no
ART” policy is between 0.88 and 3.39, at least eleven times greater than the reduction under the
status quo ART policy. As compared to the status quo policy, the number of new infections avoided
due to ART-induced testing would be more than ten times as large if the non-random rationing
mechanism were eliminated: between 139,000 and 161,000 new infections avoided.

2.3.5

Eliminating the inefficiency while expanding the supply of ART

The fifth policy scenario is to expand the supply of ART in combination with eliminating the existing
source of inefficiency. In particular, I consider an expansion in the supply of ART that is sufficient to
ensure that all individuals that are HIV positive would be able to begin ART. This policy yields the
both the mechanical benefit of increasing ART-induced testing by eliminating the random lottery for
ART as well as the composition benefit of inducing additional testing among individuals expected

37
to demonstrate greater post-testing reduction in risky behavior. Of course this policy requires
the greatest increase in expenditure. The increase in expenditure on antiretroviral drugs (and the
associated labor costs) is greater than that under the policy of expanding the supply of ART while
retaining the existing source of inefficiency. Moreover, eliminating the source of inefficiency may
itself be costly.
To simulate the effect of this new policy, I set post = 1, σ = 0, replace α with

1
b,
0.59 α

and substitute

the parameter estimates from Table 1 for the remaining parameters. The simulation results indicate
that testing demand under this policy would be between 52.1 and 54 percent, a more than seven-fold
increase as compared to the “no ART” policy. The mean choice of risky behavior would be between
5.51 and 8.96 percent smaller than under the “no ART” policy. In terms of new infections avoided,
this policy would have very large effects: the number of new infections would be at least 23 percent
(28 percent) smaller than that under the “no ART” scenario (“status quo” scenario).

2.4

Conclusion

The debate about appropriate HIV/AIDS policy often features an analytic framework in which
interventions are classified as treatment interventions or as prevention interventions (see Canning
2006). Antiretroviral drugs and care for opportunistic infections (e.g., tuberculosis) are usually seen
as treatment interventions. HIV testing and condoms are usually seen as prevention interventions.
However, many proponents of HIV testing argue that antiretrovirals are not only a treatment intervention and that the availability of ART will induce substantial increases in testing demand and
subsequent reductions in risky behavior.
The results of the current analysis indicate that the prevention impact of antiretrovirals under the
existing ART policy in Zambia is modest at best. The observed changes in testing rates associated
with the introduction of ART are not consistent with the prediction of a model of testing demand
where the incentive effect of ART is increasing in the prior probability that the individual is HIV
positive. Instead, there appears to exist a source of inefficiency in the process determining the
allocation of ART. Because of this inefficiency, the number of new infections avoided due to ARTinduced testing is relatively small; those individuals largely favored in the ART allocation process are
predicted to demonstrate only small reductions in risky behavior subsequent to testing. Simulation
results indicate that ART-induced testing under the existing ART policy reduces the number of new
infections by approximately 1 percent.
Eliminating the inefficiency in the process determining the allocation of ART yields substan-

38
tial increases in the prevention impact of ART. Simulation results indicate that this intervention
would increase the number of new infections avoided due to ART-induced testing by more than
ten-fold. Moreover, this intervention requires no additional expenditure on antiretroviral drugs. In
comparison, expanding the supply of ART without eliminating this inefficiency would only double
the number of new infections avoided due to ART-induced testing while requiring a substantial increase in expenditure on antiretroviral drugs. The results of this paper confirm the emphasis among
policymakers on the central role of ART-induced HIV testing in reducing the spread of HIV/AIDS
and suggest that future research examine the process determining the allocation of ART in regions
experiencing a rapid change in the availability of ART.

39

2.5

References

Canning, David. 2006. The Economics of HIV/AIDS in Low-Income Countries: The Case for
Prevention. Journal of Economic Perspectives, 20(3): 121-142.

40

2.6

Appendix
Figure 2.1: Predicted and Observed Testing-Age Profiles
Model 1, Males

0

0

Tested for HIV in Pas t 12 Months
.02 .04 .06 .08 .1
.12

Tested for HIV in Pas t 12 Months
.02 .04 .06 .08
.1
.12

Model 1, Females

20

15

35
30
Age

25

Observed values, pre-ART
Predicted values, pre-ART

40

45

50

15

Observed values, ART era
Predicted values, ART era

25

30

35
40
Age

45

50

55

60

Observed values, ART era
Predicted values, ART era

Tested for HIV in Pas t 12 Months
.02 .04 .06 .08
.1
.12

Model 2, Males

0

0

Tested for HIV in Pas t 12 Months
.02 .04 .06 .08
.1
.12

Model 2, Females

20

Observed values, pre-ART
Predicted values, pre-ART

20

15

25

30

Age

35

Observed values, pre-ART
Predicted values, pre-ART

40

45

50

15

Observed values, ART era
Predicted values, ART era

25

30

35
40
Age

45

Observed values, pre-ART
Predicted values, pre-ART

50

55

60

Observed values, ART era
Predicted values, ART era

Tested for HIV in Pas t 12 Months
.02 .04 .06 .08
.1
.12

Model 3, Males

0

Tested for HIV in Pas t 12 Months
.1
0
.02 .04 .06 .08
.12

Model 3, Females

20

15

20

25

35
30
Age

Observed values, pre-ART
Predicted values, pre-ART

40

45

50

Observed values, ART era
Predicted values, ART era

15

20

25

30

35
40
Age

Observed values, pre-ART
Predicted values, pre-ART

45

50

55

60

Observed values, ART era
Predicted values, ART era

41

Figure 2.2: Estimated Risky Behavior Demand Function
p1

p^1(p0)

(1- p0)p^1(0)+p0p^1(1)

p0

42

Table 2.1: Parameter Estimates

Dependent variable:
b
e
c
θ
α
σ
δ
post
male
province controls
post*province controls

tested in past 12 months
(1)
(2)
(3)
1.686
1.673
1.690
(0.137)
(0.132)
(0.141)
0.541
0.549
0.543
(0.151)
(0.159)
(0.158)
0.084
0.075
0.077
(0.031)
(0.027)
(0.035)
0.862
0.850
0.824
(0.095)
(0.098)
(0.092)
0.499
0.488
0.482
(0.235)
(0.249)
(0.237)
0.971
0.964
0.973
(0.143)
(0.146)
(0.144)
0.021
0.022
0.021
(0.021)
(0.022)
(0.019)
0.001
0.001
0.001
(0.034)
(0.036)
(0.031)
0.002
0.003
0.002
(0.037)
(0.038)
(0.037)
no
yes
yes
no
no
yes

Observations
288
288
288
Note: Tested in past 12 months is a binary variable indicating took HIV
test in 12 months period preceeding survey.
"Post" is an indicator variable equal to one if the observation comes
from the 2005 survey round of the ZSBS.
"Male" is an indicator variable equal to one if the observation is a male
demographic group.

43

Table 2.2: Testing Rates Under Alternative Policy Scenarios

Empirical specification:

model 1
(1)

model 2
(2)

model 3
(3)

Policy scenario
No ART

6.1

5.6

5.5

Status quo
Percent change compared to "No ART"

8.4
38

7.8
39

7.5
36

Expand supply
Percent change compared to "No ART"

10.3
69

9.7
73

9.2
67

Eliminate inefficiency
Percent change compared to "No ART"

36.2
493

33.8
504

33.5
509

Expand supply and eliminate inefficiency
54.0
52.3
Percent change compared to "No ART"
785
834
Note: Testing rates reported as percent of females age 15-49 and
males age 15-59 choosing to take a HIV test.

52.1
847

44

Table 2.3: Mean Choice of Risky Behavior Under Alternative Policy Scenarios

Empirical specification:

model 1
(1)

model 2
(2)

model 3
(3)

Policy scenario
No ART

0.1361

0.1651

0.1554

Status quo
Percent change compared to "No ART"

0.1360
-0.07

0.1646
-0.30

0.1550
-0.26

Expand supply
Percent change compared to "No ART"

0.1359
-0.15

0.1641
-0.61

0.1546
-0.51

Eliminate inefficiency
Percent change compared to "No ART"

0.1349
-0.88

0.1595
-3.39

0.1508
-2.96

Expand supply and eliminate inefficiency
0.1286
0.1503
0.1423
Percent change compared to "No ART"
-5.51
-8.96
-8.43
Note: Mean choice of risky behavior is mean probability individual acquires HIV
through choice of risky behavior, where all individuals treated as if HIV negative
at time of choice.

45

Table 2.4: New Infections Under Alternative Policy Scenarios

Empirical specification:
Policy scenario
No ART

model 1
(1)

model 2
(2)

model 3
(3)

927

1,000

940

Status quo
Percent change compared to "No ART"

916
-1.19

989
-1.10

931
-0.96

Expand supply
Percent change compared to "No ART"

906
-2.27

979
-2.10

923
-1.81

Eliminate inefficiency
Percent change compared to "No ART"

766
-17.37

850
-15.00

801
-14.79

Expand supply and eliminate inefficiency
Percent change compared to "No ART"
Note: New infections reported in thousands.

686
-26.00

768
-23.20

723
-23.09

Chapter 3

Identity, Parochial Institutions,
and Career Decisions: Linking the
Past to the Present in the
American Midwest1
3.1

Introduction

The textbook economic model assigns individuals to occupations on the basis of their ability, the
opportunities they have to invest in human capital, and the types of jobs that are subsequently made
available to them in the labor market. Outside of economics, social scientists have long argued that
factors seemingly unrelated to economic incentives and constraints can also determine career choices.
Weber’s (1930) writings on the Protestant ethos are the classic reference in this literature. More
recently, sociologists have described how working class communities can shape the aspirations their
members (Gans 1962). Our objective in this paper is to document the presence of such non-economic
carrer motivations in the U.S. labor market, explore the reasons why such motivations could arise,
and understand why they might persist across many generations.
We start with the idea that social groups can endogenously instill values in their members
that increase loyalty and reduce the propensity to exit, particularly when they are vulnerable to
1 This

is joint work with Kaivan Munshi.

46

47
acculturation (Bisin and Verdier 2000) or in competition with other groups (Sumner 1906). The
negative consequences of population heterogeneity, typically measured by fractionalization along
various economic and social dimensions, are well documented in the economics literature.2 While
ethnic or racial fractionalization may adversely affect the performance of secular institutions, we
argue that such fractionalization can at the same time strengthen within-group loyalty and parochial
institutions.
In the setting that we consider – the American Midwest – the labor market was the primary
domain in which ethnic groups historically interacted. Migrant workers had to bear most of the
fluctuations in labor demand in the early stages of industrialization in the United States (Hoerder
1991). This is precisely the situation in which labor market networks are most useful and there is
a wealth of historical evidence documenting the role played by ethnic migrant networks in finding
jobs for their members at this time. With the arrival of the railroads around 1850, the Midwest
in particular witnessed a large influx of European migrants in response to the new occupational
opportunities that became available. Historians describe the efforts made by migrant communities
to establish “occupational beachheads” and “toe-holds” in new locations, and to subsequently work
hard to maintain these coveted positions once they were established. We will show that the cost
to the network from the exit of one of its members and, hence, efforts to instill a sense of loyalty
to the local community would have been greatest in fractionalized labor markets with many ethnic
networks competing for coveted permanent jobs.
Once such a loyalty to the local community was instilled, how and why was it maintained over
many generations in the future? One institution that could have maintained local values from one
generation to the next is the family. Bisin, Topa, and Verdier (2004), for example, describe how
parents in minority communities in the United States socialize their children more strictly to prevent
them from assimilating. In the current context, however, it is not clear why parents acting on their
own would place the welfare of the community above their own children’s desire to be mobile. Given
the high rates of inter-county migration in the Midwest (close to 60 percent) that we document
below, few families would have remained in the same local area from the time of initial settlement in
any case. A more promising mechanism to explain the persistence of local values relies on a central
role for the church. The migrant church was historically the focal point for economic activity in
the new community, providing a domain in which information could flow and commitments could
be enforced (Gjerde 1991). While all churches would have instilled denomination-specific beliefs
2 Alesina, Baqir, and Easterly (1999), for example, document the negative relationship between racial fractionalization and public good provision in the United States.

48
in their congregations, migrant churches in highly fractionalized labor markets should, in addition,
have put special effort into instilling a loyalty to the local community among their members. It is
the long-term effect of these efforts, within particular denominations historically dominated by the
migrants, that we attempt to identify in this paper.
Labor market networks and civic institutions organized around European origin-countries are
largely irrelevant in the Midwest today (Gans 1979, Alba 1990). However, it would be incorrect
to assume that these institutions disappeared without a trace. The church, in particular, remains
at the center of the typical Midwestern community, providing important forms of support for its
members such as social activity and assistance when they were sick or infirm (Elder and Conger
2000). Well-functioning churches require loyal and committed members. If members’ inputs in the
church are complements, as suggested by Iannaccone (1998), and if there is sufficient overlap across
generations, then the first-generation members of well-functioning migrant churches, with a strong
local loyalty, would have had a greater incentive to instill this value in the generation that followed.
This process could have repeated itself from one generation to the next, in the church and the
broader community that formed around it, linking initial settlement patterns to a loyalty to place
today.
Loyalty to place or local identity (Hunter 1975, Guerson, Stueve and Fischer 1977, Hummon
1990) is well-described by the following quote from a resident of Bloomington, Indiana who moved
to that city to live with his wife who was born nearby, “... my chief ambition, I discovered during
our early years in Bloomington, was not to make a good career but to make a good life. And such a
life as I came to understand it, meant being a husband and a father first, and an employee second;
it meant belonging to a place rather than to a profession ... So as I came to recognize my children’s
need and my own need for a firm home place, I came to understand my community’s need for citizens
who stay put. Most of what I valued in Bloomington was the result of efforts by people who loved
the place, either because they grew up here and chose to stay, or because they landed here and chose
to remain” (Sanders 2007: 67-68).
Two aspects of local identity emerge from the preceding quote. First, local identity is associated with values that stress loyalty to a specific place. As in Akerlof and Kranton (2000), we
expect that individuals born and socialized in communities with a strong local identity will incur
a particularly high utility cost when they leave their communities. In the U.S. labor market, this
implies that individuals endowed with a strong sense of local identity will be less likely to choose
professional occupations, which are associated with significantly greater career-related moves than
non-professional occupations. Second, we see that the core value associated with local identity, no-

49
tably a commitment to stay, must be shared by other members of the community. Strong identity
and tight-knit communities are mutually reinforcing and, as discussed above, one cannot persist
without the other (Bénabou and Tirole [2006] make the same argument to explain the persistence
of mutually reinforcing collective beliefs and political equilibria).
The homogenization of production that is characteristic of the Midwest only began around 1880.
Thus, between 1850 and 1880 there was a window of time during which this region was rapidly settled
and during which there was substantial fluctuation in labor market conditions and, hence, ethnic
rivalry across local areas. These initial conditions are crucial because once community institutions
and a level of individual commitment were established, these mutually reinforcing characteristics
could have remained in place over many generations in the future. We consequently exploit variation
in initial conditions within the Midwest to identify the effect of historical networks, and the identity
and parochial institutions they engendered, on career decisions and geographical mobility long after
the networks themselves had disappeared.
The framework we have outlined generates the prediction that counties with greater ethnic rivalry
at the time of settlement should be associated with stronger local identity and better functioning
churches in select denominations historically denominated by the migrants. This implies that individuals born in those counties and belonging to those migrant denominations should be less likely to
enter a more mobile professional career today. Because occupational choice and religious affiliation
are jointly determined, we cannot estimate the effect of historical fractionalization on current career
decisions separately for members of migrant and non-migrant denominations. What we do instead
is to test the implied hypotheses that (i) higher fractionalization should be associated with a lower
supply of professionals and lower mobility on average in the county, and (ii) that higher fractionalization should be associated with greater church participation in select migrant denominations
alone.
To test the first hypothesis we combine data from the U.S. census and the National Longitudinal
Survey of Youth 1979 (NLSY79). A major advantage of situating the analysis in the Midwest is
that individual level data on occupations and ethnicity can be obtained from the census going back
to the time when this region was starting to develop. We focus on the 1860 census when rapid
growth was just commencing and construct a measure of ethnic fractionalization, one minus the
Herfindahl concentration index, within each broad occupational category in each county in that
census year. Averaging over all occupational categories in each county we arrive at a measure of
ethnic fractionalization in 1860 that is positively correlated with ethnic rivalry in the labor market
at that time and, hence, with the strength of local identity in the county as it subsequently emerged.

50
Matching the county-level measure of ethnic fractionalization with individual data from the NLSY79,
we find that individuals born in high fractionalization counties are significantly less likely to be
employed in professional occupations and are significantly less likely to have migrated out of their
birth county in 2000.
The values that we believe are responsible for the career choices described above could not
have persisted in a local community over so many generations without institutional support. The
second testable hypothesis is that high fractionalization counties should have been associated with
better functioning migrant churches and related parochial institutions. If these institutional features
persisted over time, supporting and being supported by local identity, then we would expect to
see greater participation in religious denominations that were historically dominated by migrants,
notably the Lutherans and the Catholics, today. Using data from the Census of Religious Bodies,
available at roughly ten-year intervals from 1860 to 2000, we successfully verify this important
prediction. The share of Lutherans and Catholics in the population is significantly larger in high
fractionalization counties by 1870 and, most importantly, this gap grows steadily wider over the
course of the twentieth century. In contrast, participation in other denominations is slightly lower in
the high fractionalization counties over the entire period. We verify that these cross-denominational
patterns in religious participation hold up with the NLSY data as well.
Our interpretation of the results described above is that local institutions and individual values,
determined endogenously by the labor market when the Midwest was first being settled in 1860,
have subsequently persisted over time and continue to shape career decisions one hundred and
forty years later. Alternative explanations assume that the ability endowment in the population,
investments in human capital, or access to particular types of jobs today vary with fractionalization in
1860. For example, although ethnic networks and specific European ancestry are no longer directly
relevant in the Midwest economy, ethnic fractionalization in 1860 could potentially be correlated
with particular features of the economy at that time, which had persistent effects and determine the
demand for professional labor today. Similarly, the well documented negative correlation between
fractionalization and public good provision could have given rise to poorly functioning schools in
high fractionalization counties.
We later show that if identity is salient and the demand for professional labor is uncorrelated with
historical fractionalization, then high fractionalization counties will supply too few professionals and
too many non-professionals as long as there is ex ante uncertainty in the demand for different types of
jobs. Professional labor must move into these counties and non-professional labor must move out of
them ex post in competitive equilibrium. It follows that individuals residing in high fractionalization

51
counties will be just as likely to hold professional jobs as individuals residing in low fractionalization
counties, once the labor market clears. It is only individuals born in high fractionalization counties
who should be less likely to be professionals. The alternative hypothesis, based on differences in the
demand for professional labor across counties, predicts that individuals born and residing in high
fractionalization counties should be less likely to be professionals. We verify that adult workers
residing in high fractionalization counties are indeed as likely to be professionals as those residing in
low fractionalization counties, ruling out the most obvious alternative explanation for our results.
Our results could also be obtained if the incoming migrants in the high fractionalization counties
had lower ability, and this ability differential persisted across generations. Human capital transmission is less of a concern with our analysis since we are going back at least five generations; even
if the ability distribution of the incoming migrants varied systematically with fractionalization in
1860, this heterogeneity would have long since disappeared, given the high rates of inter-county
migration in the Midwest.3 A more relevant concern is that fractionalization at the time of initial
settlement could be correlated with social heterogeneity many decades later in the twentieth century
when the school system was being established in the Midwest. Goldin and Katz (1999) document a
negative relationship between social heterogeneity and early investments in higher education. At the
same time, they show that greater church participation was associated with larger investments in
human capital, and so the net effect of 1860 fractionalization on education expenditures is ambiguous. The NLSY provides information on AFQT scores and educational attainment. Information on
local public good provision, including education expenditures, can also be obtained at the county
level in 1990. Reassuringly, we find no relationship between historical fractionalization and current
investments in schooling or educational attainment. The only exception is that individuals born in
high fractionalization counties are significantly more likely to choose college majors that lead to less
mobile careers.
This paper makes a contribution to the rapidly expanding literature on culture and religion in
economics as well as a more established literature on institutions. While the literature on religion
has tended to focus on beliefs or ideologies and their effect on individual behavior and aggregate
outcomes (Barro and McCleary 2003, Bénabou and Tirole 2006), we model the church as a collective
3 The empirical strategy that we employ is related to the strategy adopted by Fernandez and Fogli (2007) in a very
interesting recent paper. Fernandez and Fogli establish that female labor force participation in the origin country
affects labor force participation and fertility rates among second-generation female migrants in the United States.
They rely on the assumption that cultural traits will be transmitted across space and so cultural heterogeneity will
continue to manifest itself in an economic environment – the U.S. labor market – that does not distinguish between
social groups. The obvious alternative explanation, which they take care to rule out, is that human capital rather than
culture is being transmitted across generations. With our strategy, the chief concern is that there is something about
the place, other than identity and its complementary institutions, that has persisted over time, which the preceding
test on the county of residence, rather than the county of birth, helps rule out.

52
institution that was fortuitously well-positioned to reduce coordination and commitment problems
historically and which continues to play that role today. Instead of focussing on differences in
occupational choice and mobility across denominations or between church participants and nonparticipants, we study how local economic conditions shaped the orientation of individual churches
within particular migrant denominations. Our analysis thus provides empirical support for Greif’s
(2006) view that institutions cannot be simply described by a set of formal rules, but are in fact
more complex arrangements whose formation and evolution are determined by history and context.
Finally, this paper complements recent research that investigate the relationship between culture
and growth (Barro and McCleary 2003, Tabellini 2005, Fernandez and Fogli 2007). Although local
identity cannot be observed directly, our analysis goes beyond previous studies to theoretically and
empirically describe the formation and the persistence of this particular cultural trait. This trait
is seen to have a large effect on important economic decisions in a region as homogeneous as the
American Midwest, long after the forces that gave rise to it had ceased to be salient, providing
evidence for a role for culture that is not subject to the usual critiques of cross-country or crossregional analyses.

3.2

The Institutional Setting

3.2.1

The Settling of the Midwest

The Midwest first began to be settled in the early nineteenth century with the expansion of the
national canal system. The Erie Canal linking the Hudson to Lake Erie was completed in 1825
and numerous inter-regional and intra-regional canals were built over the next two decades (Fishlow
2000). However, it was only with the arrival of the railroad that the Midwest took off on a steeper
growth trajectory. Before 1850 the Midwest had less than one thousand miles of track, but almost
ten thousand were added by 1857 (Meyer 1989).
Improved rail transportation stimulated industrialization and the Midwest’s share of national
manufacturing increased rapidly between 1860 and 1920, with almost half of this increase occurring
in the 1860’s (Meyer 1989). This increase in economic activity led, in turn, to an increase in the
demand for labor. In 1810, approximately 6 percent of the labor force (outside the southern states)
resided in the Midwest. By 1860, this share had increased to 41 percent, with a further increase to
51 percent by 1880, after which regional growth converged to the national average (Margo 1999).
In this paper, the Midwest is comprised of the states of Illinois, Indiana, Iowa, Michigan, Min-

53
nesota, Ohio, and Wisconsin (Missouri, the only pre-Civil War slave state in the Midwest, is excluded
from the analysis). Using county-level census data we see that the number of incorporated counties
increases sharply from 1850 to 1860 and then flattens out by 1880 in Figure 1. Information on railroads, obtained from the Historical Map Archive at the University of Alabama, indicates that the
number of these counties with a railroad also increases steeply over the 1850-1870 period, growing
thereafter at a slower rate.4 The rapid expansion of the railroad system and the economic activity
that accompanied it led to a steep increase in the population of the Midwest as well as an influx of
foreign migrants. Using county-level census statistics, the total population in our seven Midwestern
states grew from less than 5 million in 1850 to 20 million in 1900. The number of foreign-born
migrants nearly tripled between 1850 and 1860, reaching close to 20 percent of the population.
Where did these migrants come from? Individual-level data, including characteristics such as
age, sex, occupation, and country of birth, are publicly available from the Population Census each
decade from 1850 to 1930. We use the 1-in-100 sample from the 1860-1900 IPUMS to study changes
in the migrant population in these critical early decades in Table 1. The English (13 percent), the
Irish (25 percent), and the Germans (32 percent), dominated the migrant population in the Midwest
in 1860, just after the first wave of migration described above, with no other ethnic group accounting
for more than a 3 percent share of the migrants in that year. Subsequently, the English and the
Irish were displaced by the Germans and the Scandinavians over the 1860-1900 period. Notice that
the Italians, Poles, and Slavs continue to be insignificant in 1900, although they would display a
substantial presence in Midwestern cities such as Cleveland, Chicago, and Pittsburgh by the first
quarter of the twentieth century.
What jobs did the migrants occupy? Table 2 reports the occupational distribution of the migrants
from the IPUMS sample in 1860, 1880, and 1900. Although agriculture was the dominant sector
in this period, the share of farm employment declines from 62 percent in 1860 to 48 percent in
1900, with manufacturing operatives and laborers accounting for much of the increase in non-farm
employment. These trends are consistent with the growth of the manufacturing sector described
above and they are similar for the foreign-born migrants and the native workers (not reported).
The apparent similarity in the occupational distribution for migrants and natives masks differences in the type of jobs they had access to within each occupational category. Labor markets in
the nineteenth century could be divided into three segments: a stable segment with permanent employment, an unstable segment with periodic short-term unemployment, and a marginal but highly
4 Railroad maps were used to construct a county-level binary variable indicating whether any part of a railroad ran
through the county in a given year. Railroad maps were unavailable in some census years in which case we used maps
that were closest in vintage to those census years (the discrepancy never exceeded three years).

54
flexible segment characterized by spells of long-term and short-term employment (Gordon, Edwards,
Reich 1982). Migrants being newcomers to the U.S. market typically ended up in the unstable and
marginal segments, where the uncertain labor demand naturally provided an impetus for the formation of ethnic networks that helped their members find jobs (Hoerder 1991). Based on the discussion
above and Figure 1, we would expect these networks to have been particularly important during the
rapid expansion phase around 1860, with the influx of migrants and the opening up of new labor
markets in both agriculture and manufacturing. The initial conditions in the empirical analysis,
measuring ethnic rivalry in the labor market at the time of settlement, will consequently be measured using data at this point in time. Although not reported, regressions that treat 1850 as the
starting point provide qualitatively similar results.

3.2.2

Ethnic Labor Networks in the Midwest

Accounts by contemporary observers and a rich social history literature indicate that friends and kin
from the origin community in Europe played an important role in securing jobs for migrants in the
Midwest in the nineteenth century and the first quarter of the twentieth century. Early historical
studies used census data, which provides fairly detailed occupational and ethnic information, to
identify ethnic clusters in particular locations and occupations. Gordon, Edwards, and Reich (1982)
note that although foreign-born workers comprised just over 20 percent of the labor force in 1870,
they accounted for 43 percent of the iron and steel operatives, 43 percent of the woolen mill workers,
and 63 percent of the miners. Nearly a quarter of railroad workers were Irish, a third of the miners
were British, and about half the workers in the baking and confectionary business were German.
While such clustering suggests that underlying ethnic networks were channeling their members into
particular occupations, it could simply reflect the fact that migrants arrived with specific skills.
Hutchinson’s (1956) analysis of 1950 census data, however, indicates that clustering continued even
among the migrants’ children, with the concentration of particular ethnic groups in some industries
actually increasing from the first to the second generation.
Although census data are a useful source of information, they do not provide details of the
migration process and its connection to ethnic networks in the United States. Over the past four
decades, however, social historians have linked parish registers and county data in specific European
sending communities to census and church records in the United States to construct the entire chain
of migration from those communities as it unfolded over time. This remarkable research effort has
documented the formation of new settlements in the Midwest by pioneering migrants, the subsequent

55
channelling of migrants from the origin community in Europe to these settlements, as well as the
movement of groups from the original settlement to new satellite colonies elsewhere in the United
States. As with the census data, this research identifies occupational and geographic clustering, but
at a dissagregate level. Over 45 percent of the Swedish emigrants from the parish of Rätvik eventually
settled in Isanti County, Minnesota (Ostergren 1976). Two-thirds of the emigrants from Balestrand
located in Norway Grove, Wisconsin in the first decade of migration from that Norwegian community
(Gjerde 1985). And one-third to one-half of the German emigrants from Westerkappeln settled in
Duden County, Missouri (Kamphoefner 1987). Although less detailed origin-country information
is available for southern European migrants, similar ethnic clustering in particular neighborhoods
of Midwestern cities has been documented for Polish, Italian, and Slovak immigrants from specific
sending regions (Alexander 1991, Bodnar, Simon, and Weber 1982).
A possibly stronger indicator of the importance of migrant networks is the maintenance of ethnic
ties over successive moves within the United States. Italians moved from Southern Illinois to the
“Italian Hill” in St. Louis when coal mining operations were reduced in the 1920’s and Slavs
moved from mines in Western Pennsylvania to Detroit’s growing automobile industry in the same
decade (Bodnar 1985). Norwegians from Balestrand initially settled in Norway Grove, Wisconsin,
but over time they established six satellite settlements in Wisconsin, Minnesota, Iowa, and Illinois
(Gjerde 1985). A similar pattern has been documented for Norwegians immigrants from Fortun,
who initially settled in Vernon and Crawford Counties, Missouri, but later established satellite
communities throughout the Midwest (Gjerde 1997).
While the preceding descriptions of ethnic clustering are informative, ethnic concentration within
specific departments or firms in a local industry provides possibly the strongest evidence that labor
networks were active. Nearly all three thousand employees of the Peninsular Gas Company in
Detroit in 1900 were Polish, and Croatians held only three jobs in Indiana’s oil refineries: stillman
helper, fireman, and still cleaner (Bodnar 1985). Italians in Pittsburgh’s steel industry dominated the
carpentry, repair, and rail shops. And, relying on friends and relatives, Poles established occupational
niches at the Jones and Laughlin and Oliver Mills on Pittsburgh’s Southside, Heppenstalls and the
Pennsylvania Railroad in Lawrenceville, and at the Armstrong Cork Company and the H.J. Heinz
Plant. As John K. a Polish immigrant put it, “The only way you got a job [was] through somebody
at work who got you in” (Bodnar, Simon, and Weber 1982: 56).
What kept ethnic networks in place so far from their origin locations? It has been argued that
“[migrants] from varying regions [in the origin country] formed a community based on common
nationality and religion centered on the central cultural institution – the church” (Gjerde 1991:

56
176). The building of a church was one of the first organized actions in the migrant community
once it arrived in an area (Barton 1975, Bodnar, Simon, and Weber 1982). Churches provided
both economic and social support to their members. Information about jobs and potential land
transactions flowed within the congregation and the church also served as a public arena in which
members who had reneged on their obligations could be sanctioned.
Given the variety of economic opportunities in the United States, individuals and small groups
drawn from the same parish in Europe often had an incentive to move and seek employment elsewhere. The stability of the local community in the United States, based on a common national
origin rather than narrower social affiliations, was thus essential for the viability of the labor market
network. One strategy to maintain stability would have been to instill a sense of loyalty to this community. The discussion that follows will describe how efforts to engender a sense of local identity by
the church and the institutions that formed around it might have varied across Midwestern counties
when the first wave of migrants arrived in the region.

3.2.3

Ethnic Fractionalization Across the Midwest

“You take in the erection department – it was mostly all Slavs ... Not Slovaks, it was Polish ... We
didn’t have Lithuanians there and the Russians were not involved there ... Now if a Russian got his
job in a shear department ... he’s looking for a buddy, a Russian buddy. He’s not going to look for
a Croatian buddy. And if he see the boss looking for a man he says, ‘Look, I have a good man,’
and he’s picking out his friends.” (Polish immigrant in Pittsburgh, quoted in Bodnar, Simon, and
Weber 1982:62).
Numerous historians have described the efforts made by ethnic groups to establish a “toe-hold”
(Thistlethwaite 1991) or a “beachhead” (Bodnar, Simon, and Weber 1982) in particular industries
or establishments when they first settled in an area. The preceding quote suggests, in addition,
that once a network had established a presence in the labor market, it was essential to maintain
that presence. The discussion that follows will describe the labor market conditions under which
migrant communities would have had the greatest incentive to restrict exit and, hence, instill a
sense of local identity among their members. An explanation for the persistence of this identity over
multiple generations, based on the complementary role of the church and other related institutions,
is postponed to the next section.
Consider a market with N migrant workers drawn from M communities. The number of workers
in each community is exogenously determined and we do not require the labor market to be in

57
equilibrium. The workers are competing for a fixed number of coveted permanent jobs, which
provide a total surplus R. Tullock’s (1980) canonical model of rent seeking can be conveniently
adapted to this setting to describe the share of total surplus captured by community i

Si =

nα
i +

nα
i
PM −1
j=1

nα
j

where ni is the number of workers belonging to community i, nj is the corresponding number of
workers from each community j 6= i, and α > 0.5 The cost to community i from the exit of of its
members at the margin can be described by the expression
PM −1 α
αnα−1
dSi
i
j=1 nj
R
= R³
´2 .
P
dni
M
−1 α
nα
i +
j=1 nj
Assuming that all networks are of equal size, ni = nj = N/M ,
dSi
Rα
R
=
dni
N

µ
¶
1
1−
,
M

which is increasing in M . Holding N and R constant, an increase in the number of communities
is associated with a greater marginal cost of exit regardless of whether the returns to network size
are increasing (α > 1) or decreasing (α < 1) at the margin. Self-interested individuals do not
internalize the cost imposed on the rest of the network by their exit. It follows that communities
in labor markets with greater ethnic competition should have invested more in instilling a sense of
local identity as a way of aligning individual incentives more closely with the community’s welfare.
Because networks vary in size in practice, we use a standard measure of fractionalization, defined
as one minus the Herfindahl index of ethnic concentration, to measure competition.6 Following the
discussion above, we assume that networks were most active in the migrant population and that they
were organized around the country of origin. These ethnic networks would have competed directly
with each other within (broad) occupational categories. The IPUMS provides the country of birth
and the occupation (where relevant) for each sampled individual. Using these data, we compute
ethnic fractionalization within the occupational categories specified in Table 2 in each Midwest
county in 1860. The weighted average of the occupation-specific statistics, where the weight is
5 Tullock’s specification is identical to the equation above except that the number of workers is replaced by the
investment in rent seeking.
6 The Herfindahl index of ethnic concentration is computed as the sum of the squared share of each ethnicity in the
occupational category. It is easy to verify that for the special case with networks of equal size, the fractionalization
measure is equal to 1 − N 2 /M , which is increasing in the number of communities M , for a given number of migrant
workers N , as above.

58
measured by the share of migrants in the occupation, then provides us with an overall measure of
ethnic fractionalization in the county. For example, suppose that there are two occupations and two
ethnicities in the county, with complete occupational segregation along ethnic lines. Our measure
of fractionalization will be zero in this case, correctly reflecting the absence of ethnic competition
in this labor market. In contrast, if the two ethnicities are of equal size and evenly distributed
across the two occupations, our measure of fractionalization will increase to 0.5. Figure 2 plots the
fractionalization measure, which has a mean of 0.5 and a standard deviation of 0.2, across the seven
Midwest states. Counties that were not incorporated and those without foreign-born migrants in
1860 are unshaded in the Figure. Notice that there is substantial variation in the fractionalization
measure, which is useful for the statistical tests that follow.
While we focus on the effect of ethnic fractionalization in strengthening within-group solidarity,
fractionalization could at the same time undermine the functioning of secular institutions. An
important advantage of using the Midwest for the analysis is that this region was settled recently
enough that accurate measures of ethnic rivalry can be constructed with census data. At the
same time, public services were provided at a rudimentary level in 1860 and the first investments
in secondary schooling, for example, occurred more than 50 years later (Goldin and Katz 1999).
While initial conditions may have determined subsequent labor in-flows and social heterogeneity,
it is possible that the relevant determinants of public good provision many decades later in the
twentieth century were largely unrelated to fractionalization in 1860. We will later verify that
1860 fractionalization is indeed uncorrelated with local public good provision, including education
expenditures, in 1990. The major long-term effect of 1860 fractionalization seems to have been
to shape the orientation of a particular set of parochial institutions that started to be established
around that time.

3.2.4

Historical Fractionalization and Current Economic Conditions

To test the hypothesis that local identity and its complementary parochial institutions shape career choices, historical fractionalization must be uncorrelated with the demand for professionals,
intrinsic ability in the population, and local public good provision today. We show in this section
that ethnic fractionalization in 1860 is uncorrelated with measures of economic activity, racial and
ethnic fractionalization, and expenditures on local public goods in 1990 once we control for a few
important characteristics of the 1860 economy. More stringent tests ruling out a link between 1860
fractionalization and current economic conditions are discussed in Section 5.

59
Labor market networks organized around the European-origin countries in Table 2 are no longer
active in the American Midwest. Nevertheless, fractionalization in 1860 could have been correlated
with particular features of the economy at the time of initial settlement that had persistent effects.
To explore this possibility, we proceed to understand what determined fractionalization in the first
place. In a rapidly expanding Midwest economy, some of the variation in fractionalization across
counties was no doubt a consequence of accidental initial settlement by ethnic groups in particular
locations, which fueled the arrival of more migrants as networks crystallized. At the same time,
fractionalization would have been determined by the demand for labor, with more ethnic groups
attracted to rapidly growing areas. We have already discussed the importance of transportation
links in the development of the Midwest and Table 3 consequently investigates the effect of railroads
and distance to canals and a Great Lakes harbor on fractionalization in 1860.7 Counties with a
railroad running through them and counties that are closer to a canal or harbor have significantly
higher fractionalization in Table 3, Column 1. Counties close to a harbor had a greater proportion
of the workforce engaged in manufacturing and a smaller proportion in agriculture in 1860 in Table
3, Columns 2-3. Improved transportation, more generally, is associated with a larger population in
1860 in Column 4.
Notice that the pattern of coefficients in Table 3, Column 1 with fractionalization as the dependent variable matches perfectly with the corresponding pattern in Column 4 with county population
as the dependent variable. Counties with superior transportation infrastructure, which were more
populated and presumably growing more rapidly, were more fractionalized in 1860. The population
of the county in 1860 could have determined subsequent agglomeration in economic activity, with
long-term implications for the growth of the local economy. Not surprisingly, we see in Table 4,
Columns 1-3 that while 1860 fractionalization has an insignificant effect on 1990 agriculture share
and manufacturing share, it is significantly and positively associated with current county population
as well as population density (not reported). More urban counties will be more racially diverse, and
it follows that 1860 fractionalization is significantly and positively correlated with racial fractionalization in 1990 in Column 5.8 In contrast, 1860 fractionalization is completely uncorrelated with
7 Data on the distance to the nearest canal (or navigable river) and the nearest Great Lakes harbor is obtained
from Jordan Rappaport’s website at the Kansas City Federal Reserve Bank. The distance is computed in each case
from the county centroid.
8 The manufacturing share in 1990 is defined as the share of the civilian labor force employed in manufacturing
in that year. The agriculture share in 1990 is computed using the farm population and the total population in the
county in that year. All these statistics, as well as the area of each county used to compute the population density,
are obtained from the 1994 County Data Book, compiled by the U.S. Bureau of the Census. Racial fractionalization
is computed from the 1990 IPUMS as one minus the Herfindahl index of racial concentration, using the same five
racial groups as in Alesina, Baqir, and Hoxby (2004). Ethnic fractionalization is computed from the 1990 IPUMS as
one minus the Herfindahl index of racial concentration, using the same 16 white ethnic groups as in Alesina, Baqir,
and Hoxby.

60
(white) ethnic fractionalization in Column 4, consistent with the idea that few individuals living in
the Midwest today could trace an unbroken line of descent to European ancestors arriving in the
same county at the time of initial settlement.
We expect the fractionalization effect to disappear once important features of the local economy
that would directly determine economic activity today, such as manufacturing share, agriculture
share, and particularly population in 1860, are accounted for. The fractionalization coefficient is
indeed small and insignificant in Columns 7-11 of Table 4, once the county controls are included.
In contrast, agricultural counties in 1860 continue to be agricultural in 1990 and more populated
counties in 1860 remain larger in 1990. Notice, however, that manufacturing in 1860 does not
have a persistent effect. Although the factory system began to replace artisan shops by 1820,
production continued to be largely organized in workshops managed by labor contractors who hired
their own employees until 1870 (Gordon, Edwards, and Reich 1982). The heavy manufacturing that
characterized the Midwest economy in the twentieth century, with its emphasis on the iron and steel
industry, only came at the turn of the century (Meyer 1989). Recently it has been argued that
the surge of foreign immigration in the second half of the nineteenth century provided the impetus
for the factory system and the subsequent industrialization of the Midwest (Kim 2007). Whatever
the explanation, it is clear that the pattern of manufacturing around 1860, spread throughout the
Midwest in small towns, had little connection with the heavy manufacturing, concentrated in large
cities, that followed in the twentieth century. This is presumably why 1860 characteristics have such
little power in predicting the share of manufacturing in the county in 1990.
For our purpose, what is important is that once we control for a few features of the nineteenth
century economy, fractionalization in 1860 has no effect on characteristics of the economy today,
such as the share of manufacturing and urbanization (measured by total population or by population
density), that are associated with the demand for professional labor. All of the results that follow will
consequently condition for these important features of the 1860 economy. We will verify that all the
results hold up without the 1860 controls as well. Finally, notice that more ethnically fractionalized
counties in 1860 have significantly lower religious fractionalization in 1990, with and without the
controls in Column 6 and Column 12.9 This intriguing result, which stands conspicuously apart
from the other regressions reported in Table 4 will be clarified in the discussion on the church and
the persistence of identity that follows in Section 3.
We complete this section by describing the relationship between local public good provision in
9 Religious fractionalization is computed from the 1990 Census of Religious Bodies as one minus the Herfindahl
index of religious concentration, using the same 18 religious denominations as in Alesina, Baqir, and Hoxby (2004).

61
1990 and ethnic fractionalization in 1860 in Table 5. Local government expenditures per capita in
1990 are regressed on ethnic fractionalization, manufacturing share, agriculture share, and county
population in 1860 in Column 1. Retaining the same set of regressors, the dependent variable is
the share of expenditure allocated to education, health, police, roads, and welfare in Columns 2-6.10
The expenditure shares for these five public goods add up to 0.79, with the education share being
as large as 0.6. The other shares range from 0.03 to 0.07. Given that the standard deviation of
the fractionalization variable is 0.2, it is evident from the point estimates that 1860 fractionalization has a negligible effect on public good provision today. Using the same shares, Alesina, Baqir,
and Easterly (1999) document a negative and significant relationship between contemporaneous
racial fractionalization and the allocation of resources to certain public goods, particularly education, roads, and welfare. We have already shown that 1860 ethnic fractionalization is uncorrelated
with 1990 racial fractionalization in Table 4, and not surprisingly we see in Table 5 that historical
ethnic fractionalization has no effect on current local public good provision, including education
expenditures.

3.3

The Church and the Persistence of Identity

In a recent paper, Akerlof and Kranton (2005) describe how firms and other economic organizations
can instill a sense of identity or loyalty among their workers to solve agency problems. In our
setting, the natural organization to instill loyalty in the community would have been the local church.
The church was among the first institutions to be established when immigrants arrived in an area
(Hoerder 1991, Barton 1975, Bodnar, Simon, and Weber 1982). The church congregation provided
many forms of mutual assistance including credit, insurance, job referrals, business information, and
social support (Gjerde 1985, 1997, Alexander 1991). Indeed, it has been argued that immigrants
participated in church communities to benefit from these economic and social services, instead of
being drawn to the church by a particular belief or ideology (Bodnar 1985).
Despite these material attractions, exit from the local church and labor market was always a
threat in nineteenth century America, especially in areas that had been recently settled. Migration
to the United States, and subsequent internal migration, was typically organized around small groups
drawn from the same parish or province in Europe. These groups were too small to maintain a viable
church congregation and labor network when they first arrived in an area and so migrant churches
in the Midwest typically brought together many regional groups with the same national origin.
10 These

data are obtained from the Annual Survey of Governments, 1990. Alesina, Baqir, and Easterly (1999) use
data drawn from the same source.

62
Inter-regional conflict was common, and there was always the possibility that groups within the
church would move again to take advantage of new opportunities that became available elsewhere
(Gjerde 1997). Under these circumstances, the migrant church had two distinct objectives: (i)
to instill denomination-specific beliefs in the congregation, and (ii) to instill a loyalty to the local
community that transcended narrower social affiliations. We saw earlier that the negative externality
associated with individual exit was greatest in ethnically fractionalized labor markets. This implies
that migrant churches should have made a special effort to instill a sense of local identity in such
markets. In contrast, we have no prediction for the effect of fractionalization on religious beliefs or
the orientation of churches belonging to non-migrant denominations.
If local identity played a role in reducing exit in the nineteenth century, how and why did it
persist long after the ethnic labor networks it supported ceased to be salient? Our explanation for
this persistence is based on the observation that churches continue to provide important forms of
social support to their members. Church activities include Sunday school service, youth groups,
pot-lucks, informal home parties, and food, visits, and other forms of support when members of the
congregation are ailing or infirm. The church also lies at the center of a cluster of inter-linked civic
institutions, including the school and various voluntary organizations. Life in a Midwest community
revolves around these institutions, which bring families, friends, and multiple generations together
on a regular basis (Elder and Conger 2000). We focus on the church in the analysis that follows
because of its central position and because church participation is available at the county level, by
denomination, from 1860 till 2000.11
There are complementarities associated with church inputs; if the rest of the congregation commits time and effort to the church, then the returns to the individual’s inputs would increase as well
(Iannaccone 1998). The presence of these complementarities introduces two problems. First, selfinterested individuals will devote a level of inputs that is sub-optimal because they do not internalize
the benefits that the rest of the congregation derives from their actions. Second, a coordination problem can arise since individuals will only make career choices that are associated with a high level of
commitment to the church if they expect other members of their generation, who are simultaneously
making these choices, to do likewise. A strong sense of local identity reduces each of these problems,
11 The number of civic and social associations in the county and the corresponding number of religious organizations
can be obtained from the County Business Patterns, 1990. The number of not-for-profit organizations in the county
can be obtained from National Center for Charitable Statistics Core Files, 1990. Controlling for manufacturing
share, agriculture share, and county population in 1860, ethnic fractionalization in that year has no effect on any
of these variables. Note, however, that these statistics are based on the number of organizations, whereas what we
require are measures of participation. For example, we will see below that 1860 fractionalization has a significant and
positive effect on religious participation in 1990 despite the fact that it is uncorrelated with the number of religious
organizations.

63
serving as a commitment device and increasing the level of church inputs in equilibrium. Given
that church inputs are complements and that generations overlap, it is easy to verify that the first
generation in a migrant church with a committed congregation would have been more willing to bear
the cost of instilling a strong sense of local identity in the generation that followed. Local identity
and church inputs could have subsequently reinforced each other in this way over many overlapping
generations.12
Local identity and church inputs cannot be observed directly. However, both the Census of Religious Bodies (CRB) and the NLSY provide information on the number of church participants. The
preceding discussion indicates that migrant churches in high fractionalization counties should have
been relatively well-functioning, with a high level of church inputs, at the time of initial settlement
and over many future generations. These churches would, in turn, have attracted more members
and supported higher rates of attendance in the congregation. Restricting attention to denominations historically dominated by the migrants, the testable prediction is that high fractionalization
counties should be associated with greater church participation overall. Moreover, the effect of fractionalization on church participation must be sustained over time, reflecting the underlying role of
the church in maintaining local identity. This prediction distinguishes our model, with identity as
a commitment device, from the model of religious cults proposed by Iannaccone (1992) in which
self-sacrifice and strict norms of behavior are used to screen out free-riders and ensure higher levels
of participation among those that remain. While this alternative mechanism may induce high levels
of participation at the intensive margin, it results in small congregations and is typically short-lived.
Based on the country of origin of the incoming migrants, reported in Table 1, most of the migrant churches would have been Lutheran or Catholic. Regressing the population-share of different
denominations, computed with data from the CRB in 1860, on the share of migrants in that year,
counties with a greater share of migrants are indeed disproportionately Lutheran and Catholic.13
A strong test of our framework is that the predictions derived above should apply to those migrant
12 The persistence in local identity that we describe is related to a paper by Bénabou and Tirole (2006) linking
investments in human capital at the household level with the political equilibrium at the macro level. In their model,
children choose a level of effort (schooling) based on their belief about the returns to this effort. If a sufficiently large
number of households invest in effort, they will form a pivotal voting block and set a low tax rate, generating, in
turn, a high return to effort that reinforces the initial beliefs. Thus, two political equilibria can arise; a low-effort
equilibrium with substantial redistribution (high taxes) and a high-effort equilibrium with little redistribution (low
taxes). Bénabou and Tirole fix the equilibrium by assuming that children have imperfect willpower and so will
underinvest in effort if left to themselves. Their parents, who provide them with information about the returns to
effort, will consequently systematically inflate these returns. This “ideological” position leads parents in the higheffort equilibrium to (optimally) ignore negative signals about the returns to effort, allowing particular equilibria and
the collective beliefs that support them to persist over many generations. In this paper we use local identity rather
than ideology to fix the level of church inputs and sustain a social equilibrium over many generations.
13 Apart from the Germans and the Irish, the English were also an important migrant group in 1860. The English
would have been disproportionately Anglican (Episcopalian). This denomination accounts for just 3 percent of church
participants in 1860 and is never a significant force in the Midwest. Inclusion of the Episcopalians, not surprisingly,
has little effect on the religious participation results reported below.

64
denominations only. In particular, higher fractionalization counties should be associated with a
greater share of participating Lutherans and Catholics in the population around the time of initial settlement and this effect of fractionalization should be sustained over time. In contrast, the
model has no prediction for the relationship between fractionalization and participation in other
denominations, providing us with a useful falsification test.
The relationship between ethnic fractionalization and religious participation can be tested with
data from the CRB, which has been conducted at roughly ten-year intervals from 1860 to 2000. This
census was conducted as part of the population census from 1860 to 1890, with census enumerators
collecting information from individual churches in each county. Subsequently, the U.S. Bureau of
the Census conducted the CRB separately from the population census in 1906, 1916, 1926, and
1936. Starting from 1952, the National Council of Churches of Christ undertook the responsibility
of conducting the CRB, with subsequent census rounds in 1972, 1980, 1990 and 2000.
The 1860-1890 census rounds collected information on the number of church seats by denomination in each county. From 1890 onwards, information was collected on the number of members
directly, and from 1972 onwards the number of adherents was collected as well. Despite these changes
in the management of the CRB and the measure of religious participation, we uncover clear changes
in the mix of denominations as well as the effect of 1860 fractionalization on religious participation
over time.
Table 6 reports changes in the mix of denominations and overall participation rates in our
Midwestern counties over the 1860-2000 period. To take account of the fact that the measure of
participation was changing over time, we report statistics in the first and the last census-year that
each measure was used. Thus, participation is measured by the number of church seats from 1860
to 1890, by the number of members from 1890 to 1952, and by the number of adherents from
1972 to 2000.14 The participation rate is then computed as the number of participants divided by
the contemporaneous population in the county. Five denominations – Baptist, Catholic, Lutheran,
Methodist, and Presbyterian – account for roughly 80 percent of church participants over the 18602000 period. Among these denominations, the Lutherans grow rapidly in popularity over the 18601890 period and the 1890-1952 period, remaining stable thereafter. In contrast, the Methodists and
the Presbyterians decline steadily over time. There is no clear trend among the Baptists and the
Catholics.
The inability of the Baptist church to increase its share of church participants contrasts with the
14 Although the number of members was also collected in the 1972-2000 census rounds, this statistic is not available
for Catholics, a major denomination in our Midwestern counties, in these rounds.

65
surge in popularity of this denomination elsewhere in the United States, as documented by Finke
and Stark (1992). The Midwest stands apart from the rest of the country in that the “traditional”
denominations, particularly the Catholics and the Lutherans, continue to dominate, accounting for
over half of all church participants by 2000. Even without the Baptist surge, religious participation increased steadily over time in the Midwest, rising over the 1860-1890 and 1890-1952 periods.
Based on the statistics in the most recent 1972-2000 period, 55-60 percent of the population in our
Midwestern counties are church adherents.15
We next proceed to estimate the relationship between ethnic fractionalization in 1860 and religious participation in the county in each round of the CRB. The religious participation regression
includes the same set of county-level controls as the regressions in Table 4, Columns 7-12 and Table
5. Because the regression is estimated over many census years, we simply report the 1860 fractionalization coefficient, together with the 95 percent confidence band, in each census year in Figure 3.
This coefficient is less precisely estimated in the early census years, but grows steadily larger, while
remaining statistically significant, all the way through to 2000.16
A stronger test of our framework is that its predictions should only apply to participation in
migrant denominations. We consequently proceed to estimate two separate regressions in each census
year; the first regression has the share of Lutherans and Catholics in the population as the dependent
variable and the second regression has the share of all other denominations as the dependent variable.
The 1860 fractionalization coefficient, with the corresponding 95 percent confidence band, is reported
for each regression in each census year in Figure 4. We now see that the share of Lutherans and
Catholics in the population is significantly greater in high fractionalization counties by 1870 and
that the fractionalization effect gets steadily larger over time. Although we do not report results
separately by denomination, this pattern is obtained for both the Catholics and the Lutherans.
A one standard deviation decline in 1860 fractionalization would increase the population share of
Lutherans and Catholics in the county by four percentage points (22 percent) in 2000. In contrast,
the fractionalization coefficient is much smaller and negative, but remains significant and stable over
15 Drawing on a number of different sources, Iannaccone (1998) reports church participation rates of 70 percent or
even higher in the United States. Our statistic may be lower because it is based on a census of churches rather than
the population, or because we divide by the entire population in the county when computing participation rates.
16 The 1860 fractionalization coefficient with standard error in parentheses is -0.66(0.39). This outlying coefficient
is omitted from Figure 3 and Figure 4 that follows to clarify changes in the fractionalization coefficient over time.
As noted, church seats and the number of members are both available in the 1890 CRB. We use the first statistic to
measure church participation in Figure 3 and Figure 4 because it is more in line with trends in the fractionalization
coefficient over time. Although not reported, the coefficient on 1860 agriculture share is also positive and significant in
the religious participation regressions and grows larger over time. Unlike ethnic fractionalization, which soon ceased
to be directly relevant, recall from Table 4 that agricultural counties in 1860 remained disproportionately agricultural
in 1990. It is well known that farming communities tend to be more religious and community-oriented and so the
persistent effect of the 1860 agriculture share may simply reflect the persistence of agricultural activity in particular
areas over time.

66
time in the companion regression. These results, taken together, provide a simple explanation for
the negative correlation between ethnic fractionalization in 1860 and religious fractionalization in
1990 that we reported in Table 4.17
Our explanation for the pattern in Figure 4 is that historical economic conditions gave rise to
particularly well-functioning Lutheran and Catholic churches in the high fractionalization counties.
An alternative explanation for this pattern is that there are increasing returns to scale in church
performance; Lutheran and Catholic churches were large to begin with in the high fractionalization
counties, simply because there were more migrants from those denominations in the area, and
subsequently grew larger over time. To rule out this explanation we experimented with augmented
specifications that included (i) the share of Lutherans and Catholics, and (ii) a full set of 32 ethnic
shares (which sum up to the share of migrants in the population), in 1860 by county, as regressors.
The initial share of Lutherans and Catholics, as well as some of the ethnic shares do have persistent
effects. However, the fractionalization coefficient remains stable over the course of the twentieth
century with both the augmented specifications.18
Initial conditions in 1860 would have determined subsequent in-migration by diverse ethnic groups
and the formation of new Catholic and Lutheran churches in the decades that followed. The analysis
in this paper is restricted to estimating the (overall) effect of initial conditions in 1860 on later
outcomes and Figure 4 thus incorporates additions to the stock of churches and to the migrant
population over time.19 The increase in the share of participating Lutherans and Catholics that we
observe could in that case simply reflect growth at this extensive margin rather than a widening of the
gap in church attendance (the intensive margin) between low and high fractionalization counties.
The mechanism that we have proposed to explain the persistence of local identity requires that
differences in church participation at the intensive margin be maintained across counties. The CRB
data do not allow us to separate church participation along the intensive and the extensive margin.
However, the NLSY collects information on both religious affiliation (inherited at birth) as well as
church attendance, which can be compared to shed light on these different marginal effects.
17 The pattern of coefficients reported in Figure 4 would also be obtained if the county controls were omitted
from the participation regression, although the point estimates would be somewhat smaller. We are using three
different measures of church participation over the 1860 period in Figure 4. One concern would be that the change
in participation over time is mechanically driven by change in these measures. Notice, however, that participation
increases steadily over time and not in three distinct jumps.
18 In a related test, we also included religious fractionalization in 1860 as a regressor to allow for the possibility that
competition between churches could have affected participation rates (Finke and Stark 1992, Gruber 2005). We find
that religious fractionalization in 1860 has no effect on future church participation, leaving the estimated coefficients
in Figure 4 unchanged.
19 The advantage of focussing on the initial conditions is that all networks had the same vintage at that point in
time and so ethnic rivalry can be measured relatively easily. To estimate the effect of social heterogeneity in later
decades we would need to consider the entire history of in-migration and account for differences in vintage and, hence,
market power across ethnic networks.

67
The NLSY consists of a nationally representative sample of American high school seniors in 1979
who were subsequently interviewed annually from 1979 to 1994 and biennially thereafter. The NLSY
collects information on the respondents’ county of birth, the religious denomination that they were
raised in, and church attendance in the 1979 and 2000 rounds. To be consistent with the results using
the CRB, we aggregate the individual data to the county level. Table 7, Columns 1-2 regress religious
affiliation, separately for the migrant and non-migrant denominations, on 1860 fractionalization
and the county-controls. The dependent variable in these regressions is measured as the share
of respondents born in each county who were raised in the migrant denominations and the nonmigrant denominations, respectively. Table 7, Columns 3-4 regress church participation, separately
for the migrant and non-migrant denominations in 1979 and 2000, on 1860 fractionalization and the
county-controls. The dependent variable is now measured as the share of the respondents born in
each county who were raised in the relevant denominations and report attending church.
Individuals born in high fractionalization counties are more likely to be raised in migrant denominations and less likely to be raised in non-migrant denominations in Columns 1-2, although the
fractionalization coefficient is not significant at conventional levels. This result indicates that some
of the variation in participation with fractionalization in Figure 4 can be explained by growth in the
migrant denominations at the extensive margin. Restricting attention to individuals who actually
attend church, we see in Columns 3-6 that the fractionalization coefficient retains the same sign,
but is now very precisely estimated for the migrant denominations. Notice that the fractionalization
coefficient in Column 5 (the 2000 round) is very similar to the corresponding coefficient in Figure 4.
Church attendance and religious affiliation are jointly determined and so we cannot estimate the
relationship between attendance (the intensity of participation) and fractionalization separately for
individuals raised in the migrant denominations. However, the results in Table 7 can be used to
test (and reject) the weaker hypothesis that attendance is uncorrelated with fractionalization in the
migrant denominations. Suppose that a fraction θi of individuals raised in those denominations in
county i attend church. The value of the dependent variable in the participation regression, Column
3 and Column 5, will then be θi multiplied by the dependent variable in the religious affiliation
regression in Column 1. It follows that the fractionalization coefficient in the participation regression
will be simply E(θi ) < 1 multiplied by the corresponding coefficient in the religious affiliation
regression if θi is uncorrelated with fractionalization.20 What we see instead is that the coefficient
20 Let the share of the population affiliated with the Lutheran and Catholic church in county i be a . Let X denote
i
i
the level of 1860 fractionalization in that county. Ignoring the additional controls and the constant term, we are
effectively estimating the following regressions:

ai = βXi + ²i

68
in Column 3 (the 1979 round) is 50 percent larger than the corresponding coefficient in Column 1,
while the coefficient in Column 5 (the 2000 round) is roughly the same size. The hypothesis that
the intensity of participation is uncorrelated with fractionalization is clearly rejected.21

3.4

Local Identity and Career Decisions

Having described a mechanism through which local identity could have persisted over many generations, we now proceed to study its economic consequences. We hypothesize that individuals
belonging to migrant denominations and born in high fractionalization counties should be endowed
with a stronger sense of local identity and will consequently shade their career choices away from
more mobile professional occupations. If the demand for professional occupations does not vary with
fractionalization, as suggested by the results in Table 4 with the 1860 county-controls, then labor
supply must respond to clear the market.
One supply-response strategy would be for individuals belonging to the non-migrant denominations in high fractionalization counties to select disproportionately into the professional occupations.
A second strategy would be for individuals from the migrant denominations to continue to select
into the professional occupations but to forego the relatively large returns to moving that are associated with those occupations. Finally, a third strategy would be for individuals born in other
counties and trained as professionals to move (at a cost) into the high fractionalization counties.
Once we introduce individual heterogeneity in the population and assume that the cost of becoming
a professional is decreasing in ability, it follows under reasonable conditions that all three strategies
will be adopted in equilibrium. This implies that individuals born in high fractionalization counties
will be less likely to become professionals on average. Professionals must move into those counties
and non-professionals must move out to clear the market.
If the objective is to avoid moving from the county of birth, why are non-professionals systematically over-supplied in the high fractionalization counties? The model developed in this section and
described in greater detail in the Appendix shows that this result will be obtained in competitive
θi ai = γXi + ui ,
where ²i , ui are mean-zero disturbance terms. It follow that γ̂ = E(θi )β̂ if cov(θi , Xi ) = 0.
21 The dependent variable in Columns 3-6 includes all individuals who attend church multiple times in the year,
roughly matching the definition of adherents in the CRB. We experimented with a more stringent measure of participation – multiple times per month – without qualitatively changing the results. The fractionalization coefficient
with the non-migrant denominations continues to be negative, small, and imprecisely estimated. The corresponding
coefficient for the migrant denominations in 1979 declines slightly to 0.23 and is no longer significant at conventional
levels, while the coefficient in 2000 remains significant and is just slightly smaller (0.18) than the point estimate in
Column 5. We continue to reject the hypothesis that the fractionalization coefficient is smaller with participation
rather than religious affiliation as the dependent variable, despite the fact that participation rates with the more
stringent measure fall by over 15 percentage points and are now below 0.7.

69
equilibrium as long as the third strategy is active and there is ex ante uncertainty in the demand
for different types of labor across counties.
To focus on the third supply-response strategy, we now ignore heterogeneity in ability and variation in local identity within counties. There are two counties in this economy with N individuals born
in each county in each period. Two types of jobs are available: professional and non-professional.
Individuals who are ex ante identical live for two periods, working in the second period of their lives.
Those individuals who choose to occupy more productive professional jobs must invest in training
in the first period of their lives. Individuals who expect to end up in non-professional jobs incur no
such cost. On average, sN professional jobs and (1 − s)N non-professional jobs are demanded in
each county in each period. However, these counties also face positive and negative demand shocks
with equal probability, which separately shift the demand for professional and non-professional labor by ²sN and ²(1 − s)N respectively, but leave the total number of professional jobs 2sN and
non-professional jobs 2(1 − s)N constant across the two counties in all periods.22
Once demand shocks are introduced, labor must flow across counties at the beginning of each
period to clear the market. In addition, we assume that professional jobs are also associated with
mid-career moves. This additional dimension of mobility in professional occupations is based on
the idea that the labor market for a school teacher or a secretary tends to be local. In contrast,
professionals such as university professors or management consultants are continually re-sorting
across local and regional labor markets as new opportunities arise and fresh cohorts of workers
enter. A professional can advance her career considerably with such a move if it becomes available
and we will later see that professionals are indeed much more likely to migrate from their county of
birth. We model this characteristic of professional occupations by assuming that the opportunity
to enhance overall productivity by switching a professional working in a particular county with a
professional working in the other county arrives with probability P in each period.
Individuals dislike moving, particularly those with a strong local identity. Let the cost of moving
be C1 for individuals born in county 1 and C2 < C1 for individuals born in county 2 with weaker
local identity. This is the only difference between workers in the two counties. We assume that
all workers are employed and that professional workers always take advantage of the productivity
enhancing career opportunities when they arise, in which case total output in this economy remains
constant across all states of the world. Notice that this assumption effectively rules out the second
labor-response strategy described above. Although wages for professionals and non-professionals will
22 Within a county, demand shocks for professionals and non-professionals could be positively or negatively correlated. We could also allow the size of the shocks to vary, ² for the professionals and ² for the non-professionals, without
changing the results reported below.

70
adjust across the two locations to clear the labor market in practice, the competitive labor allocation
can be conveniently derived as the solution to the Central Planner’s problem of minimizing expected
training and moving costs across both locations.
Let the supply of professional labor in county 1 be x1 . We show in the Appendix that expected
cost is a piece-wise linear function of x1 , as described in Figure 5, which is minimized at x∗1 =
sN − ²sN for P < 1/2(C1 + C2 )/(C1 − C2 ) and at x∗∗
1 = sN − ²sN − ²N (1 − 2s) for P > 1/2(C1 +
C2 )/(C1 −C2 ). The equilibrium supply of professional labor in county 1, with stronger local identity,
falls short of the expected demand sN for all values of P . Moreover, the share of professionals
supplied by county 1 is strictly less than the corresponding share in county 2.23
The intuition for this result is that as long as there is ex ante uncertainty in the type of jobs
demanded, some individuals will have to move in both counties with positive probability to clear
the market. By allocating a surplus of non-professionals to county 1, the Central Planner increases
the probability of such moves, but trades this off against the lower probability of having individuals
born in county 1 move ex post on the job. We show in the Appendix that the supply of professionals
in county 1 could also fall short of the expected demand without uncertainty, but only if P >
(C1 + C2 )/(C1 − C2 ).
In our set up, individuals born in county 1 are less likely to be professionals because they incur a
greater cost when they move (C1 > C2 ). However, this result could also be obtained if C1 = C2 and
the demand for professional labor is lower in county 1, s1 N < s2 N . The second prediction of the
model, which allows us to rule out this alternative demand-side explanation, is that once the labor
market has cleared, the share of professionals residing in the two counties should on average be the
same (sN ). In contrast, if the demand for professional labor is lower in county 1, then individuals
born and subsequently residing in that county should be less likely to be professionals.

3.5
3.5.1

Empirical Analysis
Individual Data

To test the predictions from the previous section we need information on the individual’s career
choice, county of birth, and county of residence (post-employment). The NLSY is the only largescale data set that we are aware of that includes this information. The data set includes basic
information on the respondent’s age, gender, race and, most importantly, county of birth. The
23 For P < 1/2(C + C )/(C − C ), the share of professionals is s − ²s in county 1 and s + ²s in county 2. For
1
2
1
2
P > 1/2(C1 + C2 )/(C1 − C2 ), the corresponding shares are s − ²(1 − s) and s + ²(1 − s).

71
Armed Forces Qualification Test (AFQT), which is designed to provide an unbiased measure of
the individual’s intelligence, was administered to all respondents in 1979. Subsequent survey rounds
collected contemporaneous information on educational attainment, employment, occupation, income,
and county of residence. We will study occupational choice and other outcomes related to that
economic decision at two points in time – 1994 and 2000 – when the respondents were old enough
to be settled in their careers and to have made some job-related moves. Occupational choices from
the NLSY in these years will be matched to census data on historical fractionalization, both in the
individual’s county of birth and the contemporaneous county of residence, to test the predictions of
the model.
Table 8 reports descriptive statistics for the individuals in our sample, who were on average
18 years old in 1979 and so around 33 years old in 1994 and 39 years old in 2000. Occupational
categories in the NLSY (up to the 2000 round) are based on the 1970 codes from the census. Professional occupations are defined to include relevant codes listed under the Professional, Technical,
and Kindred Workers category [1-196]. Job-related geographical mobility is the chief property that
distinguishes professional and non-professional occupations in our model. We consequently exclude
technical occupations and other occupations where career moves are unlikely to be important from
this category.24
Based on this occupational classification, 9 percent of the respondents hold professional jobs, with
little change from 1994 to 2000. 56 percent of all respondents had migrated out of their birth-county
by 1994, with an increase to 59 percent by 2000. Consistent with the assumption that professional
occupations are associated with greater mobility, 75 percent of the professionals and 53 percent of
the non-professionals had migrated out of their birth-county by 1994 (these differences are significant
at the 5 percent level and similar in 1994 and 2000).
Individuals in the sample are on average 33 years old by 1994 and should be established in the
labor market. Nevertheless, employment levels continue to increase over time, from 81 percent in
1994 to 92 percent in 2000. Conditional on being employed, annual income (in 2000 dollars) also
increases from 28,000 in 1994 to 33,000 in 2000. These changes in employment and income are
presumably life-cycle effects, but they could, in principle, be due to selective attrition since this is
a longitudinal survey. Notice, however, that racial composition and the proportion of women in the
sample are very stable over the 1994-2000 period. Thus, we do not observe selective attrition from
the sample, at least with respect to two important demographic characteristics that are associated
24 These occupations include Librarians [32], Nurses, dieticians, and therapists [74-76], Religious workers [86,90],
Social and recreation workers [100,101], Teachers, except college and university [141-145], and Technicians [150-174].

72
with income and employment.

3.5.2

Fractionalization in the County of Birth and Occupational Choice

The first prediction from the model of occupational choice incorporating local identity is that individuals born in counties with greater ethnic fractionalization in 1860 should be less likely to hold
professional jobs. Including race, gender, and age as regressors (although their omission would not
affect the results) we see in Table 9, Columns 1-2 that individuals born in high fractionalization
counties are indeed less likely to hold professional jobs in the 1994 and 2000 rounds of the NLSY.
The individuals in our sample are drawn from 150 of the approximately 400 Midwestern counties
that were incorporated and had attracted foreign migrants by 1860. While it thus seems unlikely
that a few outlying counties are driving the results, we nevertheless report nonparametric estimates
of the relationship between occupational choice and historical fractionalization in Figure 6. We see
that the probability that the individual is a professional declines steadily with fractionalization, both
in 1994 and in 2000, verifying the robustness of our results.25
Our test of the hypothesis that identity shapes occupational choice relies on the assumption that
fractionalization is uncorrelated with the demand for professional workers in the county. Recall from
Table 4, however, that historical fractionalization had a positive and significant effect on current
economic characteristics, measured by population in the county as well as racial and ethnic fractionalization, but that this effect disappeared once important characteristics of the 1860 economy were
included as regressors. More urban counties will have a greater demand for professional workers,
shifting up the coefficient on the fractionalization variable in Table 9, Columns 1-2. The regressions that follow will thus include the same 1860 characteristics as in Tables 4 and 5 – population,
manufacturing share and agriculture share – to control for variation in the demand for professional
workers across counties.
The prediction that high fractionalization counties should supply fewer professional workers is
derived conditional on the total surplus that was historically available to migrants in the labor market
R and the total number of migrant workers N . Everything else equal, an increase in R would increase
the cost of exit at the margin, encouraging efforts to instill local identity historically and lowering
the propensity to enter professional occupations today. Superior transportation infrastructure in
1860 was associated with high fractionalization and a larger county population in Table 3 and
25 The nonparametric regressions are estimated using the Epanechnikov kernel smoothing function. Less than 5
percent of the observations in Table 9, Columns 1-2 are drawn from counties with fractionalization below 0.45 and so
too much weight should not be placed on the extremely steep initial decline in Figure 6. Although all the parametric
regressions that follow will use the full set of counties, we verified that the fractionalization coefficient is unchanged
when the sample is restricted to individuals drawn from counties in the 0.45-0.8 range.

73
so we expect fractionalization to be positively correlated with R. It follows that exclusion of R
from the occupation choice regression in Table 9, Columns 1-2 would have shifted the coefficient
on fractionalization down. The same 1860 characteristics that we described above could be used
to control for R, but since the professional demand effect and the surplus effect work in opposite
directions, the effect of their inclusion on the fractionalization coefficient would now be ambiguous.
By the same argument, we can no longer predict the sign of the coefficients on the 1860 characteristics
when they are included in the occupation choice regression. For example, larger 1860 population
is associated with urbanization today and, hence, an increased demand for professional jobs. At
the same time, a larger 1860 population is associated with a larger R, stronger identity and, hence,
a reduced supply of professional workers. Finally, we would want to account for the number of
migrants N when estimating the effect of fractionalization on occupational choice. This variable is
highly correlated with county population (the correlation is 0.9) and so will be omitted from the
regressions that follow. 1860 population captures the professional demand effect, the surplus effect,
and the effect of N on occupational choice.
We see in Table 9, Columns 3-4 that the coefficient on 1860 fractionalization becomes more
negative and is more precisely estimated once the additional county-level controls are included in
the occupational choice regressions. Based on the estimates in Column 4, a one standard deviation
decline in ethnic fractionalization would increase the probability of holding a professional job from 9
percent to as much as 15 percent. Individuals born in counties with a greater share of manufacturing
and agricultural jobs in 1860 are also less likely to hold professional jobs. Population in 1860, in
contrast, has no effect on occupational choice in both the 1994 and the 2000 rounds. Finally, women
and Non-Whites are significantly less likely to hold professional jobs in Columns 1-4.
Figure 4 traces a continuous link between initial conditions in 1860 and participation in select
religious denominations historically dominated by the migrants over the subsequent 140 years. Table
9 brings the different pieces of the empirical analysis together by documenting the relationship
between 1860 fractionalization and occupational choice in 2000. We close this section by discussing
a number of tests that we conducted to verify the robustness of the key occupational choice result.
1. Alternative construction of the fractionalization variable: First, we computed ethnic
fractionalization in 1860 with men only. Both men and women participated in the workforce in
the nineteenth and early twentieth centuries, with ethnic networks channeling women into jobs as
well. Bodnar (1980), for example, cites a 1930 study of two thousand foreign-born women, most
of whom reported that they had secured their first job through social connections. Nevertheless,

74
we might expect labor networks to have been organized along gender lines within ethnic groups,
with male networks occupying a dominant position in the labor market and in the communities they
were drawn from. The coefficient on the alternative fractionalization measure (computed using men
alone) continues to be negative and significant, although it is smaller in size.
Inspection of the occupation categories in Table 2 indicates that farmers made up 50 percent
of the workforce in 1860. Ethnic competition may have been less relevant in this category and so
we recomputed the county-level fractionalization statistic placing zero weight on fractionalization
among the farmers as a second robustness test.26 The fractionalization coefficient becomes slightly
larger in absolute magnitude and is more precisely estimated.
As a third test, we constructed the fractionalization variable using alternative occupational classifications. The county-level statistic that we use in all the regressions reported in the paper is
computed as the weighted average of ethnic fractionalization within each of the 11 occupational categories listed in Table 2. We measured labor market rivalry within a coarser set of four occupational
categories as well as within a finer set of 139 occupations, without changing any of the results.27
2. Alternative construction of the occupation variable: First, we expanded the set of
professional occupations by including individuals assigned to the “managers and administrators”
code [245] within the broader Managers and Administrators, except Farm category [201-245]. Most
of the specific occupations listed in this broad category do not conform to our definition of a professional occupation and the “managers and administrators” classification does not provide much
information on the actual nature of jobs that it covers. However, 74.7 percent of the individuals
assigned to that category migrated out of the county of their birth, which is close to the level of
migration for the other occupations that we include in the professional category. Inclusion of the
managers and administrators in the professional classification increased the share of professionals
in the sample to 17.7 percent. Although these results need to to interpreted with caution since the
additional professional occupations are selected on the basis of an outcome (migration) rather than
their fundamental characteristics, we nevertheless verified that the fractionalization coefficient in
26 We include all individuals who report being employed, regardless of their age, when computing the fractionalization
statistic since there were no age restrictions on employment at that time. Apart from the 11 broad occupational
categories in Table 2, some individuals in the census were also assigned to an undefined occupational category.
Women were disproportionately represented in this category, which presumably covers home production and other
informal activities. As with the farmers, we assume that networks were less relevant in this category and assign it
zero weight when computing ethnic fractionalization in all the regressions that we report. Nevertheless, we verified
that assigning a weight to this category based on its share of the migrant workforce in the county had no effect on
the estimated fractionalization coefficient in the occupational choice regression.
27 The four aggregate categories are white collar, agriculture, manufacturing, and service and laborers. These
categories correspond to the broad headings in Table 2 except that Blue collar, nonfarm is divided into manufacturing
(craftsman, operative) and service and laborers (household service, service, laborer non-farm). The 139 disaggregate
occupations are defined at the three-digit census code level. 84 percent of the listed occupations in the 1860 data are
drawn from just 11 of these occupations.

75
the occupational choice regression with this expanded definition of professional activity remained
negative and significant.
As a second test we ran the occupational choice regression in other years – 1993, 1996, 1998 –
generating a pattern of fractionalization coefficients very similar to the point estimates obtained in
1994 and 2000.
3. Alternative specification of the occupation regression: First, we replaced total population with the migrant population and, separately, with the workforce, in the occupational choice
regression. These variables are all highly correlated, and not surprisingly the results were unchanged.
Second, we included the share of migrants in the workforce, and the interaction of this variable
with ethnic fractionalization, as additional regressors. This specification allows for the possibility
that ethnic fractionalization had a larger effect in counties where migrants made up a larger share
of the total population. However, both these variables had an insignificant effect on occupational
choice, leaving the (uninteracted) fractionalization coefficient unchanged.
Third, we included the population share of each ethnicity in the 1860 census, computed at the
level of the county, as additional regressors to allow for the possibility that individuals with particular
ethnic ancestry continue to be concentrated in specific occupations today. Given the high rate of
inter-county migration, we do not expect the ethnic shares in 1860 to be relevant today and not
surprisingly the fractionalization coefficient was hardly affected by the inclusion of these additional
regressors.

3.5.3

Fractionalization and Outcomes Related to Occupational Choice

An individual born in a county with higher ethnic fractionalization in 1860 is less likely to select
into a professional occupation in 1994 and 2000. Our interpretation of this result is that individuals
born in high fractionalization counties identify strongly with their local communities and so wish to
avoid the spatial mobility that comes with professional occupations. The regression results reported
in Table 10, Columns 1-2 indicate that individuals from high fractionalization counties are indeed
significantly less likely to migrate from the county of their birth.28 On average, around 58 percent
of the individuals in the sample migrate from the county of birth. The point estimates indicate
that a one standard deviation increase in fractionalization reduces migration by 8 percentage points
28 Recall that one supply-response strategy was for individuals from migrant denominations to continue to select
into the professional occupations but to forego the relatively large returns to moving that are associated with those
occupations. Although it is tempting to estimate the relationship between migration and fractionalization separately
for the professionals, note that occupational choice is jointly determined with migration in our framework. By the
same argument, we cannot estimate the relationship between occupational choice and fractionalization separately by
religious denomination.

76
(a 14 percent decline). Among the other regressors, none of the 1860 county-level controls significantly affect migration, but Whites are significantly more likely to move. Age is also (mechanically)
positively associated with migration.
Does the effect of fractionalization on occupational choice and migration that we have just
described have financial consequences? The results in Table 10, Columns 3-4 indicate that while
individuals born in counties with a greater share of manufacturing and agriculture in 1860 are
significantly more likely to hold a job by 2000, employment levels do not vary significantly with
ethnic fractionalization in 1994 or 2000. Whites and males are, not surprisingly, significantly more
likely to be employed, although the importance of these individual characteristics declines over time.
In contrast with the results for employment, the income regressions reported next in Table 10,
Columns 5-6 indicate that high fractionalization is associated with significantly lower income (in
2000), conditional on being employed. Average annual income in 2000 was 33,000 dollars, and so
our estimates indicate that a one standard deviation increase in fractionalization would have reduced
income by 2,300 dollars (a 7 percent decrease). A greater share of manufacturing and agriculture in
1860 is associated with lower income in 2000, with Whites and males earning significantly more in
1994 and 2000.

3.5.4

Alternative Explanations

1. Fractionalization and the demand for professionals: The second prediction of the model
is that historical fractionalization in the county of residence should have no effect on occupational
choice. This implies that non-professional workers must flow out of the high fractionalization counties, while professional workers flow in to clear the labor market. This prediction rules out the
possibility that differences in the demand for professional labor across counties are driving our
results.
To test this prediction we regress occupation on fractionalization in the county of residence in
Table 11. The fractionalization coefficient is small and insignificant in 1994 and 2000, with and
without the county controls in Columns 1-4. It actually changes sign and is positive in three of the
four specifications. Individuals working in high fractionalization counties are at least as likely to be
professionals as individuals working in low fractionalization counties. It is only individuals born in
those counties who are less likely to select into those occupations.29
29 One remaining concern is that the mismatch between the supply and the demand for professional labor in the
high fractionalization counties could have been driven by unexpected changes in the demand for professionals in those
counties rather than a strong local identity. Suppose, for example, that the demand for professionals was systematically
lower in high fractionalization counties historically and then increased in the 1980’s, with the restructuring of the U.S.

77
2. Fractionalization and educational attainment: Could variation in individual ability or
the quality of the school system have generated the observed variation in the supply of professionals
across counties? High fractionalization counties had superior transportation infrastructure in 1860
and so were most likely growing relatively rapidly at that time. It is entirely possible that particular
types of migrants were attracted to those counties, although we would expect migrants drawn to
areas with many competing ethnic groups to have been positively selected on innate talent. Given
the high rates of inter-county migration, we would in any case expect few families to have maintained
an unbroken line of descent to the present day in the same county. We noted earlier that while 1860
fractionalization may have been positively correlated with relevant measures of social heterogeneity
many decades later when the secondary school system was being established, fractionalization was
also positively correlated with religious participation. The net effect of fractionalization on investments in education was thus seen to be ambiguous, and reassuringly, 1860 fractionalization was
uncorrelated with education expenditures in 1990 in Table 5.
To provide direct support for the idea that individuals born in high fractionalization counties
are no less capable and no less prepared to invest in the human capital that is necessary to secure
professional jobs, we regress measures of educational attainment obtained from the NLSY on 1860
fractionalization in Table 12. We see in Table 12, Columns 1-3 that fractionalization, and all the
1860 county-controls for that matter, have no effect on AFQT scores, high school completion, and
college completion.30 In contrast, individual characteristics have a strong effect on AFQT scores
and college completion.
Why are individuals born in high fractionalization counties as likely to attend and complete
college even though they expect to end up in non-professional occupations? To explore this result
further, we construct a binary variable that indicates whether the college major would have channeled
the individual into a less mobile occupation.31 Individuals born in high fractionalization counties and
females, not surprisingly, are significantly more likely to choose such majors in Table 12, Column 4.
On average, 39 percent of the individuals in the sample who complete college choose majors that are
economy, after our NLSY cohort had made their career choices. To rule out this possibility, we regressed manufacturing
share, agriculture share, and county population in 1970, obtained from the 1972 County Data Book, on the same 1860
characteristics as in Table 4, Columns 7-9. 1860 fractionalization has no effect on economic conditions in 1970, just
as we saw earlier with the same conditions in 1990.
30 The fact that fractionalization is uncorrelated with education also rules out an alternative channel through which
fractionalization could affect religious participation (see Sacerdote and Glaeser [2001] for an analysis of the complex
relationship between education and religious participation).
31 The non-professional (less mobile) majors include agriculture and natural resources, accounting, hotel and restaurant management, marketing and purchasing, transportation and public utilities, real estate, insurance, secretarial
studies, personnel management, labor and industrial relations, communications, data processing, education, health
professions (except allied health, other, veterinary medicine specialties, speech pathology and audiology), home economics, library science, psychology for counseling, rehabilitation counseling, community services, parks and recreation
management, social work and helping services, law enforcement, administration of justice, and theology.

78
associated with lower mobility. The point estimates indicate that a one standard deviation increase
in fractionalization would increase the likelihood of such a major being chosen by four percentage
points (a 10 percent increase). Once again, we uncover evidence that suggests that individuals born
in high fractionalization counties are making conscious choices that restrict their future mobility.

3.6

Conclusion

This paper draws a connection between ethnic labor networks in the American Midwest when it was
first being settled, the local identity that emerged endogenously to support these networks and then
persisted over many generations, and occupational choice today. Individuals born in counties with
greater ethnic fractionalization in 1860, which we expect to be associated with stronger local identity
and better functioning parochial institutions, are significantly less likely to select into professional
occupations, which come with greater geographical mobility, in 2000.
The results in this paper are relevant to the ongoing debate on “economic institutions” (Acemoglu, Johnson and Robinson [2001, 2002]) versus “culture” (Barro and McCleary 2003, Tabellini
2005, Fernandez and Fogli 2007) as determinants of growth. We find that a local cultural trait
generates significant variation in occupational choice within a relatively homogenous region – the
American Midwest – and it is possible that cultural effects of this sort would be even larger across
countries or regions with very different histories. At the same time, culture cannot be sustained
without institutional support. Social institutions such as the church and the family help sustain
cultural traits, which in turn keep these institutions alive. As long as these institutions continue to
be useful, cultural traits can persist long after the economic circumstances that gave rise to them
have ceased to be relevant. While the economics literature has focussed much of its attention on
economic and political institutions, social institutions, with their complementary cultural traits,
might have important effects on growth as well.

79

3.7

References

Acemoglu, Daron, Simon Johnson, and James A. Robinson. 2001. The Colonial Origins of Comparative Development: An Empirical Investigation. American Economic Review, 91(5): 1369-1401.
Acemoglu, Daron, Simon Johnson, and James A. Robinson. 2002. Reversal of Fortune: Geography
and Institutions in the Making of the Modern World Income Distribution. Quarterly Journal of
Economics, 117(4): 1231-1294.
Akerlof, George A. and Rachel E. Kranton. 2000. Economics and Identity, Quarterly Journal of
Economics 105(3): 715-53.
Akerlof, George A. and Rachel E. Kranton. 2005. Identity and the Economics of Organizations,
Journal of Economic Perspectives 19(1): 9-32.
Alba, Richard D. 1990. Ethnic Identity: The Transformation of White America. New Haven and
London: Yale University Press.
Alesina, Alberto, Reza Baqir, and William Easterly. 1999. Public Goods and Ethnic Division,
Quarterly Journal of Economics 114(4): 1243-84.
Alesina, Alberto, Reza Baqir, and Caroline Hoxby. 2004. Political Jurisdictions in Heterogeneous
Communities, Journal of Political Economy 112(2): 348-396.
Alexander, June Granatir. 1991. Moving into and out of Pittsburgh: Ongoing Chain Migration, in
A Century of European Migrants, 1830-1930, eds. Rudolph J. Vecoli and Suzanne M. Sinke, Urbana:
University of Illinois Press.
Barro, Robert J. and Rachel M. McCleary. 2003. Religion and Economic Growth Across Countries,
American Sociological Review 68: 760-781.
Barton, Josef J. 1975. Peasants and Strangers: Italians, Rumanians, and Slovaks in an American
City, 1890-1950. Cambridge, Massachusetts: Harvard University Press.
Bénabou, Roland and Jean Tirole. 2006. Belief in a Just World and Redistributive Politics, Quarterly Journal of Economics 121(2): 699-746.
Bisin, Alberto and Thierry Verdier. 2000. “Beyond the Melting Pot”: Cultural Transmission, Marriage, and the Evolution of Ethnic and Religious Traits, Quarterly Journal of Economics 115(Au-

80
gust): 955-988.
Bisin, Alberto, Georgio Topa, and Thierry Verdier. 2004. Religious Intermarriage and Socialization
in the United States, Journal of Political Economy 112(3): 615-664.
Bodnar, John. 1980. Immigration, Kinship, and the Rise of Working-Class Realism in Industrial
America, Journal of Social History. 45-65.
Bodnar, John. 1985. The Transplanted: A History of Immigrants in Urban America. Bloomington:
Indiana University Press.
Bodnar, John, Roger Simon, and Michael P. Weber. 1982. Lives of Their Own: Blacks, Italians,
and Poles in Pittsburgh, 1900-1960. Urbana: University of Illinois Press.
Elder Jr., Glen H. and Rand D. Conger. 2000. Children of the Land: Adversity and Success in Rural
America. Chicago and London: University of Chicago Press.
Fernandez, Raquel and Alessandra Fogli. 2007. Culture: An Empirical Investigation of Beliefs,
Work, and Fertility. New York University Department of Economics mimeograph.
Finke, Roger and Rodney Stark. 1992. The Churching of America, 1776-1990. New Brunswick, NJ:
Rutgers University Press.
Fishlow, Albert. 2000. Internal Transportation in the Nineteenth and Early Twentieth Centuries,
in The Cambridge Economic History of the United States, eds. Stanley L. Engerman and Robert E.
Gallman, Cambridge, England: Cambridge University Press.
Gans, Herbert. 1962. The Urban Villagers: Group and Class in the Life of Italian-Americans. New
York: The Free Press of Glencoe.
Gans, Herbert. 1979. Symbolic Ethnicity: The Future of Ethnic Groups and Cultures in America,
Ethnic and Racial Studies 2(January): 1-20.
Gerson, Kathleen, C. Ann Stueve, and Claude S. Fischer. 1977. Attachment to Place, in Networks
and Places: Social Relations in the Urban Setting, eds. Claude S. Fischer et al., New York: The
Free Press.
Gjerde, Jon. 1985. From Peasants to Farmers: The Migration from Balestrand, Norway to the
Upper Middle West. Cambridge, England: Cambridge University Press.

81
Gjerde, Jon. 1991. Chain Migrations from the West Coast of Norway, in A Century of European
Migrants, 1830-1930, eds. Rudolph J. Vecoli and Suzanne M. Sinke, Urbana: University of Illinois
Press.
Gjerde, Jon. 1997. The Minds of the West: Ethnocultural Evolution in the Rural Middle West,
1830-1917. Chapel Hill: University of North Carolina Press.
Goldin, Claudia and Lawrence Katz. 1999. Human Capital and Social Capital: The Rise of Secondary Schooling in America, 1910-1940, Journal of Interdisciplinary History 29 (March): 683-723.
Gordon, David M., Richard Edwards, and Michael Reich. 1982. Segmented Work, Divided Workers:
The Historical Transformation of Labor in the United States. Cambridge England: Cambridge
University Press.
Greif, Avner. 2006. Institutions and the Path to the Modern Economy: Lessons from Medieval
Trade. Cambridge England: Cambridge University Press.
Gruber, Jonathan. 2005. Religious Market Structure, Religious Participation, and Outcomes: Is
Religion Good for You? NBER Working Paper 11377.
Hoerder, Dirk. 1991. International Labor Markets and Community Building by Migrant Workers
in the Atlantic Economies, in A Century of European Migrants, 1830-1930, eds. Rudolph J. Vecoli
and Suzanne M. Sinke, Urbana: University of Illinois Press.
Hummon, David M. 1990. Commonplaces: Community Ideology and Identity in American Culture.
Albany: State University of New York Press.
Hunter, Albert. 1975. The Loss of Community: An Empirical Test Through Replication, American
Sociological Review 40(5): 537-552.
Hutchinson, E. P. 1956. Immigrants and Their Children 1850-1950 New York: John Wiley and
Sons, Inc.
Iannaccone, Laurence R. 1992. Sacrifice and Stigma: Reducing Free-Riding in Cults, Communes,
and Other Collectives, Journal of Political Economy 100(2): 271-297.
Iannaccone, Laurence R. 1998. Introduction to the Economics of Religion, Journal of Economic
Literature 36: 1465-1496.

82
Kamphoefner, Walter D. 1987. The Westfalians: From Germany to Missouri. Princeton: Princeton
University Press.
Kim, Sukkoo. 2007. Immigration, Industrial Revolution and Urban Growth in the United States,
1820-1920: Factor Endowments, Technology and Geography, NBER Working Paper No. 12900.
Margo, Robert A. 1999. Regional Wage Gaps and the Settlement of the Midwest, Explorations in
Economic History 36: 128-143.
Meyer, David R. 1989. Midwestern Industrialization and the American Manufacturing Belt in the
Nineteenth Century, Journal of Economic History 49(4): 921-937.
Sacerdote, Bruce and Edward L. Glaeser. 2001. Education and Religion, NBER Working Paper No.
8080.
Sanders, Scott Russell. 2007. To Stay or Move On: The Artist’s Dilemma. Bloom, February/March:
65-68.
Sumner, William G. 1906. Folkways. New York: Ginn.
Tabellini, G. 2005. Culture and Institutions: Economic Development in the Regions of Europe,
CESifo Working Paper No. 1492.
Thistlethwaite, Frank. 1991. Migration from Europe Overseas in the Nineteenth and Twentieth
Centuries, in A Century of European Migrants, 1830-1930, eds. Rudolph J. Vecoli and Suzanne M.
Sinke, Urbana: University of Illinois Press.
Tullock, Gordon. 1980. Efficient Rent Seeking, in Toward a Theory of the Rent-Seeking Society, eds.
James M. Buchanan, Robert D. Tollison, and Gordon Tullock, College Station: Texas A&M Press.
Weber, Max. 1930. The Protestant Ethic and the Spirit of Capitalism. London, England: Allen and
Unwin.

83

3.8
3.8.1

Appendix: Labor Market Equilibrium
Population and Production Technology

Two types of jobs are available in this economy: professional and non-professional. Individuals
who are ex ante identical live for two periods, working in the second period of their lives. Those
individuals who choose to occupy professional jobs incur a training cost Ce in the first period of their
lives. Individuals who expect to end up in non-professional jobs incur no such cost. The expected
output obtained from a professional worker who takes advantage of career opportunities when they
arise is θ and the output obtained from an non-professional worker with certainty is θ < θ.
There are two counties in this economy with N individuals born in each county in each period.
On average, sN professional jobs and (1 − s)N non-professional jobs are demanded in each county in
each period. However, these counties also face demand shocks, with two states of the world occurring
with equal probability:
State 1: sN + ²sN professional and (1 − s)N + ²(1 − s)N non-professional jobs in county 1.
sN − ²sN professional and (1 − s)N − ²(1 − s)N non-professional jobs in county 2.
State 2: sN − ²sN professional and (1 − s)N − ²(1 − s)N non-professional jobs in county 1.
sN + ²sN professional and (1 − s)N + ²(1 − s)N non-professional jobs in county 2.
Notice that these demand shocks are skill neutral, in the sense that the probability of receiving
a shock is the same for professional and non-professional workers within each county. The shocks
are also positively correlated for professional and non-professional workers within a county. We
could relax each of these assumptions without changing any of the results that follow.32 In addition,
the opportunity to enhance overall productivity by switching a professional working in a particular
county with a professional working in the other county arrives with probability P in each period.
Let the cost of moving be C1 for individuals born in county 1 and C2 < C1 for individuals
born in county 2 with weaker local identity. We assume that all workers are employed and that
professional workers always take advantage of the productivity enhancing career opportunities when
they arise, in which case total output in this economy remains constant across all states of the
£
¤
world: 2N sθ + (1 − s)θ . The competitive labor market equilibrium can be obtained in that case
as the solution to the Central Planner’s problem of minimizing training and moving costs across
both locations.
32 Specifically, we allowed shocks to be perfectly negatively correlated for professional and non-professional workers
within each county. We also allowed the size of the shocks to vary; ² for professionals and ² for non-professionals.

84

3.8.2

The Central Planner’s Problem

Let the supply of professional labor in county 1 be x1 . From the structure of the demand shocks it
then follows that the supply of professional labor in county 2 will be 2sN − x1 and that the supply
of non-professional labor in county 1 will be N − x1 . We derive x1 as the solution to the Central
Planner’s cost minimization problem. Expected cost turns out to be a piece-wise linear function of
x1 and so it will be convenient to solve what is essentially a linear programming problem in three
regimes:
Regime 1: x1 ∈ [sN − ²sN, sN + ²sN ]
The supply of professional labor in each county is sufficient to satisfy the minimum demand in
that county but does not exceed the maximum demand.
The labor flow in each state of the world can then be derived as:
Flow in state 1: sN + ²sN − x1 professional labor from county 2 to county 1.
(1 − s)N + ²(1 − s)N − (N − x1 ) non-professional labor from county 2 to county 1.
Flow in state 2: x1 − [sN − ²sN ] professional labor from county 1 to county 2.
(N − x1 ) − [(1 − s)N − ²(1 − s)N ] non-professional labor from county 1 to county 2.
The Central Planner chooses x1 to minimize expected cost
1
E(C) = 2sN Ce + ²N (C1 + C2 ) + P x1 C1 + P (2sN − x1 )C2 ,
2

(3.1)

where the second term on the right hand side is the cost associated with movement at the
start of the period and the last two terms reflect movement of professional labor during the period.
Because C1 > C2 it is easy to verify that E(C) is increasing linearly in x1 and will be minimized at
x∗1 = sN − ²sN . Labor flows with x1 = x∗1 are then obtained as:
Flow in state 1 (x1 = x∗1 ): 2²sN professional labor from county 2 to county 1.
²N (1 − 2s) non-professional labor from county 2 to county 1.33
Flow in state 2 (x1 = x∗1 ): No flow of professional labor.
²N non-professional labor from county 1 to county 2.
Regime 2: x1 ∈ [sN − ²sN − ²N (1 − 2s), sN − ²sN ]
We now reduce the supply of non-professional labor in county 2, but at most to the point where no
non-professional labor flows to county 1 in state 1. In our set up, any reduction in non-professional
33 To generate a positive labor flow we require s ≤ 1/2. This is a reasonable assumption since just a small fraction
of jobs (9 percent in our data) are professional. For the case with asymmetric shocks, ² for the professionals and ² for
the non-professionals, the corresponding condition is s ≤ 1/(1 + ²/²).

85
labor supply in county 2 must lead to a reduction in professional labor in county 1 by the same
amount. It then follows that the supply of professional labor in county 1 will no longer be sufficient
to meet even the minimum demand in that county, while the supply of professional labor in county
2 will exceed the maximum demand in that county. Labor flows at the beginning of the period will
necessarily increase, with an accompanying increase in moving costs, but we will see that this may
be outweighed by the reduced cost of recounty for professional workers from county 1.
The labor flow in each state is derived as:
Flow in state 1: sN + ²sN − x1 professional labor from county 2 to county 1.
(1 − s)N + ²(1 − s)N − (N − x1 ) non-professional labor from county 2 to county 1.
Flow in state 2: sN − ²sN − x1 professional labor from county 2 to county 1.
(N − x1 ) − [(1 − s)N − ²(1 − s)N ] non-professional labor from county 1 to county 2.
The expected cost can then be expressed as:

E(C) = 2sN Ce +

1
[²N + sN (1 − ²) − x1 ] (C1 + C2 ) + P x1 C1 + P (2sN − x1 )C2 .
2

(3.2)

Collecting terms, E(C) is declining linearly in x1 and continues to be minimized at x∗1 = sN −²sN
if P < 1/2(C1 +C2 )/(C1 −C2 ). However, the local minimum is obtained at x∗∗
1 = sN −²sN −²N (1−
2s), if the sign of the inequality is reversed. This will be the case if P and C1 − C2 are sufficiently
large. Substituting x∗∗
1 in equation (3.2) above, it is easy to verify that the term in square brackets
is greater than ²N , the term corresponding to it in equation (3.1), which implies that moving costs
at the beginning of the period increase when going from x∗1 to x∗∗
1 . However, moving costs during
the period decrease with the reduction in x1 , and this effect dominates under the conditions on P
and C1 − C2 derived above. Labor flows with x1 = x∗∗
1 are obtained as:
Flow in state 1 (x1 = x∗∗
1 ): ²N professional labor from county 2 to county 1.
No flow of non-professional labor.
Flow in state 2 (x1 = x∗∗
1 ): ²N (1 − 2s) professional labor from county 2 to county 1.
2²N (1 − s) non-professional labor from county 1 to county 2.
Regime 3: x1 ∈ [0, sN − ²sN − ²N (1 − 2s)]
We now reduce the supply of professional labor in county 1, with an accompanying increase in
non-professional labor, even further so that professional labor flows from county 2 to county 1 and
non-professional labor flows in the opposite direction in both states of the world.

86
Labor flows are now derived as:
Flow in state 1: sN + ²sN − x1 professional labor from county 2 to county 1.
(N − x1 ) − [(1 − s)N + ²(1 − s)N ] non-professional labor from county 1 to county 2.
Flow in state 2: sN − ²sN − x1 professional labor from county 2 to county 1.
(N − x1 ) − [(1 − s)N − ²(1 − s)N ] non-professional labor from county 1 to county 2.
The corresponding expected cost expression is obtained as:

E(C) = 2sN Ce +

1
[2(sN − x1 )] (C1 + C2 ) + P x1 C1 + P (2sN − x1 )C2 .
2

(3.3)

It is straightforward to verify that E(C) is unambiguously decreasing in x1 and, hence, is minimized at x∗∗
1 = sN − ²sN − ²N (1 − 2s).

3.8.3

Equilibrium Labor Allocation

As described in Figure 5 and derived above, E(C) is increasing in x1 to the right of sN − ²sN and
decreasing in x1 to the left of sN − ²sN − ²N (1 − 2s) for all values of P . For x1 ∈ [sN − ²sN −
²N (1 − 2S), sN − ²sN ], E(C) is declining in x1 for P < 1/2(C1 + C2 )/(C1 − C2 ), whereas E(C) is
increasing in x1 when the sign of the inequality is reversed. The global minimum is consequently
obtained at x∗1 = sN − ²sN for P < 1/2(C1 + C2 )/(C1 − C2 ) and at x∗∗
1 = sN − ²sN − ²N (1 − 2s)
for P > 1/2(C1 + C2 )/(C1 − C2 ). The supply of professional labor in county 1 falls short of the
expected demand sN for all values of P . Without uncertainty, the supply of professionals in county
1 could still fall short of the expected demand, but only if P > (C1 + C2 )/(C1 − C2 ).34

34 Without uncertainty in labor demand, sN professional and (1 − s)N non-professional jobs are available in each
county in each period. Let x1 ∈ [0, sN ] measure the supply of professional workers in county 1. It then follows that
sN − x1 professional workers would flow from county 2 to county 1 and (N − x1 ) − (1 − s)N non-professional workers
would flow in the opposite direction at the beginning of each period. Using the same notation as above, the Central
Planner chooses x1 to minimize

E(C) = 2sN Ce + (sN − x1 ) (C1 + C2 ) + P x1 C1 + P (2sN − x1 )C2 .
It follows that x1 = sN if P < (C1 + C2 ) / (C1 − C2 ). x1 = 0 and professional labor is under-supplied in county 1 if
the sign of the inequality is reversed.

0

100

200

300

400

500

600

700

1860

1870

1880

1890

1900

Counties with railroads

Year

Incorporated counties

1850

Figure 3.1: Early Growth in the Midwest

87

Iowa
Illinois

Minnesota

Lake
Michigan

Wisconsin

Indiana

Ohio

Michigan

0.81 - 0.90

0.71 - 0.80

0.61 - 0.70

0.51 - 0.60

0.41 - 0.50

0.31 - 0.40

0.21 - 0.30

0.11 - 0.20

0.00 - 0.10

N/A

Ethnic Fractionalization

Figure 3 . 2: Ethnic Fractionalization in 1860

88

Fractionalization coefficient in each
census year (with 95% confidence band)

-0.3

-0.2

-0.1

0

0.1

0.2

0.3

1870

Figure 3.3: Religious Participation

1890
1906
1916
1926
Year

1936
1952
1972
1980
1990
2000

89

Fractionalization coefficient in each
census year (with 95% confidence band)

-0.4

-0.3

-0.2

-0.1

0

0.1

0.2

0.3

0.4

Figure 3.4: Religious Participation by Denomination

1870
1890
1906
1916
1926

Year

1936
1952
1972
1980
1990
2000

90

expected
cost

sN-εsN-εN(1-2s)

sN+εsN

share of professionals in county 1

sN-εsN

P< ½ (C1+C2)/(C1-C2)

P> ½ (C1+C2)/(C1-C2)

Figure 3.5: The Central Planner's Problem

91

92

Figure 3.6

93

Table 3.1: Ethnic Distribution, 1860-1900

Census year:

1860
(1)

1880
(2)

1900
(3)

Scandinavia
Danish
Finish
Norwegian
Swedish

0.01
0.00
0.03
0.02

0.02
0.00
0.07
0.06

0.02
0.01
0.07
0.10

British Isles
English
Irish
Scottish
Welsh

0.13
0.25
0.03
0.01

0.11
0.19
0.03
0.01

0.09
0.11
0.02
0.01

Western Europe
Belgian
Dutch
French
German
Italian
Swiss

0.00
0.01
0.03
0.32
0.00
0.02

0.01
0.01
0.02
0.37
0.00
0.02

0.01
0.01
0.01
0.41
0.01
0.02

Eastern Europe
Czech
Hungarian
Polish
USSR
Other

0.00
0.00
0.00
0.00
0.14

0.01
0.00
0.01
0.00
0.05

0.02
0.00
0.02
0.01
0.04

Total

1.00

1.00

1.00

Source: IPUMS 1:100 sample, including all foreign-born individuals.

94

Table 3.2: Occupational Distribution, 1860-1900

Census year:

1860
(1)

1880
(2)

1900
(3)

White collar
Professional
Manager
Clerical
Sales

0.04
0.04
0.00
0.01

0.04
0.04
0.01
0.02

0.05
0.06
0.02
0.03

Farm
Farmer
Laborer, Farm

0.50
0.12

0.41
0.17

0.31
0.17

Blue collar, nonfarm
Craftsman
Operative
Household Service
Service
Laborer, Non-Farm

0.10
0.05
0.05
0.00
0.09

0.08
0.08
0.05
0.01
0.10

0.09
0.09
0.05
0.02
0.12

Total

1.00

1.00

1.00

Source: IPUMS 1:100 sample, including all foreign-born individuals
who report that they are employed and report an occupational category.

0.049
(0.022)
-0.668
(0.191)
-0.252
(0.073)
401

0.001
(0.008)
0.169
(0.097)
-0.066
(0.029)
401

manufacturing
share
(2)
(3)

-0.013
(0.013)
-0.034
(0.125)
0.100
(0.048)
401

agriculture
share

(4)
0.102
(0.012)
-0.469
(0.102)
-0.136
(0.056)
401

population

Note: Robust standard errors in parentheses.
Distance to canal and distance to Great Lakes harbor measured in thousands of kilometers.
Fractionalization is one minus the the Herfindahl index of ethnic concentration, averaged across occupational categories.
Manufacturing share and agriculture share in 1860 computed using IPUMS.
Population divided by 100,000.

Observations

Distance to Great Lakes harbor

Distance to canal, 1890

Railroad through county, 1860

Dependent variable:

ethnic
fractionalization
(1)

Table 3.3: Transportation Infrastructure and County Characteristics, 1860

95

437

--

--

0.024
(0.019)
--

manufac
share
(2)

437

--

--

2.590
(0.792)
--

pop
(3)

437

--

--

0.045
(0.106)
--

ethnic
frac
(4)

437

--

--

0.072
(0.020)
--

437

--

--

-0.080
(0.033)
--

1990
racial
religious
frac
frac
(5)
(6)
0.003
(0.015)
-0.090
(0.038)
0.136
(0.027)
-0.087
(0.031)
437

agri
share
(7)
0.018
(0.020)
-0.056
(0.053)
0.057
(0.037)
0.030
(0.026)
437

manufac
share
(8)

Note: Robust standard errors in parentheses.
Fractionalization is one minus the Herfindahl index of ethnic concentration, averaged across occupational categories.
Manufacturing share in 1860 and agriculture share in 1860 are computed using IPUMS.
Population is divided by 100,000.
Manufacturing share in 1990 defined as share of civilian labor force employed in manufacturing.
Agriculture share in 1990 is computed using farm population and total population in county.
Ethnic fractionalization in 1990 is one minus the Herfindahl index of (white) ethnic concentration based on 16 ethnicities.
Racial fractionalization in 1990 is one minus the Herfindahl index of racial concentration based on 5 racial groups.
Religious fractionalization in 1990 is one minus the Herfindahl index of religious concentration based on 18 denominations.

437

--

Population, 1860

Observations

--

-0.022
(0.014)
--

agri
share
(1)

Agriculture share, 1860

Manufacturing share, 1860

Fractionalization, 1860

Dependent variable:

Year:

-0.269
(0.854)
1.636
(1.063)
-0.962
(0.746)
11.716
(5.372)
437

pop
(9)

Table 3.4: Fractionalization in 1860 and County Characteristics in 1990

-0.027
(0.113)
0.176
(0.309)
0.049
(0.225)
0.326
(0.110)
437

ethnic
frac
(10)

0.011
(0.017)
0.084
(0.055)
-0.080
(0.030)
0.246
(0.052)
437

racial
frac
(11)

-0.124
(0.034)
-0.257
(0.110)
-0.119
(0.074)
0.124
(0.076)
437

religious
frac
(12)

96

0.143
(0.145)
0.432
(0.469)
-0.042
(0.410)
0.168
(0.160)
437

total expenditure
per capita
(1)
(2)

-0.055
(0.047)
0.065
(0.155)
0.020
(0.106)
-0.167
(0.047)
437

education
share
(3)

1990

-0.011
(0.026)
-0.019
(0.060)
0.047
(0.048)
-0.005
(0.016)
437

health
share
(4)

0.014
(0.004)
0.004
(0.014)
-0.011
(0.009)
0.026
(0.007)
437

police
share

Note: Robust standard errors in parentheses.
Fractionalization is one minus the Herfindahl index of ethnic concentration, averaged across occupational categories.
Manufacturing share in 1860 and agriculture share in 1860 are computed using IPUMS.
Population is divided by 100,000 and population density is measured in thousands per square mile.
Total expenditure per capita is measured in thousands of dollars (1990).
Shares in columns (2) through (6) are computed as fraction of total expenditure.

Observations

Population, 1860

Agriculture share, 1860

Manufacturing share, 1860

Fractionalization, 1860

Dependent variable:

Year:

(5)
0.019
(0.012)
-0.050
(0.030)
-0.0001
(0.026)
-0.046
(0.020)
437

roads
share

Table 3.5: Fractionalization in 1860 and Local Government Expenditure in 1990

0.006
(0.009)
-0.034
(0.029)
-0.025
(0.031)
0.000
(0.010)
437

welfare
share
(6)

97

1.00
0.59

Total

Proportion religious

0.67

1.00

0.11
0.13
0.12
0.27
0.08
0.28
0.30

1.00

0.08
0.29
0.14
0.21
0.06
0.22

0.50

1.00

0.06
0.27
0.22
0.18
0.05
0.22

0.58

1.00

0.06
0.30
0.21
0.19
0.05
0.19

0.54

1.00

0.09
0.33
0.20
0.13
0.03
0.22

church adherents
1972
2000
(5)
(6)

Source: Census of Religious Bodies.
Proportion religious is computed as the number of church seats divided by the population 1860-1890, the number
of church members divided by the population 1890-1952, and the number of church adherents divided by the
population 1972-2000.

0.14
0.11
0.05
0.38
0.14
0.18

church seats
1860
1890
(1)
(2)

Distribution of denominations
Baptist
Catholic
Lutheran
Methodist
Presbyterian
Other

Measure of religious participation:
Census year:

religious participation
church members
1890
1952
(3)
(4)

Table 3.6: Distribution of Denominations and Religious Participation, 1860-2000

98

222

0.205
(0.160)
0.456
(0.359)
0.292
(0.313)
-0.276
(0.135)
222

-0.235
(0.159)
-0.373
(0.360)
-0.244
(0.304)
0.273
(0.126)

share of population raised
in church denomination
-migrant
non-migrant
(1)
(2)

222

0.292
(0.155)
0.637
(0.348)
0.195
(0.321)
-0.314
(0.136)

222

-0.124
(0.161)
-0.537
(0.378)
-0.377
(0.291)
0.198
(0.134)

Migrant denominations are Catholic and Lutheran.
Partipation defined as attend church at least "several times" in year prior to interview date.

222

0.210
(0.112)
0.226
(0.302)
0.153
(0.227)
-0.090
(0.087)

222

-0.116
(0.129)
0.138
(0.370)
-0.088
(0.247)
0.133
(0.089)

share of population participating
in church denomination
1979
2000
migrant
non-migrant
migrant
non-migrant
(3)
(4)
(5)
(6)

Note: Standard errors in parentheses.
Fractionalization is one minus the Herfindahl index of ethnic concentration, averaged across occupational categories.
Manufacturing share in 1860 and agriculture share in 1860 are computed using IPUMS.
Population divided by 100,000.

Observations

Population, 1860

Agriculture share, 1860

Manufacturing share, 1860

Fractionalization, 1860

Year:
Church denomination:

Dependent variable:

Table 3.7: Fractionalization in 1860 and Religious Participation in NLSY79

99

100

Table 3.8: NLSY79 Descriptive Statistics

year
1994
(1)
Professional
Migrated out of county of birth
Employed
Income
White
Female
Age

0.10
(0.01)
0.56
(0.01)
0.81
(0.01)
27.80
(0.51)
0.79
(0.01)
0.50
(0.01)
33.36
(0.05)

2000
(2)
0.10
(0.01)
0.59
(0.01)
0.92
(0.01)
33.06
(0.58)
0.79
(0.01)
0.51
(0.01)
39.38
(0.05)

Note: Standard errors in parentheses.
Professional occupations are relevant codes in the Professional, Technical
and Kindred Workers category.
All variables except income and age are binary.
Income is measured in thousands of dollars (2000).

1122

--

--

-0.163
(0.088)
0.045
(0.019)
-0.023
(0.017)
0.003
(0.004)
--

professional
2000
(2)

1209

-0.253
(0.096)
0.047
(0.020)
-0.064
(0.020)
0.001
(0.004)
-0.213
(0.172)
-0.272
(0.144)
0.018
(0.017)

1994
(3)

Note: Standard errors in parentheses are clustered at the county level.
Fractionalization is one minus the Herfindahl index of ethnic concentration, averaged across occupational categories.
White, female, and age are individual-level characteristics.
Manufacturing share in 1860 and agriculture share in 1860 are computed using IPUMS.
Population divided by 100,000.
Professional is a binary variable indicating whether the individual is employed in a professional occupation.
Professional occupations are relevant codes in the Professional, Technical, and Kindred Workers category.

1209

--

Population, 1860

Observations

--

-0.146
(0.083)
0.031
(0.022)
-0.065
(0.020)
0.0003
(0.004)
--

1994
(1)

Agriculture share, 1860

Manufacturing share, 1860

Age

Female

White

Fractionalization, 1860

Dependent variable:
Year:

Table 3.9: Fractionalization in the County of Birth and Occupational Choice

1122

-0.316
(0.109)
0.059
(0.020)
-0.021
(0.016)
0.003
(0.004)
-0.521
(0.213)
-0.431
(0.160)
0.011
(0.025)

2000
(4)

101

(1)

migrated

-0.388
(0.155)
-0.071
(0.305)
-0.171
(0.311)
0.021
(0.039)
0.232
(0.048)
-0.014
(0.017)
0.013
(0.005)
1598

1994
(2)

-0.414
(0.138)
-0.242
(0.260)
-0.317
(0.222)
0.065
(0.040)
0.284
(0.052)
-0.014
(0.024)
0.015
(0.005)
1437

2000
(3)

employed

0.114
(0.108)
0.078
(0.230)
0.031
(0.164)
-0.015
(0.026)
0.153
(0.024)
-0.155
(0.021)
0.001
(0.004)
1614

1994
(4)

0.037
(0.065)
0.362
(0.136)
0.253
(0.107)
0.013
(0.015)
0.043
(0.021)
-0.059
(0.015)
0.003
(0.004)
1332

2000

Note: Standard errors in parentheses are clustered at the county level.
Fractionalization is one minus the the Herfindahl index of ethnic concentration, averaged across occupational categories.
Fractionalization is measured in the county of birth.
Manufacturing share in 1860 and agriculture share in 1860 are computed using IPUMS.
Population is divided by 100,000.
White, female, and age are individual-level characteristics.
Migrated is a binary variable that indicates whether the individual resides outside the county of birth.
Employed is a binary variable that indicates whether the individual currently holds a job.
Income is measured in thousands of dollars (2000).

Observations

Age

Female

White

Population, 1860

Agriculture share, 1860

Manufacturing share, 1860

Fractionalization, 1860

Dependent variable:
Year:
(5)

income

1.659
(4.804)
-1.195
(10.268)
-8.197
(7.808)
-0.210
(1.041)
5.773
(1.850)
-11.986
(1.122)
0.764
(0.208)
1251

1994

Table 3.10: Fractionalization and Outcomes Related to Occupational Choice

(6)
-11.722
(6.183)
-26.379
(12.231)
-21.016
(9.359)
0.449
(1.691)
4.791
(1.567)
-14.476
(0.988)
0.211
(0.312)
1122

2000

102

1130

--

--

0.088
(0.070)
0.066
(0.019)
-0.031
(0.017)
0.002
(0.004)
--

professional
2000
(2)

1205

-0.053
(0.078)
0.090
(0.015)
-0.062
(0.016)
-0.004
(0.004)
-0.321
(0.136)
-0.183
(0.111)
0.055
(0.016)

1994
(3)

Note: Standard errors in parentheses are clustered at the county level.
Fractionalization is one minus the Herfindahl index of ethnic concentration, averaged across occupational categories.
White, female, and age are individual-level characteristics.
Manufacturing share in 1860 and agriculture share in 1860 are computed using IPUMS.
Population divided by 100,000.
Professional is a binary variable indicating whether the individual is employed in a professional occupation.
Professional occupations are relevant codes in the Professional, Technical, and Kindred Workers category.

1205

--

Population, 1860

Observations

--

0.080
(0.073)
0.073
(0.016)
-0.062
(0.016)
-0.0041
(0.004)
--

1994
(1)

Agriculture share, 1860

Manufacturing share, 1860

Age

Female

White

Fractionalization, 1860

Dependent variable:
Year:

Table 3.11: Fractionalization in the County of Residence and Occupational Choice

1130

0.067
(0.075)
0.063
(0.020)
-0.031
(0.017)
0.002
(0.004)
-0.210
(0.188)
-0.075
(0.111)
-0.001
(0.017)

2000
(4)

103

(1)

-0.535
(7.457)
11.570
(17.828)
-3.015
(14.672)
-2.769
(1.981)
22.608
(1.626)
-2.728
(1.248)
2.880
(0.332)
2187

AFQT score
0.016
(0.100)
0.197
(0.160)
-0.017
(0.143)
-0.052
(0.023)
0.010
(0.024)
0.000
(0.017)
0.006
(0.003)
2023

high school completion
(2)
-0.095
(0.104)
0.003
(0.241)
-0.261
(0.193)
-0.018
(0.023)
0.083
(0.018)
-0.039
(0.020)
0.005
(0.004)
2023

college completion
(3)

non-professional occupation.

0.207
(0.107)
0.306
(0.224)
0.327
(0.229)
0.017
(0.040)
-0.029
(0.048)
0.283
(0.022)
-0.002
(0.005)
1239

non-professional
major
(4)

Note: Standard errors in parentheses are clustered at the county level.
Fractionalization is one minus the Herfindahl index of ethnic concentration, averaged across occupational categories.
Fractionalization is measured in the county of birth.
Manufacturing share in 1860 and agriculture share in 1860 are computed using IPUMS.
Population is divided by 100,000.
White, female, and age are individual-level characteristics.
AFQT is the score on the Armed Forces Qualification Test.
High school completion is a binary variable indicating whether the individual completed high school, including GED.
College completion is a binary variable indicating whether the individual completed a four-year college/university degree.
Non-professional major is a binary variable indicating whether the individual's college major is associated with a

Observations

Age

Female

White

Population, 1860

Agriculture share, 1860

Manufacturing share, 1860

Fractionalization, 1860

Dependent variable:

Table 3.12: Fractionalization and Educational Attainment

104