ESSAYS IN SOCIAL POLICY AND ECONOMIC
WELL-BEING
A Dissertation
Presented to the Faculty of the Graduate School
of Cornell University
in Partial Fulfillment of the Requirements for the Degree of
Doctor of Philosophy
by
Caroline Maura Walker
August 2019
©c 2019 Caroline Maura Walker
ALL RIGHTS RESERVED
ESSAYS IN SOCIAL POLICY AND ECONOMIC WELL-BEING
Caroline Maura Walker, Ph.D.
Cornell University 2019
This is a dissertation in three parts examining the impact of social policies on the economic
well-being of communities in the United States. In particular, I study the consequences
of policies such as prison construction, criminal sentencing, and subsidized housing devel-
opment that disproportionately impact low-income communities.
In the first chapter, I examine the spillover effects of rising male incarceration on
women’s participation in safety net programs designed for families with children. To
estimate this causal effect, I instrument for male incarceration rates using variation in
the treatment of drug-related offenses across state felony court systems and demographic
groups. Results suggest that higher rates of male incarceration lead to fewer women
participating in cash welfare.
The second chapter revisits the regression discontinuity (RD) estimates used in the
literature evaluating the impact of Low Income Housing Tax Credit (LIHTC) subsidized
housing construction on neighborhood characteristics. I find that although global cubic
RD specifications find a significant estimated effect, these results are highly sensitive to
polynomial specification, inclusion of controls, and disappear when estimated nonpara-
metrically.
In the third chapter, I examine two changing features of the Texas criminal justice
system: prison proliferation and sentencing harshness. In particular, I identify the features
of communities chosen as sites for prison construction, finding evidence for prisons being
located in whiter, less educated, though wealthier areas. I also find rising sentencing
harshness is associated with higher rates of white and female incarceration.
BIOGRAPHICAL SKETCH
Caroline Walker was born in Washington, DC and grew up in Chevy Chase, Maryland.
She graduated from the University of California, Santa Cruz in 2013 with a bachelor’s
degree in economics and mathematics. After graduating, she enrolled in the graduate
program in economics at Cornell, with interests in addressing poverty and inequality.
After graduating from Cornell, she plans to return to the DC area to continue research
as an economist at the US Census Bureau.
iii
Dedicated to John.
iv
ACKNOWLEDGEMENTS
I’d like to first thank my committee for their time and encouragement over the years.
Thanks to my advisor, Fran Blau, for your compassion and patience in taking on a
struggling student. Your dedication to research on socially important issues like gender
inequality was one of the reasons I came to Cornell and your mentorship was one of the
main reasons I stayed.
Thanks to Larry Blume for challenging me to justify why my research projects matter,
your generosity with your time, and your willingness to discuss any topic.
Thanks also to Mallika Thomas – it has truly been a pleasure and inspiration to work
with someone in research and in teaching with such a commitment to rigor but rooted in
concern for people’s well-being.
Though not a formal member of my committee, I’d like to thank Ron Ehrenberg for
encouraging me to come to Cornell. Your continued kindness, wisdom, and humor over
the past six years have meant a lot to me.
I’m especially grateful to my peers in Labor Office – Angela Cools, Jorgen Harris, and
Miriam Larson-Koester. You made working in a crammed, windowless closet a pleasure
and there is no doubt I could not have made it through without you.
I’d also like to thank all the members of Cornell Graduate Students United. My time
spent organizing with fellow grads in CGSU taught me far more about labor, unions,
democracy, and collective power than any classroom ever could.
I’m grateful to my family, particularly my parents Richard and Mary, for lending a
sympathetic ear, wisdom, and the requisite cheerleading when needed.
I’m also grateful for the near endless patience of my friends Vera, Tom, Tucker,
Suzanne, Isha, Rhiannon, Andrew, Anne, Charlie, and Rachel, to name a few.
Finally, I’d also like to thank my partner Brian for his love, support, and levelhead-
edness, and my beloved cat Henry for being a constant source of comfort and calm.
v
TABLE OF CONTENTS
Biographical Sketch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii
Dedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii
1 Male Incarceration and Female Welfare Participation 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3.1 Sentencing Policy and the Prison Boom . . . . . . . . . . . . . . . . 8
1.3.2 Safety Net Programs and Welfare Reform . . . . . . . . . . . . . . 9
1.4 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.5 Theoretical Framework and Empirical Strategy . . . . . . . . . . . . . . . 14
1.5.1 Instrumenting for the Male Incarceration Rate . . . . . . . . . . . . 17
1.5.2 Instrument Validity . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.6 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.7 Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.7.1 Female Incarceration . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.7.2 Male Public Assistance Participation . . . . . . . . . . . . . . . . . 23
1.7.3 State of Birth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.8 Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.8.1 Family Formation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.8.2 Labor Supply and Human Capital Accumulation . . . . . . . . . . . 26
1.8.3 Incomplete Take-up . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
1.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
1.10 Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
1.11 Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2 Revisiting Regression Discontinuities and Affordable Housing Supply 47
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.2 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
2.3 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.4 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
2.5 Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
2.6 Validity and First Stage Results . . . . . . . . . . . . . . . . . . . . . . . . 58
2.7 Case Study: Labor Market Outcomes . . . . . . . . . . . . . . . . . . . . . 60
2.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.9 Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
2.10 Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
vi
3 Prison Proliferation and Criminal Sentencing in Texas 76
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.2 Institutional Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.4 Prison Siting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
3.5 Sentencing Variation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
3.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
3.7 Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
3.8 Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
References 98
vii
LIST OF TABLES
1.1 Instrumental Variables First Stage Regression: The Effect of the Share of
Incarcerations that are Drug-related on the Change in Male Incarceration
Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
1.2 The Effect of Male Incarceration on Female Welfare Participation: First
Difference Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
1.3 IV Second Stage Regression: The Effect of a Change in Male Incarceration
on Female Welfare Participation . . . . . . . . . . . . . . . . . . . . . . . 40
1.4 IV Robustness: The Effect of a Change in Male Incarceration on Female
Welfare Participation Controlling for Female Incarceration . . . . . . . . 41
1.5 IV Robustness: The Effect of a Change in Male Incarceration on Female
Welfare Participation Controlling for Male Welfare Participation . . . . . 42
1.6 IV Robustness: The Effect of a Change in Male Incarceration on Female
Welfare Participation Using State of Birth . . . . . . . . . . . . . . . . . . 43
1.7 IV Mechanisms: The Effect of a Change in Male Incarceration on Female
Family Formation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
1.8 IV Mechanisms: The Effect of a Change in Male Incarceration on Female
Labor Supply and Human Capital Accumulation . . . . . . . . . . . . . . 45
1.9 IV Mechanisms: The Effect of a Change in Male Incarceration on Female
Welfare Participation Conditional on Eligibility . . . . . . . . . . . . . . . 46
2.1 First Stage: The Effect of QCT Designation on LIHTC Housing Supply. . 71
2.2 Robustness to Functional Form: The Effect of QCT Eligibility on LIHTC
Housing Supply. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
2.3 Local Polynomial Regression Discontinuity: The Effect of QCT Eligibility
on LIHTC Housing Supply. . . . . . . . . . . . . . . . . . . . . . . . . . . 73
2.4 Reduced Form: The Effect of QCT Designation on Local Labor Market
Conditions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
2.5 Local Polynomial Reduced Form: The Effect of QCT Designation on Local
Labor Market Conditions. . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
3.1 Predictors of 1990s Prison Construction by Security Level: Counties . . . 96
3.2 Predictors of 1990s Prison Construction by Security Level: Tracts . . . . . 97
viii
LIST OF FIGURES
1.1 Male incarceration rate by race/ethnicity, 1980-2013 . . . . . . . . . . . . 30
1.2 Male incarceration rates ages 18-35, 1990. . . . . . . . . . . . . . . . . . . 31
1.3 Violent crime, property crime, and imprisonment rates, 1980-2013. . . . . 32
1.4 Number of States Adopting Habitual Offender Laws . . . . . . . . . . . . 33
1.5 Share of Incarcerations for Drug-related Offenses . . . . . . . . . . . . . . 34
1.6 Share of Eligible Families Participating in AFDC/TANF, 1981-2011 . . . . 35
1.7 Drug Share of Arrests and Incarceration Sentences, by race. . . . . . . . . 36
1.8 Correlation between state policies. . . . . . . . . . . . . . . . . . . . . . . 37
2.1 LIHTC Projects and Low-Income Units Put in Service Over Time. . . . . 65
2.2 Washington, DC data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
2.3 Density Test: Density at the QCT Eligibility Threshold . . . . . . . . . . 67
2.4 Response of Tract Characteristics at the QCT Eligibility Threshold, 1990. 68
2.5 First Stage: Response of Developments at the QCT Eligibility Threshold. 69
2.6 Second Stage: Aggregate Labor Market Response at the QCT Eligibility
Threshold. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.1 Violent and Property Crime Rates in Texas, 1970-2014. . . . . . . . . . . 87
3.2 Public and Private Prisons Constructed in Texas, 1980-2005. . . . . . . . . 88
3.3 Prisons Constructed in Texas, 1991-1999. . . . . . . . . . . . . . . . . . . 89
3.4 New Sentences for Recent Crimes by Year . . . . . . . . . . . . . . . . . . 90
3.5 New Drug Sentences and Drug Share of Sentences by Year . . . . . . . . . 91
3.6 Share of Drug Sentences in Texas, by Race. . . . . . . . . . . . . . . . . . 92
3.7 Share of Sentences in Texas, by Race. . . . . . . . . . . . . . . . . . . . . 93
3.8 Share of Arrests for Drug Related Offenses in Texas, by Race. . . . . . . . 94
3.9 Male Share of Sentences. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
ix
CHAPTER 1
MALE INCARCERATION AND FEMALE WELFARE PARTICIPATION
1.1 Introduction
Over the past fifty years, the United States has experienced unprecedented growth in
incarceration. At its peak in 2008, the stock measure of individuals in prisons and jails
totaled over 2.3 million people – nearly one percent of the population – with incarceration
borne disproportionately by men relative to women and particularly by black men (see
Figure 1.1).1 Incarceration rates also vary substantially geographically, with the highest
levels occurring primarily in the southern states (see Figure 1.2). Despite the widespread
prevalence of incarceration in communities across the United States, relatively little is
known about its impacts on community well-being, and particularly how communities
cope with the removal of such a large fraction of their male population.
While only 17 percent of state prison inmates were married in 2004, 65 percent reported
having children.2 Moreover, 66 percent of male prison inmates reported receiving income
from a job in the months prior to incarceration, suggesting a potential loss in household
resources. Even if a father was not living with his child prior to incarceration, he may have
been making child support payments which cease during an incarceration spell. Together,
these facts suggest potential for direct within-family effect of prison sentencing policies on
households losing the resources of one member. One important consequence of sentencing
policy, then, may manifest through changes in the caseloads of public assistance programs
like cash welfare.
1In 2010, 0.1% of women were incarcerated compared to 1.4% of men according to the Bureau of
Justice Statistics, and approximately 10% of the prison population were female.
2Calculated from the Survey of Inmates in State and Federal Correctional Facilities, 2004.
1
On the other hand, a large fraction of the male population becoming incarcerated may
also have equilibrium effects on the probability of partnerships forming to begin with.
Because women without children are categorically ineligible for cash welfare benefits, if
male incarceration reduces female fertility, this may reduce eligibility and participation
in these public assistance programs for families. In addition, diminishing expectations
among women of forming stable partnerships with men may also affect their labor supply
or human capital investment decisions. If they participate in the labor force at higher rates
or graduate from secondary or post-secondary education at higher rates, this may improve
their labor market opportunities and financial independence, reducing the likelihood that
they participate in cash assistance programs.
Finally, incarceration may also generate fear or mistrust in government agencies, ef-
fectively imposing a transaction cost on engaging with welfare offices.3 The perception
of data sharing between different government agencies may expose a household to direct
scrutiny by social workers, potentially revealing criminal activity of other family members
or friends.4 These effects may also reduce public assistance participation, even conditional
on a household being eligible for benefits.
The implications of higher rates of male incarceration on welfare participation are
important for several reasons. First, if safety net programs are meant to address financial
hardship experienced by children with absent parents, the failure of welfare to aid families
with incarcerated fathers may be a cause for concern. Second, spillovers of sentencing
policy onto other program caseloads are important for the finances of state and local
3The passage of PRWORA in 1996, for example, granted law enforcement agencies the authority
to link felony fugitive files with records of food stamp receipt. This culminated in Operation Talon
which, through coordination between state and local law enforcement and the Department of Agriculture
(USDA), resulted in hundreds of arrests of fugitive felons.
4Some precedent for the impact of policing behavior on the willingness of Hispanic populations to take
up public assistance can be found in Alsan and Yang (2018). The authors demonstrate that expanded
immigration enforcement policies generate fear among Hispanic citizens of exposing non-citizens in their
social networks to law enforcement scrutiny, making them less likely to take up both food stamp and
health insurance benefits.
2
governments. Finally, welfare participation may give an indication of the level of trust
and engagement with government institutions in communities.
In this paper, I explore the effects of rising incarceration rates of men on participation
in the cash welfare programs Aid to Families with Dependent Children (AFDC) and – after
1996 – Temporary Assistance for Needy Families (TANF) among women in those same
communities. Measuring the causal relationship between rising incarceration rates and
public assistance participation is complicated by the fact that many of the same factors
that drive higher rates of incarceration may also generate higher rates of participation
in welfare programs. For example, poverty and poor legal job opportunities within a
community may mean women are more dependent on welfare income to support their
families and men are more likely to choose occupations in the illegal sector (Becker,
1968), producing higher rates of crime and hence higher rates of incarceration.
To address this source of endogeneity and building off the work on marriage markets by
Becker (1974), I follow a recent literature beginning with Charles and Luoh (2010) which
uses variation in the severity of sentences for drug-related offenses by the criminal justice
system as a plausibly exogenous source of variation in male incarceration rates across
marriage markets (defined by age group, race, and state). The 1990s saw the adoption
of a range of policies at federal and state levels which impacted prosecutor and judge
decision-making, leading to an elevated probability that convicted felons were sentenced
to prison or jail, particularly for drug-related offenses. I first document that the increase
in incarceration in the 1990s was driven at least in part by these changes in the treatment
of drug offenses, and that this had differential impacts across geography, race/ethnicity,
and age groups. I then use the treatment of drug-related offenses as an instrument for
male incarceration using two stage least squares.
My main results indicate that higher rates of male incarceration reduced female par-
3
ticipation in cash assistance programs in the 1990s. A one percentage point increase in the
share of incarcerated men within a marriage market led to a 0.6 percentage point decline
in female welfare participation. I find that this result can be explained to a large extent
by reductions in the probability that women have a child in their household who is under
18. This suggests that fewer children were born in communities with higher rates of male
incarceration. I also find an increase in labor supply, which may also reduce women’s eli-
gibility for welfare programs, lowering participation. Both of these results are consistent
with prior work finding increases in labor force participation (Charles and Luoh (2010))
and reductions in teenage fertility among black women (Mechoulan (2011)) in particular.
This paper makes several contributions to the literature. First, I provide the first
causal estimates of the aggregate relationship between male incarceration and female cash
assistance participation. Prior work has primarily looked at the effect of imprisonment
on public assistance take-up of inmates post-release or on the families directly affected by
a parent becoming incarcerated. In this paper, I consider the equilibrium effect of rising
shares of incarcerated men on public assistance participation through its effect on family
formation, labor supply, and take-up conditional on eligibility. Second, I contribute to
the literature on the community-level spillover effects of higher rates of incarceration by
supplying a new potential outcome: cash welfare participation. Finally, I contribute to the
literature on the interactions between different public policy programs with potentially
conflicting goals, particularly the interactions between criminal justice policy and welfare
policy.
The structure of the paper proceeds as follows: Section 2 briefly discusses the liter-
ature on the effects of incarceration and criminal justice policy on ex-prisoners and on
communities. Section 3 gives context on the rise in incarceration and the changes in public
assistance programs that occurred during the sample period. Section 4 describes the data
4
and sample restrictions I use. Section 5 characterizes the theoretical framework linking
male incarceration to public assistance participation, then relates this framework to my
empirical strategies. Section 6 presents the main results, Section 7 discusses robustness
checks to those main results, and Section 8 discusses potential mechanisms of changes in
eligibility and incomplete take-up in explaining those results. Finally, Section 9 discusses
the social welfare implications of these results and identifies some remaining questions.
1.2 Literature Review
At the individual level, most of the causal work on the effects of criminal justice policy
have exploited random assignment of judges with differing degrees of harshness in criminal
cases to identify the effect of criminal sentencing at both the intensive and extensive
margin on the post-release outcomes of the formerly incarcerated.5 While some early
work did not find negative impacts of harsher sentencing (see Grogger (1995) and Kling
(2006)), more recent studies often find worse outcomes for ex-prisoners including increased
recidivism6, decreased employment and earnings, and lower high school completion rates
for juvenile offenders (Mueller-Smith (2015), Aizer and Doyle (2015)). One explanation
for higher rates of recidivism and poorer labor market outcomes in recent studies is that
as sentencing has become more harsh, the marginal offender may have had better legal
options and thus experienced greater consequences post-release. Interestingly, in the
Norwegian context Bhuller, Dahl, Løken, and Mogstad (2016) find prisons successfully
reduce recidivism and improve employment and earnings post-release, however this work
suggests Norwegian prisons may function differently compared to prisons in the United
States, focusing more on rehabilitation.
5Intensive margin meaning the length of time a convicted offender is sentenced to prison or jail, and
extensive margin meaning the probability of receiving an incarceration sentence for a given crime.
6Recidivism is the probability of an ex-prisoner committing a new crime post-release.
5
Community-level analyses of the impact of criminal justice policy have historically
focused primarily on crime as the outcome variable. Studies evaluate the effect of sen-
tencing through two primary channels: (1) deterrence, where harsher sentencing increases
the costs of engaging in criminal activity reducing crime, and (2) incapacitation, where
crime is reduced through the removal of criminal individuals from the community. While
the literature has found some evidence supporting the incapacitation hypothesis7, the
evidence on deterrence is much more mixed. For a full review of deterrence effects of
sentencing, see Chalfin and McCrary (2017).
Recently, there has been an expansion of the literature to include spillover effects of
incarceration to the non-incarcerated, particularly in sociology. Using the Fragile Fami-
lies data and propensity score matching techniques, a range of papers document sugges-
tive evidence of the deleterious effects that paternal incarceration has on child support
payment compliance (Geller, Garfinkel, & Western, 2011), behavioral problems (Geller,
Garfinkel, Cooper, & Mincy, 2009; Wakefield & Wildeman, 2011; Wildeman & Turney,
2014), non-cognitive skills, school readiness, special education placement, and grade re-
tention (Haskins, 2014, 2015; Turney & Haskins, 2014). Cho (2009) examines the impact
of maternal imprisonment on children’s grade retention, finding children with incarcer-
ated mothers are less likely to be held back compared to their matched counterparts.
In Sweden, Dobbie, Grönqvist, Niknami, Palme, and Priks (2018) find that teen preg-
nancy, crime, and welfare receipt of children are all increased and early-life employment
is decreased by parental incarceration using random judge assignment.
Given the difficulty of administratively linking incarcerated men to their partners and
children in the United States, most work on spillover effects of incarceration in economics
have used the incidence of incarceration across different racial and geographic populations
7For example, see Drago and Galbiati (2012), Buonanno and Raphael (2013), and Barbarino and
Mastrobuoni (2014) for the effect of several collective pardons undertaken by the Italian government on
crime.
6
to estimate the effect on communities. Finlay and Neumark (2010) use incarceration as
an instrument for never-married motherhood and conclude that out-of-wedlock childbirth
results in improved outcomes for children, particularly Hispanic children. Charles and
Luoh (2010) estimate how the sex ratio imbalance generated by high male incarceration
rates affects the probability of women to marry, quality of partner, and the distribu-
tion of “gains” to marriage between men and women. Consistent with Becker’s marriage
market model, increasing male incarceration rates results in decreasing probability of fe-
male marriage, “worse quality” partners for women8, and increases in female labor supply
and education. Craigie, Myers Jr., and Darity (2018) further show that the decline in
the supply of marriageable men arising from higher rates of male incarceration increases
never-married female family headship for black women but not white women. As an ex-
pansion of this, Mechoulan (2011) looks at how the sex ratio imbalance generated by high
black male incarceration rates affects young black women’s non-marital teenage fertility,
school attainment and early employment. He finds that fertility drops, high school com-
pletion increases, and employment rates are higher for young black women in communities
with high incarceration rates. DeFina and Hannon (2010) and DeFina and Hannon (2013)
examine the effects of expanding incarceration on child poverty, finding incarceration in-
creases poverty using several different poverty measures.These studies together suggest a
strong case for why male incarceration may impact female use of cash welfare benefits
through impacts on family structure and through dynamic effects on women’s investment
decisions, although the sign of the effect is ambiguous.
To my knowledge, only two empirical papers exist which examine the link between in-
carceration and public assistance in the United States. Mueller-Smith (2015) uses random
courtroom assignment to estimate the impact of incarceration on ex-offender outcomes
post-release, finding that incarceration increased recidivism, decreased labor market op-
8As measured by spouse’s educational attainment.
7
portunities, and increased public assistance dependency as measured by food stamps and
cash welfare. However, this analysis does not include potential spillover effects on non-
incarcerated household members. To that end, Sugie (2012) uses the Fragile Families data
and propensity score matching methods to examine the effect of paternal incarceration on
women’s use of public assistance, finding no effect on cash welfare take-up, but increased
use of food stamps and Medicaid/SCHIP. My work expands on Sugie (2012) by extend-
ing the unit of analysis from within-family effects of paternal incarceration to consider
aggregate effects, which may differ.
1.3 Context
1.3.1 Sentencing Policy and the Prison Boom
Since 1970, a number of legislative changes at both the federal and state level have altered
the punishments that convicted offenders receive. There has been a dramatic increase in
the adoption of policies such as mandatory minimums, truth in sentencing, and habitual
offender laws, all of which both increase the probability that convicted offenders receive
incarceration sentences and increase the length of time spent incarcerated.9 In this paper,
I focus on one aspect of the increasingly punitive criminal justice system in the United
States – the treatment of drug-related offenses.10
At the federal level under the Reagan administration, the Comprehensive Crime Con-
9Mandatory minimums require that offenders serve a predefined term for certain crimes. Truth-in-
sentencing laws aim to reduce early parole release and ensure convicted offenders serve a larger fraction of
their sentences. Habitual offender laws require state courts to impose mandatory and extended sentences
for offenders who commit crimes repeatedly. For example, three-strikes laws require that a person guilty
of committing a violent felony and two other felony convictions must serve a mandatory life sentence in
prison.
10For a more detailed general history of the relationship between sentencing policy and rising incarcer-
ation rates, see Neal and Rick (2014) and Neal and Rick (2016).
8
trol Act of 1984 and the Anti-Drug Abuse Acts of 1986 and 1988 were passed, partially
motivated by increasing crime rates associated with the emergence of crack cocaine. These
laws introduced federal penalties for marijuana cultivation, possession, or trafficking, and
mandatory minimum sentences for possession of crack cocaine11, with even lower posses-
sion thresholds for repeated offenders. These efforts persisted and intensified under the
Clinton administration, which passed the Violent Crime Control and Law Enforcement
Act of 1994, even as violent and property crime rates were falling (see Figure 1.3). This
law allocated funding for prison construction, established additional federal-level habitual
offender laws, implemented drug testing for individuals on probation, and established an
incentive program whereby states would receive federal grants for implementing Truth in
Sentencing laws.
At the state level, a number of changes were made to criminal sentencing as well. For
example, Figure 1.4 shows a large number of states passed habitual offender laws in the
1990s, many of which included provisions requiring mandatory life in prison. Importantly,
these laws increase the probability of being sentenced to prison for drug-related felonies
if the offender has any prior felony convictions. If there are racial differences in the
probability of prior felony convictions, this may be driver of differences in incarceration
growth among different racial groups for the same criminal offense. The National Judicial
Reporting Program data series provides evidence that criminalization of drug offenses was
in effect throughout the sample period of 1990 to 2000. In Figure 1.5 we can see a dramatic
increase in the share of incarcerations that were for drug-related offenses, doubling from
around 20 percent in the first half of the 1990s to over 40 percent by 1996.
11Notably, the law mandated a minimum sentence of five years for possession of 5 grams of crack cocaine
while mandating five years for 500 grams of powder cocaine, generating wide disparities in sentencing for
similar crimes that produced racially disparate effects.
9
1.3.2 Safety Net Programs and Welfare Reform
Aid to Families with Dependent Children (AFDC) was a means-tested cash benefit es-
tablished under the Social Security Act of 1935 as a grant program to states to provide
welfare payments for children deemed needy as a result of their fathers or mothers be-
ing absent from the home, incapacitated, deceased, or unemployed. “Need” and benefits
levels were established at the state level, along with income and resource limits (within
federal guidelines). States were required to provide aid to anyone meeting the federal
eligibility criteria and whose income and resources were below state-imposed limits.
Beginning in the 1990s, under waivers provided by the federal government, states
were able to disregard portions of federal AFDC requirements in order to experiment
with strategies to reduce welfare dependency. This experimentation culminated in the
Personal Responsibility and Work Opportunity Reconciliation Act of 1996 (PRWORA),
which replaced AFDC with a block grant program called Temporary Assistance for Needy
Families (TANF). TANF expanded state discretion in program design, however it imposed
lifetime limits on the time a family with an adult can receive assistance using federal
funds (maximum of five years) and increased work participation rate requirements among
recipients. These changes have been shown to have led to substantial declines in welfare
caseloads, as well as increases in female labor force participation and earnings. For a
comprehensive review of the effects of welfare reform, see Blank (2002).
In conjunction with eligibility declines, take-up rates of cash welfare also declined sub-
stantially (see Figure 1.6). A frequent explanation for this decline has been an increase in
the stigma (or transaction) costs imposed on welfare recipients following welfare reform
and worsening public attitude towards benefit recipients.12 However, an additional ex-
12It may also reflect measurement error in classifying households as eligible, given the difficulty of
observing that a household is reaching the maximum time limit for welfare receipt.
10
planation, and one I begin to explore in this paper, is that increasing incarceration rates
may impact willingness to make use of welfare programs.
1.4 Data
This study makes use of several sources of data. The primary dataset comes from the
IPUMS Decennial Census microdata files for 1970, 1980, 1990, and 2000 combined with
the 5-year American Community Survey (ACS) microdata files for 2009-2013. The ex-
planatory variable of interest is the male incarceration rate. While the Census and ACS
survey individuals living in group quarters, after 1980 they do not distinguish between
kinds of institutionalization – correctional facilities, mental institutions, or institutions
for the elderly, handicapped, and poor. The Census uses a separate category for non-
institutional group quarters which include military barracks, college dormitories, and
rooming houses. By restricting the sample to 18 through 44 year olds, most residents
of nursing homes or institutions for the elderly are excluded.13 The incidence of mental
institutions were also on a decline by 1980, suggesting that by 1990 “institutionalized”
populations in this age group are primarily incarcerated.14 All other variables are defined
based on the non-institutionalized population.
In order to assign a relevant male incarceration rate to particular group of women,
I exploit the race/ethnicity, age group, and geographic clustering of relationships into
“marriage markets” (Charles and Luoh (2010)). I match groups of men to groups of
13According to recent data from the Kaiser Family Foundation, only around 15% of nursing home
residents were under 65 in 2011.
14Between 1970 and 1980, the share of the male institutionalized population between 20 and 44 years
old in mental institutions fell from 21 to 14 percent, a trend which likely persisted as public opinion shifted
more towards community-based care for the mentally ill and developmentally disabled. Meanwhile, the
share of the institutionalized population between ages 20 and 44 in correctional facilities rose from 59
percent to 69 percent between 1970 and 1980.
11
women who are of the same race/ethnicity, a similar age group, and reside in the same
state (for incarcerated men, this means the current state of institutionalization, rather
than state of conviction). The race/ethnicity groups I use are white non-Hispanic, Black
non-Hispanic, and Hispanic. Consistent with Charles and Luoh (2010), I link women ages
18 to 25 to men ages 20 to 27, women ages 26 to 33 to men ages 28 to 35, and women
ages 34 to 41 to men ages 36 to 43. Age ranges do not overlap perfectly because women
typically form relationships with men who are slightly older. Finally, I assign groups of
men to groups of women based on state of current residence – for men, this means state
of institutionalization. In a robustness check, I find that estimates are similar if I instead
use state of birth, given that lower income populations have lower rates of mobility.15
In addition, I impose two sample restrictions worth noting. First, I examine men
and women who have less than a high school diploma, although results are robust to the
inclusion of all men and women. I restrict the sample in this manner because the share
of men who are incarcerated is highly skewed towards less-educated populations, and the
same is true for female welfare recipients. Secondly, I focus on native-born populations.
I do this in order to make results consistent across defining geography by state of birth
rather than state of residence, and also to abstract from separate laws governing the
treatment of immigrant populations.16
To construct the instrument for changes in male incarceration rates, I use sentenc-
ing data come from the National Judicial Reporting Program (NJRP) series, which is
a restricted-use dataset derived from state court and prosecutor offices of 300 counties
in the United States. It includes demographic (age, race/ethnicity, sex) and sentencing
information for individuals convicted of a felony in state courts for even-numbered years
between 1986 and 2006. Sentencing details include worst crime convicted of (allowing me
15Approximately 70% of incarcerated men were incarcerated in their state of birth.
16I hope to examine the impact of more punitive criminal justice policies on immigrant populations,
particularly Hispanic immigrant populations in the future.
12
to distinguish between drug related and non-drug related offenses), sentence length, and
whether they were sentenced to probation, jail, or prison.17
In order to isolate the variation in sentencing across marriage markets that comes
from the behavior of prosecutors and courts, I include two additional sources of data
on crime and arrests. Data on violent and property crime rates at the state-level come
from the Uniform Crime Reporting (UCR) series produced by the Federal Bureau of
Investigations. These data include instances of violent and property crime per 100,000
residents. To supplement these measures, I also use public-use annual Uniform Crime
Reporting Program arrest data which breaks down arrests by age, sex, and race for
reporting law enforcement agencies within states. These allow me to control for within-
state changes in policing or arrest behavior for different crime types so that the variation
I am left with is a primarily a result of decisions made by prosecutors or judges.18
Because my analysis occurs over a period in which there were substantial changes to
welfare programs, it is very important to include controls for welfare policy parameters
that may be correlated with criminal sentencing. To do this, I use AFDC/TANF program
rules as compiled by the Transfer Income Model, version 3 (TRIM3).19 Variables of
interest include the maximum benefit amount for various household compositions, the
need standard for qualifying for welfare payment, and maximum time limit before a
household is no longer eligible for payments. These parameters also allow me to simulate
household eligibility by comparing their reported income in the Census data to the income
required for a family of their type.
17Typically, offenders are sent to jail when the sentence length is less than a year and sent to prison
when the sentence is a year or more.
18The impact of policing intensity on the the propensity of households to participate in public assistance
programs is an interesting question, but outside the scope of this current paper.
19TRIM3 project website, trim3.urban.org, accessed on July 2, 2018.
13
1.5 Theoretical Framework and Empirical Strategy
For simplicity, consider a household consisting of a woman with one or more children,
who has some relationship with the father of the children, married or not.20 An increase
in the male incarceration rate increases the probability that the partner of such a woman
becomes incarcerated. If the male had non-zero income and made transfers to his female
partner, his incarceration would generate a decrease in the woman’s non-labor income. In
a standard static labor supply model, this would result in simultaneous effects on female
labor supply and eligibility for (and thus participation in) public assistance programs.
Depending on the changes in labor supply, it may induce more women to participate in
welfare programs.
In addition to direct monetary costs of a household member becoming incarcerated,
sociological work has documented a lower propensity for those with criminal justice system
interaction to engage with “surveilling” institutions – including hospitals, banks, and wel-
fare offices – for fear of that information being transmitted to local law enforcement.21 In
a Moffit stigma model of welfare program participation, higher rates of male incarceration
would translate to higher stigma – or transaction – costs of public assistance participation,
even holding eligibility fixed, and thus reducing take-up of programs ((Moffitt, 1983)).
At the same time, the skew in sex ratio arising from higher male incarceration rates also
influences the probability of partnerships occurring to begin with. If higher male incarcer-
ation rates reduces female fertility (Mechoulan, 2011), the share of eligible women would
shrink since households without children are categorically ineligible for AFDC/TANF ben-
efits. Similarly, women may alter their expectations of the stability and financial returns
20Welfare offices do not distinguish between married and cohabiting partners in the time period con-
sidered, so long as the male is the father of the qualifying child.
21See Brayne (2014) for more details on what she terms “system avoidance”. See also Alsan and Yang
(2018) for recent work in economics quantifying the effects of immigration enforcement on the willingness
of Hispanic citizens to take up food stamp and health insurance benefits.
14
of future relationships, potentially impacting their decision to participate in the labor
force or invest in human capital if they believe it is unlikely their future male partners
will be able to contribute to the household due to incarceration. Depending on the mag-
nitudes of the effects of each of these factors, the relationship between male incarceration
and female public assistance participation is ambiguous.
Since I am interested in the causal impact of male incarceration on women through its
effects on household income and formation, and in continuity with prior literature, I break
men and women into “marriage markets” which consist of individuals most likely to form
partnerships with one another. The literature on marriage suggests that marriages – and
by extension non-marital partnerships – are likely to form within specific race, age, and
geographical cells (Charles & Luoh, 2010). In what follows, I focus on marriage markets
which match native-born men and women of a particular race (defined as non-Hispanic
white, non-Hispanic black, or Hispanic), residing in a given state, and grouped into three
age cohorts. From this, one can estimate a naive regression of the impact of a change in
male incarceration, ∆IncRate, on the change in female welfare participation, ∆Welf :
∆Welfars = β1∆IncRatears + β2∆Xars + σs + εars (1.1)
∆Xars is a vector of time-varying observable characteristics of marriage markets de-
fined by age group a, race/ethnicity group r, and state of residence s, and σs is a state
fixed effect. The first difference controls for any fixed differences within marriage markets
and σs control for time trends within a state, effectively comparing different race and
age groups within a state. If incarceration rates were allocated randomly across mar-
riage markets, the parameter β1 would capture the causal effect of an increase in male
incarceration on female public assistance participation.
15
However, the causal interpretation of parameter β1 is complicated by the fact that
similar factors influence both the incarceration rate and public assistance participation.
Since the incarceration rate is driven in part by the crime rate, many of the same unob-
served factors which affect the decision to engage in criminal occupations (for example,
poor legal job opportunities) may also affect participation in public assistance. This means
communities with poor labor market opportunities and high poverty may experience both
high crime rates (and incarceration rates) and high public assistance participation shares.
To the extent that the controls inadequately absorb the endogenous economic and social
factors, these omitted variables are likely to bias β1 upwards.
At the same time, public assistance availability might affect the decision to engage in
crime, generating a simultaneity problem. More generous welfare benefits or a higher
proportion of a high-poverty community receiving benefits may reduce economically-
motivated crime.22 Finally, because male incarceration rates are assigned to women based
on assumptions over the age, race, and geographic distribution of relationship formation,
these measures of incarceration rate contain measurement error. For example, although
they are more rare, interracial and age-disparate relationships clearly do form, biasing
estimates of β1 towards zero (attenuation). These factors together suggest that the first
difference estimates are likely to be higher (more positive) than the true population pa-
rameter.
1.5.1 Instrumenting for the Male Incarceration Rate
To understand the process which generates incarceration, suppose there exists a monotonic
ranking of the severity of felony crime on a scale that is common to all communities – for
22For more on the relationship between crime and generosity of public assistance programs, see
Fishback, Johnson, and Kantor (2010), Foley (2011), Liebertz and Bunch (2017), and Agan and Makowsky
(2018).
16
example, homicide is most severe and drug possession is consistently less severe.23 Now
suppose that the community’s criminal justice system can select where on the severity scale
to place the cutoff for prison sentences. More punitive communities will draw that line at a
lower point on the severity scale, sentencing more individuals with minor offenses to prison.
More lenient communities may award probation and send fewer minor offenders to prison.
Thus, differences in sentencing severity will produce differences in incarceration rates.
Further, differences in punitiveness will produce differences in composition of incarcerated
offenders, holding composition of crimes committed fixed. A more punitive community
will have a higher share of minor offenders in the incarcerated population. For the purposes
of this paper, consider these minor offenders to be those convicted of felony drug charges.
Now consider a natural experiment comparing communities which, ceteris paribus,
have set the dividing line between imprisonment and community supervision at different
points on the severity scale. Women in communities with more punitive sentencing policy
will experience an elevated probability of a male partner (or a potential male partner)
becoming incarcerated relative to a community where the probability of male partner
incarceration is low, holding economic opportunities of men, the propensity of men to
commit crime, and eligibility rules for public assistance programs fixed.
Consistent with the literature on spillover effects of male incarceration, this paper
exploits the change in the treatment of drug-related offenders throughout the 1990s to
instrument for changes in male incarceration. Several laws were adopted by states, such
as mandatory minimums and habitual offender laws, which increased the likelihood of
sending individuals to prison for drug crimes. A more punitive state should experience a
larger increase in the share of drug-related sentences in prison. Indeed, referring to Figure
1.5, we can see a sharp increase in the share of incarcerations that were for drug-related
23In the National Corrections Reporting Program data for every year between 1991 and 2015, for
example, the median maximum sentence for non-negligent manslaughter was more than 25 years in
prison, while the median sentence for drug-related offenses were between 2 to 4.9 years.
17
offenses occurring around the same time as a large number of states implemented habitual
offender laws.
Following Charles and Luoh (2010), the final regression model takes the differenced
equation 1.1 and then instruments for the change in male incarceration rate at the mar-
riage market-level between 1990 and 2000 using the average share of incarcerations that
were for drug-related offenses between 1990 and 1998. The result is a first stage of the
following form:
∆IncRatears = α1DrugSharears + α2∆Xars + σs + uars (1.2)
Here, IncRatears is the male incarceration rate, DrugSharears is the percent of incar-
ceration sentences that are for drug-related offenses,24 Xars remains the set of controls for
time-varying marriage market-specific characteristics, and σs are state fixed effects. The
second stage uses the predicted incarceration rates from the first stage to estimate the
relationship between incarceration and public assistance in the first differenced equation
1.1 using two-stage least squares:
∆Welfars = β1∆ÎncRatears + β2∆Xars + σs + εars (1.3)
For this instrument to be valid, the change in the percent of drug-related convictions
must both be a good predictor of a change in incarceration (relevance) and also be un-
correlated with welfare participation through any channel other than increases in male
incarceration (exogeneity).
1 ∑1998 # of drug incarcerations24 ars,tMore specifically, DrugSharears = t=1990 , where the instrument is5 # total incarcerationsars,t
divided by five because the NJRP is reported every two years.
18
1.5.2 Instrument Validity
First, I test the strength of the instrument using an F-test. We can see in Table 1.1 that
the share of incarcerations that were for drug-related offenses has a large and statistically
significant impact on the male incarceration rate. These regressions correspond to Equa-
tion 1.2. In all three regression specifications I use in the remainder of the paper, the
F-statistic is well above the rule of thumb measure of 10 (Stock & Yogo, 2002).
The instrument is defined as the number of felony drug-related incarcerations divided
by the total number of felony incarcerations. Therefore, both increases in the incidence
of drug-related crime and declining incidence of other felony offenses would cause the
instrument to rise and could potentially be correlated with the error term in equation
1.1, violating the exclusion restriction. Since the exclusion restriction cannot be tested
directly when the model is just-identified, I present some suggestive evidence supporting
my claim that the change in drug-related incarceration is driven by changes in criminal
justice policy rather than changes in household behavior.
One concern is that the incidence of non-drug-related crimes are declining due to
improving economic opportunities, rather than an increase in the severity of sentencing
for drug-related offenses. If this were the case, then we would not expect an increase in
the share of drug-related convictions to predict an increase in the incarceration rate (since
other criminal categories, such as violent offenses, typically have a much higher probability
of carrying an incarceration sentence). The first stage regression indicates a strong positive
relationship between the change in convictions that are drug-related and incarceration.
Further, the regression controls for changes in property and violent crime within each
state, ensuring that estimates are the effect of marriage market (particularly state and
race differences) in the treatment of drug offenses rather than changes in behavior.
19
Another concern is that even holding other crimes constant, drug-related crime may
be increasing throughout the 1990s at differential rates across marriage markets in ways
that might be related to women’s labor market opportunities, leading to increases in in-
carceration rates. For this to be true, improved economic opportunities would both need
to reduce welfare participation among women and also increase the probability of men
committing drug-related offenses. To the extent that drug-related crime is economically
motivated, this relationship is likely to go in the opposite direction to my results. Addi-
tionally, if you compare the trends in drug share of arrests to drug share of incarcerations
split by racial group, you can see that although drug-related arrests were increasing for
both white and black populations throughout the 1990s, they trended very similarly (see
Figure 1.7a). Then, when comparing the share of incarcerations that were for drug-related
offenses across racial groups, you see an increase among both black and white populations
between 1994 and 1996, but a much larger increase for black populations (see Figure 1.7b).
This suggests that the racial differences in incarceration likely arose primarily from the
behavior of prosecutors and judges, rather than by decisions of law enforcement. I also
later control for the share of drug-related arrests in my regressions.25
An additional source of concern is that the changes in drug sentencing policy may
be correlated with changes in harshness of welfare reform in the 1990s. For example,
more punitive states may be more likely to increase the severity with which they treat
drug offenses of men and impose harsher eligibility rules for welfare receipt, causing the
observed negative relationship between incarceration and welfare participation. Indeed,
when plotting the date of AFDC waiver against the drug-related incarcerations instru-
ment, we see a positive relationship albeit a weak one (see Figure 1.8a). Additionally,
since high rates of incarceration also mean increased costs of corrections, there may be
25An additional test for adult drug-related criminal activity may be to look at Vital Statistics of drug-
related deaths over this time period to see if overdose deaths are correlated with incarceration rates
among marriage markets. These results forthcoming.
20
crowd-out of expenditures on social programs.26 To the extent that these features of states
are time-varying (and hence not absorbed by the differencing), I also include state fixed
effects in my regressions. This controls for a within-state time trend in unobserved factors
driving incarceration and welfare, such as changing harshness or crowdout effects in state
finances. This means that I am then comparing the differential changes in incarceration
and welfare participation of different demographic groups within-state.
Finally, to believe that these reflect causal estimates, it should be the case that racially-
biased practices in the criminal justice system must not be correlated with racially-biased
administration of cash benefits. To the extent that both the criminal justice system
and welfare offices treat black felons and welfare applicants more punitively relative to
whites, this would bias results in the negative direction. There is no direct way to test
for whether this is the case, however there appears to be more evidence of judicial and
prosecutor discretion in criminal cases relative to the administration of welfare rules,
which have strict family composition and income rules.27 In a robustness check, I control
for the share of non-incarcerated men reporting positive welfare income within a marriage
market in order to control for any trends in participation within-state and race group as
well. The inclusion of these controls also does not appear to alter the coefficient estimates
significantly.
1.6 Main Results
Main results for the first difference regression are reported in Table 1.2. Regressions
are performed at the level of the marriage market, which are defined by age group,
26For a discussion of federal prison overcrowding litigation in the 1970s and 1980s and its impact on
cash welfare spending, see Boylan and Mocan (2014)
27Future work may want to investigate whether there is racial bias in the behavior of welfare offices.
21
race/ethnicity, and state of residence, and are weighted by the population the male incar-
ceration rate was calculated based on. In column (1), I regress the change in female welfare
participation within a marriage market on the change in male incarceration, including a
variety of state-level controls including welfare policy parameters (monthly need standard
for a family of three, maximum monthly benefit for a family of three, welfare lifetime time
limit) and controls for violent and property crime rates (all differenced). These results
suggest that a one percentage point increase in the male incarceration rate is associated
with a reduction in female welfare participation of approximately 0.3 percentage points.
As discussed previously, these results may suffer from both positive bias due to similar
economic factors driving both crime and welfare participation, and attenuation bias aris-
ing due to error in the assignment of marriage markets. As a result, in Table 1.3 I report
the coefficient estimates and robust standard errors for the instrumental variables speci-
fication. In column (1), again I directly include state-level measures of welfare program
parameters and crime rates, finding that a one percentage point increase in male incarcer-
ation results in a 0.5 percentage point decline in female welfare participation off of a base
of 19 percent (this accounts for roughly a 2.6 percent reduction in welfare participation).
In column (2), I include state fixed effects which control for state-level trends in welfare
participation within state that may be correlated with changes in male incarceration.
Then, to isolate the impact of solely the behavior of prosecutors and judges, I include a
control for the share of arrests that are drug-related.28 We can see in column (3) that
coefficient estimates remain at around a 0.6 percentage point reduction in female public
assistance participation in response to an instrumented one percentage point increase in
male incarceration (approximately a 3.2 percent decline).
28The sample size decreases because the arrest data only exists for black and white offenders.
22
1.7 Robustness
1.7.1 Female Incarceration
A first order concern in interpreting these estimates is that the treatment of male drug
offenders by the criminal justice system may be correlated with the treatment of female
drug offenders. Because one of the stipulations of welfare reform in 1996 restricted welfare
payments to people convicted of drug offenses, the decline in welfare participation may
come from declining eligibility of women due to drug convictions rather than because of
higher rates of male incarceration. In order to address this, I include a control on the
right-hand side for female incarceration rate. The inclusion of this control assumes the
female incarceration rate within a marriage market is an appropriate proxy for the number
of women who become ineligible for welfare benefits due to felony convictions. Results
are reported in Table 1.4. We can see that the addition of female incarceration does not
significantly alter coefficient estimates, suggesting these results are not being driven by
the treatment of female offenders.
1.7.2 Male Public Assistance Participation
As discussed previously, if there is racial bias occurring across different marriage markets
in both the criminal justice system and in the welfare offices, the negative coefficient esti-
mate may overstate the negative impact of male incarceration on female public assistance
participation. For example, a cause for concern would be if judges are likely to sentence
black men more harshly and welfare offices are more likely to impose sanctions on black
women requesting welfare benefits in a particular state. One way to address this is by
including a control for the welfare participation rate of non-incarcerated males within a
23
marriage market. Inclusion of this control would effectively control for this bias if the
treatment of black men by welfare offices is similar to their treatment of black women.
On the other hand, changes in male public assistance participation have been shown to be
a consequence of higher rates of public assistance participation.29 Despite the endogeneity
of the inclusion of this measure, we can see in Table 1.5 that it also does not significantly
alter coefficient estimates.
1.7.3 State of Birth
As an additional robustness check, I estimate results defining marriage market by state of
birth rather than current state of residence. This accounts for any migration patterns that
might be correlated with welfare generosity and criminal sentencing. In general, we can
see in Table 1.6 that the results change very little when marriage markets are redefined in
this way. This makes sense since the populations of interest (less educated) have relatively
low rates of migration.
1.8 Mechanisms
In the following section, I consider three potential mechanisms for the negative relationship
between male incarceration rates and female welfare participation rates shown in the
previous section.
29Mueller-Smith (2015) finds that offenders randomly assigned to incarceration by a harsher judge have
increased level of welfare dependency post-release.
24
1.8.1 Family Formation
Prior literature suggests that increases in male incarceration reduces marriage and fertility,
suggesting fewer women may be eligible for welfare simply because they are now choosing
not to have children (or have fewer children) as a result of reduced availability of men.
A female-headed household without children is categorically ineligible for welfare, and a
female-headed household with fewer children will be eligible for a shorter period of time.
To investigate this potential channel, I examine the impact of an increase in male in-
carceration on female fertility. These results are reported in Table 1.7. First, I investigate
the impact of an increase in the male incarceration rate on the probability that a woman
is married but has an absent spouse. If any men who become incarcerated are married,
this estimate should be positive. Indeed, I find that a one percentage point increase in
the share of men who are incarcerated increases the share of women who report an absent
spouse by 0.13 percentage points. This estimate matches survey estimates of the marriage
rates of incarcerated state prisoners which put that figure at around 17 percent.30
Next, I investigate the impact of an increase in male incarceration on the probability
that a woman reports at least one own-child in her household under the age of 18. I choose
the age of 18 because to be categorically eligible for welfare benefits, a household must
contain at least one child 18 or below. Consistent with my predictions and prior estimates
in the literature, I find a large and statistically significant decline in the probability that
a woman has an own child below 18 years old in her household. A one percentage point
increase in the male incarceration rate reduces the probability a woman has a child under
18 by 0.7 percentage points. In results I do not report here, I find that this result is driven
primarily by the younger cohorts of women (ages 18 to 33). I also estimate the effect on the
average number of children women report having, finding a negative coefficient estimate,
30Source: Survey of Inmates in State and Federal Correctional Facilities, 2004.
25
however the standard error is quite large. This relationship also appears to be strongest for
women in the youngest cohort. Together, these results suggest that higher rates of male
incarceration reduce fertility, which may help to explain declining welfare participation.
1.8.2 Labor Supply and Human Capital Accumulation
Next, I examine how female labor supply, both at the intensive and extensive margins,
respond to an increase in male incarceration. Higher rates of male incarceration represent
both a negative income effect to households from which the inmate originated and a shift
in expectations for shared income of single women with future partners. Both of these
should increase women’s participation in the labor force and increase hours of work.
Regression results are displayed in Table 1.8. Consistent with these predictions and
previous literature, I find that increases in male incarceration result in increases in la-
bor force participation and increases in average hours worked last week among women,
although the results are only weakly significant for hours. However, when I restrict the
sample to women in the younger age cohorts (aged 18 through 33), I find that a one per-
centage point increase in the male incarceration results in a 0.49 percentage point increase
in female labor force participation. If this increase in labor force participation generates
household income that pushes a woman out of eligibility for welfare, or makes the returns
to participating too small to outweigh the costs, this may also contribute to the negative
relationship between male incarceration and female welfare participation.
26
1.8.3 Incomplete Take-up
Finally, I present preliminary estimates of the impact of incarceration on the decision of
women to participate in welfare conditional on eligibility. Although a complete simulation
of eligibility would require categorizing women’s eligibility based on family income, for
these initial estimates I focus on the probability of a woman with a potentially eligible child
choosing to participate. This means I include a large fraction of women who have children
under 18 but who have earnings that disqualify them from receiving welfare payments.
Because higher rates of male incarceration affect women’s labor force participation as
shown, this effect may still be capturing some combination of declining eligibility and the
decision to participate conditional on eligibility.31
These results are reported in Table 1.9. The first column reproduces the same results
reported in Table 1.7, where conditional eligibility is simply defined as the share of women
within a marriage market who have an own-child under 18 in their household. Column
(2) estimates the impact of an increase in male incarceration on the probability that a
woman within a marriage market reports positive welfare income, conditional on being
categorically eligible. This means that conditional on having a child under 18 in the
household, a one percentage point increase in the male incarceration rate reduces women’s
probability of reporting positive welfare income by 0.8 percentage points.
1.9 Conclusion
In this paper, I examine the relationship between rising male incarceration rates in the
1990s and female cash welfare participation. Using an instrumental variables approach,
I find that an increase in the share of men in a marriage market who are incarcerated
31The complete simulation results are in progress.
27
by one percentage point reduces the share of women who participate in cash welfare
by approximately half of a percentage point. I provide evidence that the increase in
incarceration is not being driven by behavioral changes, but rather comes from changes
in the treatment of drug-related offenses by the criminal justice system. I also find that
the result is not being driven by trends in female criminal justice system involvement
(rendering them ineligible for cash assistance benefits), nor is it driven by a demographic-
specific trend in welfare participation (as proxied by male cash welfare participation).
This result appears to be driven by reductions in the probability of women having an
own-child under 18 living in their household with them. This suggests that fewer children
were born in communities with higher rates of male incarceration, reducing the share of
eligible women. In addition, I find higher rates of male incarceration also increase labor
force participation of younger women. However, conditional on having a child under 18,
women are still less likely to participate in welfare. This may be either because their
earnings rose, or because conditional on eligibility they are less likely to participate.
The social welfare implications of these results are not self-evident. Although lower
public assistance caseloads may decrease financial burdens to state departments of health
and human services, these estimates do not necessarily translate to the well-being of
communities nor are they likely to be offset by increases in criminal justice system costs.32
Additionally, women who wanted to have children may be made worse off by the decreasing
availability of male partners, and if the increase in labor supply is the result of a negative
income effect either directly or through expectations, she is by definition worse off in
utility terms. To the extent that policymakers value the well-being of low-income women,
higher rates of male incarceration may negatively affect them. Second, this paper measures
aggregate effects of higher rates of male incarceration on women – the within-family causal
32As a crude benchmark, the average cost of incarceration for federal inmates in 2015 was around
$32,000 per year. By contrast, the maximum annual benefit amount for a family of three in Alaska, the
state with the highest maximum monthly benefit, would amount to $923 × 12 = $11, 000.
28
effects of a household member becoming incarcerated remains largely unexplored.
These conflicting effects on well-being suggest that more work is needed to under-
stand the relationship between cash assistance programs, incarceration, and their impact
on communities. In particular, how are families coping with the effects of a family member
becoming incarcerated? The present study could be augmented by examining participa-
tion in public assistance programs besides cash welfare, including food stamps, EITC,
and Medicaid. Additionally, a natural next step would be to administratively link judges’
sentencing severity towards male offenders to records on cash benefit receipt among their
families in order to estimate the direct within-family causal effect of paternal incarceration
in the United States. Finally, policymakers likely care about the well-being of children,
hence outcomes such as children’s performance in school may be of interest.
29
1.10 Figures
Figure 1.1: Male incarceration rate by race/ethnicity, 1980-2013
Note: This graph plots the share of men who are incarcerated (“institutionalized”) between ages 20 and
44, separated by white non-Hispanic, black non-Hispanic, and Hispanic populations. Data come from
IPUMS Decennial Census microdata files.
30
Figure 1.2: Male incarceration rates ages 18-35, 1990.
Note: This figure shows a heatmap of incarceration rates for men between ages 18 and 35 for each state
in the US, calculated using Census data. A legend is forthcoming.
31
Figure 1.3: Violent crime, property crime, and imprisonment rates, 1980-2013.
Note: This graph plots the property and violent crime rates, as well as the imprisonment rates from 1980
to 2013. Crime data comes from FBI Uniform Crime Reporting Program statistics and imprisonment rates
come from various annual Prisoners in the US reports from the Bureau of Justice Statistics. Imprisonment
rates account only for individuals incarcerated in state or federal prisons, but excludes jail populations.
32
Figure 1.4: Number of States Adopting Habitual Offender Laws
Note: This graph plots the number of states in a given year who had adopted a Habitual Offender law.
Data comes from “Impact of State Sentencing Policies on Incarceration Rates in the United States” by
Don Stemen at the Vera Institute (2007).
33
Figure 1.5: Share of Incarcerations for Drug-related Offenses
Note: This graph plots the share of men who received an incarceration sentence for drug-related of-
fenses, including drug trafficking and drug possession. Data comes from restricted-use National Judicial
Reporting Program data files.
34
Figure 1.6: Share of Eligible Families Participating in AFDC/TANF, 1981-2011
Note: Data from this graph comes from Table IND 4a. Number and Percentage of Eligible Families
Participating in the AFDC/TANF Cash Assistance Program: Selected Years included in the “Welfare
Indicators and Risk Factors: Thirteenth Report to Congress”. It was calculated using the Urban Institute
model (TRIM3) that uses CPS data to simulate eligibility.
35
Figure 1.7: Drug Share of Arrests and Incarceration Sentences, by race.
(a) Drug Share of Arrests
(b) Drug Share of Incarcerations
Note: Panel (a) plots the share of arrests that were for drug related offenses for Black and white adults.
Data comes from FBI Uniform Crime Reporting Series: Arrests by Age, Sex, and Race, Summarized
Yearly. Panel (b) shows a similar plot for the share of incarcerations of black and white offenders that
were for drug-related offenses. Data comes from the National Judicial Reporting Program series.
36
Figure 1.8: Correlation between state policies.
(a) Share of Drug Incarcerations vs. Date of AFDC Waiver
(b) Share of Drug Incarcerations vs. Welfare Time Limits
Note: Panel (a) plots the share of incarcerations for drug-related offenses for the sample population
against the year an AFDC waiver was requested. Panel (b) plots the share of incarcerations for drug-
related offenses against the state welfare time limit in 2000. Data come “State Implementation of Major
Changes to Welfare Policies, 1992-1998” published by the U.S. Department of Health and Human Services.
37
1.11 Tables
Table 1.1: Instrumental Variables First Stage Regression: The Effect of the Share of
Incarcerations that are Drug-related on the Change in Male Incarceration Rate
(1) (2) (3)
∆Incarceration rate ∆Incarceration rate ∆Incarceration rate
Drug Share of Incarcerations 0.220*** 0.310*** 0.285***
(0.040) (0.046) (0.059)
Drug Share of Arrests 0.376***
(0.062)
Baseline welfare participation 19.10 19.10 19.11
Initial welfare caseload control yes yes yes
State fixed effect no no yes
Marriage markets 308 308 197
F-statistic 30.80 44.44 23.83
Robust standard errors in parentheses.
*** p < 0.01, ** p < 0.05, * p < 0.1
Note: This table reports the first stage of a first differenced regression of the change in male incarceration
between 1990 and 2000 on the average share of incarcerations for drug-related offenses within a marriage
market. Regressions are performed at the level of the marriage market, which are defined by age group,
race/ethnicity, and state of residence, and are weighted by the population male incarceration rate was
calculated on. Regression 1 includes controls for baseline welfare caseload, maximum monthly benefit
for a family of three, monthly need standard for a family of three, welfare time limit, violent crime rate,
and property crime rate (all differenced). Regressions 2 and 3 include state fixed effects which absorb
the state-level controls included in regression 1. Regression 3 adds an additional control of the share of
arrests that were for drug-related offenses.
38
Table 1.2: The Effect of Male Incarceration on Female Welfare Participation: First Dif-
ference Model
(1) (2) (3)
∆ Welfare Rate ∆ Welfare Rate ∆ Welfare Rate
∆ Male Incarceration Rate -0.278*** -0.350*** -0.294***
(0.086) (0.082) (0.087)
∆ Violent crime rate 0.013 -0.006 -0.857***
(0.038) (0.039) (0.182)
∆ Property crime rate -0.008 -0.008 0.084**
(0.007) (0.007) (0.036)
∆ Monthly need standard 0.000 0.000
(0.000) (0.000)
∆ Maximum monthly benefit 0.000 0.000***
(0.000) (0.000)
∆ Time limits 0.000 0.009***
(0.000) (0.002)
Baseline welfare participation 19.50 19.50 19.50
Initial welfare caseload control yes yes yes
State fixed effect no no yes
Marriage markets 421 421 421
R-squared 0.062 0.076 0.382
Robust standard errors in parentheses.
*** p < 0.01, ** p < 0.05, * p < 0.1
Note: This table reports the a first differenced regression of the change in female welfare participation
between 1990 and 2000 on the change in male incarceration within a marriage market. Regressions are
performed at the level of the marriage market, which are defined by age group, race/ethnicity, and state
of residence, and are weighted by the population male incarceration rate was calculated on. Regression
1 includes controls for violent crime rate and property crime rate. Regressions 2 add welfare program
parameters including maximum monthly benefit for a family of three, monthly need standard for a family
of three, and welfare time limit. Regression 3 adds a state fixed effect to the differenced regression,
effectively adding a within-state time trend.
39
Table 1.3: IV Second Stage Regression: The Effect of a Change in Male Incarceration on
Female Welfare Participation
(1) (2) (3)
∆ Welfare Rate ∆ Welfare Rate ∆ Welfare Rate
∆ Male incarceration rate -0.493** -0.705*** -0.591**
(0.218) (0.161) (0.240)
∆ Drug share of arrests 0.039
(0.122)
Baseline welfare participation 19.10 19.10 19.11
Initial welfare caseload control yes yes yes
State fixed effect no no yes
Observations 308 308 197
R-squared 0.096 0.369 0.435
Robust standard errors in parentheses.
*** p < 0.01, ** p < 0.05, * p < 0.1
Note: This table reports the results of a first differenced regression of the change in female cash assistance
participation rate between 1990 and 2000 on the change in male incarceration, as instrumented by the
average share of incarcerations for drug-related offenses within a marriage market. Regressions are
performed at the level of the marriage market, which are defined by age group, race/ethnicity, and state
of residence, and are weighted by the population male incarceration rate was calculated on. Regression 1
includes controls for baseline welfare caseload, maximum monthly benefit for a family of three, monthly
need standard for a family of three, welfare time limit, violent crime rate, and property crime rate (all
differenced). Regressions 2 and 3 include state fixed effects which absorb the state-level controls included
in regression 1. Regression 3 adds an additional control of the share of arrests that were for drug-related
offenses.
40
Table 1.4: IV Robustness: The Effect of a Change in Male Incarceration on Female
Welfare Participation Controlling for Female Incarceration
(1) (2) (3)
∆ Welfare rate ∆ Welfare rate ∆ Welfare rate
∆ Male incarceration rate -0.500** -0.710*** -0.558**
(0.248) (0.176) (0.245)
∆ Female incarceration rate 0.062 0.042 -0.326
(0.389) (0.281) (0.244)
Drug share of arrests 0.036
(0.125)
Baseline welfare participation 19.10 19.10 19.11
Initial welfare caseload control yes yes yes
State fixed effect no no yes
Marriage markets 308 308 197
Robust standard errors in parentheses.
*** p < 0.01, ** p < 0.05, * p < 0.1
Note: This table reports the results of a first differenced regression of the change in female cash assistance
participation rate between 1990 and 2000 on the change in male incarceration, as instrumented by the
average share of incarcerations for drug-related offenses within a marriage market. It also includes controls
for the change in female incarceration rates within a marriage market. Regressions are performed at the
level of the marriage market, which are defined by age group, race/ethnicity, and state of residence, and
are weighted by the population male incarceration rate was calculated on. Regression 1 includes controls
for baseline welfare caseload, maximum monthly benefit for a family of three, monthly need standard
for a family of three, welfare time limit, violent crime rate, and property crime rate (all differenced).
Regressions 2 and 3 include state fixed effects which absorb the state-level controls included in regression
1. Regression 3 adds an additional control of the share of arrests that were for drug-related offenses.
41
Table 1.5: IV Robustness: The Effect of a Change in Male Incarceration on Female
Welfare Participation Controlling for Male Welfare Participation
(1) (2) (3)
∆ Welfare rate ∆ Welfare rate ∆ Welfare rate
∆ Male incarceration rate -0.506*** -0.641*** -0.605***
(0.184) (0.153) (0.234)
∆ Male welfare rate 0.869*** 0.517*** 0.385**
(0.151) (0.147) (0.190)
Drug Share of Arrests 0.048
(0.118)
Baseline welfare participation 19.10 19.10 19.11
Initial welfare caseload control yes yes yes
State fixed effect no no yes
Marriage markets 306 306 195
Robust standard errors in parentheses.
*** p < 0.01, ** p < 0.05, * p < 0.1
Note: This table reports the results of a first differenced regression of the change in female cash assistance
participation rate between 1990 and 2000 on the change in male incarceration, as instrumented by the
average share of incarcerations for drug-related offenses within a marriage market. It also includes controls
for the change in male welfare participation within a marriage market. Regressions are performed at the
level of the marriage market, which are defined by age group, race/ethnicity, and state of residence, and
are weighted by the population male incarceration rate was calculated on. Regression 1 includes controls
for baseline welfare caseload, maximum monthly benefit for a family of three, monthly need standard
for a family of three, welfare time limit, violent crime rate, and property crime rate (all differenced).
Regressions 2 and 3 include state fixed effects which absorb the state-level controls included in regression
1. Regression 3 adds an additional control of the share of arrests that were for drug-related offenses.
42
Table 1.6: IV Robustness: The Effect of a Change in Male Incarceration on Female
Welfare Participation Using State of Birth
(1) (2) (3)
∆ Welfare Rate ∆ Welfare Rate ∆ Welfare Rate
∆ Male incarceration rate -0.515** -0.693*** -0.651**
(0.182) (0.145) (0.207)
Drug share of arrests 0.001
(0.128)
Baseline welfare participation 18.27 18.27 18.27
Initial welfare caseload control yes yes yes
State fixed effect no no yes
Observations 305 305 195
Robust standard errors in parentheses.
*** p < 0.01, ** p < 0.05, * p < 0.1
Note: This table reports the results of a first differenced regression of the change in female cash assistance
participation rate between 1990 and 2000 on the change in male incarceration, as instrumented by the
average share of incarcerations for drug-related offenses within a marriage market. Regressions are
performed at the level of the marriage market, which are defined by age group, race/ethnicity, and state
of birth, and are weighted by the population male incarceration rate was calculated on. Regression 1
includes controls for baseline welfare caseload, maximum monthly benefit for a family of three, monthly
need standard for a family of three, welfare time limit, violent crime rate, and property crime rate (all
differenced). Regressions 2 and 3 include state fixed effects which absorb the state-level controls included
in regression 1. Regression 3 adds an additional control of the share of arrests that were for drug-related
offenses.
43
Table 1.7: IV Mechanisms: The Effect of a Change in Male Incarceration on Female
Family Formation
∆ Spouse absent ∆ Children under 18 ∆ No. children
∆ Male incarceration rate 0.132*** -0.725*** -0.698
(0.034) (0.199) (0.835)
Baseline dep. var. mean 1.53 13.60 1.22
Initial welfare caseload yes yes yes
State fixed effect yes yes yes
Marriage markets 308 308 308
Robust standard errors in parentheses.
*** p < 0.01, ** p < 0.05, * p < 0.1
Note: This table reports the results of a first differenced regression of the change in various family forma-
tion measures for women between 1990 and 2000 on the change in male incarceration, as instrumented by
the average share of incarcerations for drug-related offenses within a marriage market. Regressions are
performed at the level of the marriage market, which are defined by age group, race/ethnicity, and state
of residence, and are weighted by the population male incarceration rate was calculated on. Regressions
all include a control for state fixed effects which absorb state-level trends in family formation outcomes.
Spouse absent is the share of women in a marriage market reporting an absent spouse. Children under
18 is defined as the share of women who report an own-child under 18 living in the household. Average
number of children women within a marriage market report having in their household.
44
Table 1.8: IV Mechanisms: The Effect of a Change in Male Incarceration on Female
Labor Supply and Human Capital Accumulation
(a) Women Ages 18 to 43
∆ Labor Force ∆ Years ∆ Hours Worked
Participation Schooling Last Week
∆ Male incarceration rate 0.204 0.832 26.848*
(0.196) (1.323) (15.116)
Baseline dep. var. mean 52.34 4.07 34.49
Initial welfare caseload yes yes yes
State fixed effect yes yes yes
Marriage markets 308 308 304
Robust standard errors in parentheses.
*** p < 0.01, ** p < 0.05, * p < 0.1
(b) Women Ages 18 to 34
∆ Labor Force ∆ Years ∆ Hours Worked
Participation Schooling Last Week
∆ Male incarceration rate 0.489*** 1.517 24.651
(0.164) (1.403) (16.981)
Baseline dep. var. mean 50.69 4.17 32.47
Initial welfare caseload yes yes yes
State fixed effect yes yes yes
Marriage markets 207 207 205
Robust standard errors in parentheses.
*** p < 0.01, ** p < 0.05, * p < 0.1
Note: These tables report the results of a first differenced regression of the change in various labor
supply and education measures for women between 1990 and 2000 on the change in male incarceration,
as instrumented by the average share of incarcerations for drug-related offenses within a marriage mar-
ket. Regressions are performed at the level of the marriage market, which are defined by age group,
race/ethnicity, and state of residence, and are weighted by the population male incarceration rate was
calculated on. Panel (a) reports results for all marriage markets, while panel (b) only reports results for
marriage markets for younger women (ages 18 to 34).
45
Table 1.9: IV Mechanisms: The Effect of a Change in Male Incarceration on Female
Welfare Participation Conditional on Eligibility
∆ Categorical Eligibility ∆ Conditional Participation
∆ Male incarceration rate -0.725*** -0.805***
(0.199) (0.364)
Baseline dep. var. mean 16.73 12.40
Initial welfare caseload yes yes
State fixed effect yes yes
Marriage markets 308 246
Robust standard errors in parentheses.
*** p < 0.01, ** p < 0.05, * p < 0.1
Note: This table reports the results of a first differenced regression of the change in categorical eligibility
for welfare and participation conditional on categorical eligibility for women between 1990 and 2000 on
the change in male incarceration, as instrumented by the average share of incarcerations for drug-related
offenses within a marriage market. Regressions are performed at the level of the marriage market, which
are defined by age group, race/ethnicity, and state of residence, and are weighted by the population male
incarceration rate was calculated on. Regressions all include a control for state fixed effects which absorb
state-level trends in family formation outcomes. Categorical eligibility is defined as the share of women
within a marriage market who have an own child under 18 in the household. Conditional participation
is the fraction of women who report positive welfare income, conditional on being categorically eligible.
46
CHAPTER 2
REVISITING REGRESSION DISCONTINUITIES AND AFFORDABLE
HOUSING SUPPLY
2.1 Introduction
Housing affordability is a widespread concern in the United States, with nearly 30 percent
of households paying over 30 percent of their income – the conventional benchmark for
housing affordability – on shelter.1 Moreover, housing affordability difficulties are dispro-
portionately borne by renters, with the share of cost-burdened renters reaching nearly 50
percent in 2016.2 To address affordability concerns, the federal government employs both
demand- and supply-side subsidies to improve the availability and accessibility of quality
housing for low-income households in the rental market.
The Low Income Housing Tax Credit (LIHTC) program is the primary policy lever
the federal government uses to increase the supply of affordable housing for lower income
households, and accounts for one-third of all recent multi-family rental construction (see
Figure 2.1 for trends in number of projects and total number of rent-restricted units put
in service over time). The program provides subsidies to developers for the construction
and rehabilitation of projects with units reserved for low-income residents. A major
challenge local housing agencies face is weighing the need to incentivize development
where there is need while also not concentrating low-income residents in high poverty,
low amenity neighborhoods. To guide the siting decisions of private developers, housing
agencies provide incentives to project proposals meeting certain criteria. For example,
projects located in sufficiently low-income tracts may receive additional tax incentives.
1The State of the Nation’s Housing 2018, Joint Center for Housing Studies of Harvard University.
2ibid.
47
This means that in contrast with demand-side programs like housing choice vouchers
which allow individual renters to choose where they will live, LIHTC projects have the
potential to significantly alter neighborhood quality.
Since LIHTC units are not randomly allocated across geography, a common empirical
approach has been to exploit discontinuities in tax credit allocation formulas to generate
plausibly exogenous variation in the presence of LIHTC projects in neighborhoods. Most
commonly, the literature has used tax credits afforded to projects developed in Qualified
Census Tracts (QCTs), defined by the share of households in the tract who are sufficiently
low-income,3 to estimate a regression discontinuity (RD) design of the impact of LIHTC
development on neighborhood outcomes such as housing prices and crowd-out, income
inequality, and crime (see Baum-Snow and Marion (2009), Freedman and McGavock
(2015), and Freedman and Owens (2011) respectively).
As RDs have grown in popularity in economics due to their simplicity and trans-
parency, so too have the methods for evaluating the validity and robustness of results
obtained from these methods.4 Because understanding the incentive structure of LIHTC
is important in and of itself as affordability crises mount in many major cities and rural
areas, revisiting these relationships using the latest RD techniques is important. If the
incentive effects of certain tax reductions are ineffective in guiding developer’s location
decisions, this suggests housing agencies may wish to pursue additional strategies for in-
creasing the supply of affordable housing. Moreover, if developers would have chosen to
build in neighborhoods even in absence of certain tax abatements, local governments may
be losing a major source of revenue.
3Specifically, 50 percent of households in the tract have incomes below 60 percent of the Area Median
Gross Income, and no more than 20 percent of the population of the metropolitan area lives in one of
those eligible tracts.
4For a discussion of challenges in regression discontinuity and tests of validity, see Lee and Lemieux
(2010).
48
In this paper, I investigate the robustness of the “first stage” effect used in LIHTC
studies – in particular, I examine the strength of the relationship between tax incentives
developers receive for projects in “qualified census tracts” and the number of rental units
built in the 1990s. I find that, when subject to a battery of specifications – both parametric
and nonparametric – the first stage relationship is very sensitive to specification. In
particular, strong global polynomial assumptions, as used in canonical papers such as
Baum-Snow and Marion (2009), produce significant positive estimates of the impact of
QCT designation on development, while local polynomial specifications indicate a much
weaker relationship.
Next, I perform a case study analysis examining the impact of LIHTC units on local
labor market conditions. Unlike HUD-subsidized public housing units or housing choice
vouchers, the rents faced by LIHTC residents are not pegged to income and aren’t sub-
ject to large rent increases like a typical private market housing unit. As a result, the
rent structure does not directly disincentivize work. However, depending on the siting of
LIHTC projects and their proximity to jobs, migration associated with construction of
a project may generate frictional or structural unemployment. Moreover, large housing
projects targeted towards low-income residents may concentrate workers with fewer years
of schooling, increasing labor supply and decreasing wages within a neighborhood. On
the other hand, improving housing quality and stability for residences may increase pro-
ductivity and reduce turnover allowing for worker earnings to rise. LIHTC housing units
also represent a form of place-based economic development, potentially making neighbor-
hoods more attractive to firms. If LIHTC units spur local economic development, this
may further improve labor market outcomes for residents. I show despite the weakness of
the first stage, using parametric RD techniques to estimate second stage effects produce
spurious positive results, easily refuted through graphical illustration. This suggests both
the importance of visual representation of RD methods and of using both parametric and
49
nonparametric techniques.
The remaining structure of this paper is as follows. In Section 2, I outline the literature
on public housing and the LIHTC program in particular to contextualize the contribu-
tions of this investigation. In Section 3, I describe the details of the LIHTC program,
with particular attention paid to the tax credit allocation formulas which generate the
discontinuity in construction that I exploit. In Section 4, I discuss the regression discon-
tinuity design. Section 5 describes the data used. Section 6 discusses first stage results,
Section 7 describes spurious results obtained from global polynomial estimates with an
example, and I conclude with Section 8.
2.2 Literature Review
Most of the literature examining the relationship between housing assistance and labor
supply has examined its effect at the individual level, using random assignment of over-
subscribed housing choice vouchers or ‘traditional’ public housing5 as treatment. Using
variation in random assignment of housing choice vouchers, Jacob and Ludwig (2012),
Mills et al. (2006), and Chyn, Hyman, and Kapustin (2019) all conclude that means-
tested voucher receipt reduces household labor supply and earnings, and increases welfare
participation. In a similar paper, Currie and Yelowitz (1998) examine both the effect of
public housing on labor supply and on participation in other public assistance programs,
finding public housing provision reduces labor force participation and increases partici-
pation in social programs such as Aid to Families with Dependent Children, Medicaid,
Supplemental Security Income, and food stamps. These results suggest that increasing
the availability of subsidized housing units in a community through LIHTC may worsen
5Government owned and operated housing projects.
50
the labor market outcomes of the individuals housed in those units, despite improving the
household’s living conditions (Currie & Yelowitz, 2000).
The literature on publicly subsidized low-income housing is large and broad, covering
a range of policies. Olsen and Barton (1983) develop most clearly a basic theoretical
framework for thinking about how public provision of housing may influence households’
consumption decisions, particularly in contrast to unconditional cash transfers. In the
public housing literature, Currie and Yelowitz (2000) use the Survey of Income and Pro-
gram Participation and an instrumental variables framework to examine what effect living
in a public housing project has on children, finding that project children lived in house-
holds residing in less overcrowded, lower-density complexes, and were less likely to have
been held back in school. Carlson, Haveman, Kaplan, and Wolfe (2012a, 2012b) find that
Section 8 housing vouchers modestly improve neighborhood quality in the long run, neg-
atively affect earnings, and leave employment unchanged. In contrast, Jacob and Ludwig
(2012) use randomized rationing of vouchers in Chicago and find that Section 8 reduces
employment and quarterly earnings and increases TANF receipt. The explanations in
these papers suggest two conflicting mechanisms which may drive the relationship between
publicly-subsidized housing and labor supply: if housing is a complement to leisure, then
public housing may serve to decrease a tenant’s labor supply, whereas if it is a substitute
for leisure (or possibly a complement for work) then it may increase labor supply.
To date, the best identified causal estimates of the impact of neighborhood quality
on family well-being come from the Moving to Opportunity (MTO) experiment, which
randomized housing mobility for low-income families living in five major U.S. cities6 for
1994 to 1998. Families were eligible if they had children and resided in public housing
or Section 8 assisted housing in high-poverty census tracts. MTO studies typically find
minimal effect on adults and older children, but find increased college attendance rates
6Baltimore, Boston, Chicago, Los Angeles, and New York.
51
and earnings for children who were under 13 at the time of the family’s move.7
In the urban economics literature, one often-contentious way federal and state govern-
ments try to improve neighborhood quality is through local place-based economic devel-
opment policies to improve opportunity. The most common place-based policy explored
in the literature are Enterprise Zones. These are designated areas which adopt policies
like lower tax rates or fewer regulations in order to entice private investment and spur
economic development in already-distressed communities. Evidence on the effectiveness of
these zones in propagating job creation are mixed, there is some evidence that it improves
employment marginally at the lower ends of the wage distribution.8
Finally, a handful of papers have been written on the LIHTC program specifically,
all making use of discontinuities in the allocation of tax credits based on census tract
characteristics to identify the effect of subsidized low-income housing development. Baum-
Snow and Marion (2009) find that construction of LIHTC-subsidized developments create
a downward drag on median incomes, however Ellen, Horn, and O’Regan (2016) and
Horn and O’Regan (2011) suggest these units have done little to concentrate poverty.
Freedman and McGavock (2015) show that rental housing subsidized through the LIHTC
program increases income inequality and slightly increases poverty within and between
communities, however this is primarily due to low-income households moving into the
developments rather than driving out more affluent residents. Freedman and Owens
(2011) find that LIHTC developments decrease the incidence of violent crimes at the
county level, but generate no effect on property crime. This paper returns to the first
stage effects of these papers – particularly Baum-Snow and Marion (2009) – to evaluate
their robustness using the latest RD techniques.
7See Sanbonmatsu, Kling, Duncan, and Brooks-Gunn (2006), Aliprantis and Richter (2012), and
Chetty, Hendren, and Katz (2015).
8E.g., see Shroder (2002), Neumark and Kolko (2010), Ham, Swenson, Imrohorǧlu, and Song (2011),
Freedman (2012), and Freedman (2013) for more on the labor market effects of place-based policies.
52
2.3 Background
The Low-Income Housing Tax Credit (LIHTC) program was established as part of the
Tax Reform Act of 1986 to incentivize the development of affordable rental units for low-
and middle-income households. LIHTC accounts for over 46,000 projects consisting of
3.05 million housing units placed in service between 1987 and 2016. As a result, it is
the largest source of federal funding for subsidizing the construction or rehabilitation of
rental units to the affordable housing stock, far outpacing traditional government owned
and operated public housing projects.
Federal tax credits for the LIHTC program are allocated to states by Congress based
on population – for example, in 2007 the allocation was $1.85 per state resident per year
– and are then distributed to projects through an application process. To obtain the
credits, developers submit proposals which are assessed by state housing agencies based
on the state’s Qualified Action Plan, which is a preference ordering for characteristics of
projects. To qualify for consideration, proposals must meet one of two criteria: either at
least 20 percent of units must be rent-restricted9 and occupied by tenants earning less
than 50 percent of the Area Median Gross Income (AMGI), or at least 40 percent of
units occupied must be occupied by tenants earning less than 60 percent of AMGI. The
units reserved for low-income residents are rent restricted, with rents capped at 30% of
the applicable income limit. The tax credit is then calculated as a percentage of costs
incurred during development of low-income units, thus creating an incentive for projects
to be primarily low-income units. Indeed, about 95% of units in most projects are reserved
for low-income tenants.
Local housing agencies then allocate points to proposals based on the state’s Qualified
9Rent-restricted means that gross rent including utilities cannot exceed 30% of the tenant’s imputed
income limit (i.e. 50% or 60% of AMGI, as defined in
53
Action Plan. For example, in New York City, preference is given to projects with a high
percentage of low-income units, housing for vulnerable populations (such as the homeless,
families with children, individuals with mental or physical impairments, persons with
AIDS, and domestic violence survivors), commitment to “green” construction, experience
of the developer, or – importantly for this project – whether the project is located in a
Qualified Census Tract (QCT). Points are tallied based on the number of criteria met,
meaning otherwise similar proposals will be approved by housing agencies if they are in a
QCT. Census tracts are designated QCTs if at least 50% of its households have incomes
below 60% of the AMGI and no more than 20% of the population of a metropolitan area
lives in one of these eligible tracts.
2.4 Data
The data for this paper comes from several different sources. The United States De-
partment of Housing and Human Development (HUD) publicly provides a comprehensive
list of all the LIHTC-subsidized projects placed in service beginning in 1987 through the
present. These data contain characteristics about individual LIHTC projects, including
geocoded project street address, state, county, census tract, and metropolitan statistical
area, as well as year of tax credit allocation, year the project was put into service, number
of housing units, number of low-income units, and QCT status. In total, the LIHTC
database contains 40,502 projects, representing a stock of 2.6 million units total, with
2.25 million reserved for low-income residents.
To illustrate the nature of the data, Figure 2.2 displays a map of the Washington, DC
census tracts, the location of LIHTC subsidized units, and the share of households within
each tract who had an income below 60% of AMGI in 1990. This figure illustrates the
54
sharp economic divide between western, wealthier Washington, DC (indicated by lighter
shades of gray) and lower-income neighborhoods in east and southeast DC, with LIHTC
units largely concentrated in the lower-income tracts. Interestingly, it also showcases how
LIHTC developments are rarely in the highest percentile eligible neighborhoods.
To generate variation across geographic regions in the presence of LIHTC-subsidized
units, this paper exploits quasi-random variation in the location of LIHTC-subsidized
projects generated by a discontinuity in the formula for allocation of tax credits. Census
tracts in which at least 50% of households are below 60% of the Area Median Income
(AMGI) receive additional credits, incentivizing developers to locate projects in these
tracts. In order to exploit this discontinuity, I also use internal data from HUD which
contains the number of eligible households by the income criteria for each census tract
as well as the number of total households, allowing me to assess which tracts should
qualify for QCT status. This data includes data for the years 2000 and 2001, for which
QCT status was determined using 1990 Decennial Census data. Since the 1990 Decennial
Census was used to establish QCT status for 1993 to 2002 with only minor modifications
in the interim period, I use the 2001 QCT designation as a proxy for QCT status between
1993 and 2002.
Baseline tract characteristics and outcome variables come from tract level population
aggregates from the National Historical Geographic Information System (NHGIS) files for
the Decennial Censuses for the years 1990 and 2000. I use the 1990 files to establish base-
line characteristics of each tract, and the 2000 files to measure the impact of LIHTC units
on labor market characteristics. Baseline characteristics and controls include racial/ethnic
composition, immigrant population, urban population, educational attainment, and labor
market characteristics such as labor force participation, unemployment, and earnings.
55
2.5 Empirical Strategy
In the standard RD formulation, the purpose of analysis is to estimate the causal effect
of construction of LIHTC-subsidized rental units has on some neighborhood outcome, y.
To this end, we are interested in estimating β1 in the following regression:
yi = β0 + β1LIunitsi +B
′Xi + i (2.1)
In this regression, i indexes census tracts, LIunitsi is a measure of stock of low-
income units constructed between 1993 and 2001 in tract i, Xi is a vector of initial tract
characteristics which might affect both the number of LIHTC-subsidized units in a tract
and neighborhood outcomes (for example, population controls10), and i is the error term.
The parameter of interest in this model is β1, which describes how an increase in
the stock of LIHTC units in a tract over the period 1993 to 2001 relates to the the
outcome variable. However, for β1 to be estimated consistently, the error term i must be
orthogonal to LIunitsi. This assumption seems unlikely since developers do not choose
tracts in which to develop at random. Since rent restrictions on LIHTC-subsidized units
within a project are lifted after a certain number of years, developers have an incentive
to locate in ‘gentrifying tracts’, where rents are forecasted to be higher in the future, but
which are observably similar to declining or stagnant tracts in the present. Other omitted
variables might include local infrastructure, population density, or migration trends that
aren’t captured by the baseline measures.
As a result of the potential endogeneity between number of subsidized rental units
and changes in neighborhood characteristics, this paper exploits a discontinuity in the
10Population controls include total population, white population share, share of 25 year olds with less
than high school degree, log median household income, and labor force participation.
56
manner tax credits are allocated to projects to generate quasi-random variation in the
location of projects. Developers are provided with additional tax incentives to develop in
tracts where more than 50% of households have an income below 60% of the AMGI, hence
tracts designated as QCTs just above the 50% cutoff will on average be the same as the
tracts just below the cutoff. Because local housing agencies fund projects based on the
number of criteria developers’ proposals meet, a tract that is a QCT but otherwise very
similar to other tracts will have a higher probability of seeing development. A “first-stage”
specification describes the relationship between the eligibility cutoff and the number of
LIHTC-subsidized rental units (or projects) constructed in a tract:
LIunitsi = α0 + α1QCTi + f(eligi) + A
′Xi + ui (2.2)
where eligi is the running variable defined as share of households within a tract in 1990
which meet the income requirement for QCT status, and QCTi is an indicator variable
which takes the value of one if the tract meets the income criteria – that is, eligi ≥ 0.500
– and zero if eligi < 0.500.
11
Substituting equation (2) into equation (1), we arrive at the reduced form relation-
ship between the percent of households eligible by the income criteria and the change in
outcomes of interest, which is given by
yi = γ0 + γ1QCTi + γ2f(eligi) +G
′Xi + ηi (2.3)
Then to uncover the parameter of interest, β1, simply divide γ1 by α1. This estimate
11Note that the QCT designation criteria requires that no more than 20% of the population of an
“area” (as defined by AMGI) reside in a QCT, hence some units may meet the eligibility criteria, but
developers may not experience an additional incentive. That should bias α1 downward, dampening the
measured affect but not biasing the second-stage.
57
gives the change in the outcome yi that results from one additional LIHTC-subsidized
unit.
2.6 Validity and First Stage Results
For identification to be valid, there cannot be any observables that are correlated with
both QCT status and with the neighborhood outcomes of interest conditional on controls.
If there was sorting at the eligibility threshold of these unobservables, the identifying
assumption would be violated. In this context, that would mean households or developers
would have needed perfect foresight of which tracts would receive QCT status based on
the 1990 census – a feat which seems extremely unlikely. Nevertheless, there are two
common ways of empirically testing this assumption in the literature: (1) testing whether
the density of the running variable is continuous through the eligibility threshold, and
(2) testing whether baseline observables are continuous through the threshold (Lee &
Lemieux, 2010).
The first test – the density test – is illustrated in Figure 2.3 where we can see that the
density of the running variable (percent of eligible households) evolves smoothly through
the eligibility cutoff (McCrary, 2008). This corresponds to the intuition that there is
not easy way for any agent involved to manipulate the share of households in a tract
meeting the income criteria. Second, I examine whether baseline tract characteristics
also evolve smoothly through the QCT eligibility threshold. This test can be seen in
Figure 2.4, where there does not appear to be any discontinuity in either demographic or
economic characteristics of tracts through the eligibility threshold. These tests together
provide some evidence that a tract’s QCT status is exogenous conditional on f(eligi).
Conditioning on f(eligi) is necessary since, as can be seen in all panels of Figure 2.4,
58
eligi is clearly correlated with tract characteristics. For identification, there cannot be a
discontinuous jump in the characteristics at the QCT threshold.
Next, I investigate the “first stage” effect. The relationship between the percent of
eligible households in each track and three different measures of LIHTC housing supply
are shown graphically in Figure 2.5 using three different bin sizes. We can see from this
diagram that there does appear to be an increase in the number of units and projects
as the share of eligible households increases, however the visibility of the discontinuity
depends a great deal on choice of binsize. Moreover, inspecting the data more closely, we
can see a cluster of high mean number of LIHTC units just to the left of the cutoff, likely
explaining the lack of relationship we find in local linear regression specifications.12
Regression results are reported in Table 2.1. These regression results correspond to
the parameter estimates for α1 of Equation 2.2. I estimate three separate specifications
for three different measures of housing supply: total LIHTC units, total rent restricted
low-income LIHTC units, and total LIHTC projects. Each regression includes a cubic
polynomial for eligi for each side of the discontinuity, consistent with global polynomial
specifications in prior literature.13
In specification (1), I include no additional controls. I find that at the discontinuity,
QCT designation increases the total number of LIHTC units in the tract by almost 7,
increases the number of low-income units by slightly less at around 5, and a small positive
though statistically insignificant effect on number of projects. These numbers shrink
somewhat with the addition of controls, though remain the same sign. In specification
(2), I add population controls for the base year 1990 including total population, percent
of households residing in urban areas, white non-Hispanic population share, percent of
12This could mean∑in practice there was a rounding rule for QCT status by local housing agencies,although I have found no qualitative evidence of this being the case.
13 3Here, f(elig pi) = p=1 [a1p(eligi − 0.5) + a2pQCTi(eligi − 0.5)p], where p is the degree of polynomial
which is cubic in this case. This specification is the one used in Baum-Snow and Marion (2009).
59
households with single mother, share of persons age 25 or older with less than a high school
diploma, labor force participation, and log median household income. Specification (3)
includes state fixed effects.14
In Table 2.2, I illustrate the sensitivity of first stage estimates to the degree of poly-
nomial and controls used. Because the RD estimator only identifies a causal effect at the
point of discontinuity, estimates that rely on data points very far from the threshold are
likely to produce biased results. We can see that parameter estimates always go in the
positive direction, however the strength and magnitude of that coefficient varies substan-
tially by specification. In particular, inclusion of controls in both the linear and quadratic
specifications (columns 2 and 4 respectively) halves the estimates for total number of
LIHTC units shown in Table 2.1.
More recently, the RD literature has cautioned against using these kinds of higher-
order global polynomial specifications for a number of reasons, instead recommending the
use of local linear or quadratic regressions (for further discussion, see Gelman and Imbens
(2017)). In response to these developments in the literature, I estimate local polynomial
regression discontinuity estimators with robust bias-corrected confidence intervals and
optimal bandwidth selection as developed in Calonico, Cattaneo, and Titiunik (2014).15
Results for these regressions using differing degrees of polynomials are displayed in Ta-
ble 2.3. We can see that the coefficient estimates shrink considerably relative to prior
estimates and we lose all statistical significance, even without the inclusion of controls.
14In future versions, I’d like to see how results change with the inclusion of county fixed effects.
15In practice, these regressions are estimated using the rdrobust command in Stata.
60
2.7 Case Study: Labor Market Outcomes
To illustrate the potential for global polynomial specifications to generate spurious esti-
mates of causal effects, I turn to an example. In this section, I estimate the aggregate
impact of increasing the supply of low-income housing units on labor market outcomes,
including labor force participation, unemployment, and earnings. The reduced form re-
sults using a a cubic polynomial defined separately for each side of the threshold can be
seen in Table 2.4 and reflects the reduced form regression shown in Equation 2.3. Re-
sults indicate that at the eligibility threshold, the availability of additional tax credits are
associated with a 2 percentage point increase in labor force participation, a result that
diminishes somewhat with the inclusion of controls, though remains positive and statis-
tically significant. The unemployment rate in the tract also increases by 0.6 percentage
points, though fades into insignificance when the controls are added. Finally, to examine
a measure of the impact of QCT designation on earnings, I estimate the impact of QCT
eligibility on the share of full-time workers who report earnings below $7,500.16 find a 0.3
percentage point decline in the share of such workers.
However, when estimating the impact of LIHTC units on labor market outcomes using
local polynomial specifications, Table 2.5 shows that coefficient estimates again shrink
and become insignificant. Specifications (1) and (2) use local linear regressions, while
specifications (3) and (4) use local quartic specifications – neither of which produce a
significant result on labor force participation, unemployment, or the share of low-income
residents. Moreover, if we compare these estimates to what can be seen visually in Figure
2.6, it should be no surprise that we find no estimated impact since all three factors appear
to smoothly increase across the threshold. Thus, together, both the local polynomial
regressions and the graphical depictions of the dependent variables of interest confirm the
16A coarse approximation for full-time workers earning below minimum wage in 1990.
61
lack of reliability of the global polynomial estimates used in prior literature.17
2.8 Conclusion
In this paper, I revisit the literature on the impacts of Low Income Housing Tax Credit
financed development on measures of neighborhood quality. In particular, I reexamine
the impact of the availability of additional tax benefits for developers on subsidized unit
construction by exploiting a discontinuity in the tax credit allocation formula. In using
the global polynomial regression discontinuity methods used in prior work, I find evidence
that additional tax incentives do lead to increased construction of affordable housing units
in qualifying census tracts.
However, when using local linear estimation techniques, I find that the increase in
subsidized units constructed as a result of additional tax credits may not be strong enough
to uncover a true causal effect. Moreover, those results are very sensitive to the degree of
global polynomial specified. I also examine a case study of labor market conditions for
an outcome variable, finding significant effects in the polynomial specification despite null
effects both graphically and using local linear regressions. As a result, this work illustrates
the importance of pairing global RD estimates with local linear or quartic regression
estimates. It also suggests how important graphically depicting the relationships are, as a
misspecified functional form may lead to identifying an effect where there clearly is none.
Because evaluating the impact of a major federal program like LIHTC is still impor-
tant, in future work, an additional course of action may be to examine the discontinuity
at a more local level – particularly within a state. Some states may make more explicit
17For completeness, future versions of this work may wish to replicate Baum-Snow and Marion (2009)
directly and plot the outcome variables of interest graphically as well.
62
the criteria for obtaining the tax credits, or may receive excess supply of applications,
meaning the QCT eligibility of a project may significantly increase its chances of being
approved and prioritized by state or local housing authorities.18
Additionally, if a more robust source of variation in LIHTC units is uncovered,19
future work may wish to explore the impact of LIHTC construction on local labor market
outcomes. In particular, it could integrate measures of firm location in order to tease out
the impact of subsidized units on the labor demand side. It may also benefit from merging
in additional county-level labor market data from the Quarterly Census of Employment
and Wages, although this level of aggregation may mean local employment effects would
be difficult to detect.
18Preliminary evidence shows a small number of states – Arizona, Tennessee, and Michigan – have
more substantial discontinuities at the threshold than the United States as a whole.
19For example, LIHTC studies conducted in later time periods which exploit similar though distinct
discontinuities.
63
2.9 Figures
64
Figure 2.1: LIHTC Projects and Low-Income Units Put in Service Over Time.
Note: The top figure plots the number of LIHTC-subsidized projects put in service beginning in 1987
through 2013. The bottom figure plots the number of rent-restricted LIHTC units put in service in the
same period. Data come from the HUD LIHTC project database.
65
Figure 2.2: Washington, DC data.
Note: This figure shows the geographical dispersion of LIHTC subsidized projects across Census tracts in
Washington, DC with tracts shaded according to the share of households within that tract that met the
income eligibility criteria in 1990. Project locations come from HUD LIHTC project database. Eligibility
estimates come from internal HUD 2001 QCT eligibility files, which were calculated using 1990 Decennial
Census data.
66
Figure 2.3: Density Test: Density at the QCT Eligibility Threshold
67
Figure 2.4: Response of Tract Characteristics at the QCT Eligibility Threshold, 1990.
68
Figure 2.5: First Stage: Response of Developments at the QCT Eligibility Threshold.
Note: Each point reflects the mean number of subsidized units, subsidized rent controlled low-income
units, or projects in the relevant bin per Census tract. The top row have a bin size of 0.001, middle row
a bin size of 0.005, and the bottom row has a bin size of 0.01.
69
Figure 2.6: Second Stage: Aggregate Labor Market Response at the QCT Eligibility
Threshold.
Note: Each point reflects the mean number of subsidized units, subsidized rent controlled low-income
units, or projects in the relevant bin per Census tract.
70
2.10 Tables
Table 2.1: First Stage: The Effect of QCT Designation on LIHTC Housing Supply.
(1) (2) (3)
LIHTC units 6.854*** 4.548** 4.602**
(1.888) (1.879) (1.883)
LIHTC low income units 5.403*** 3.413** 3.383**
(1.671) (1.667) (1.673)
LIHTC projects 0.027 0.002 0.006
(0.024) (0.024) (0.024)
Census tracts 59,459 59,449 59,449
Cubic polynomial yes yes yes
Population controls no yes yes
State fixed effects no no yes
Note: Each cell represents a separate regression estimating the parameter α1 in Equation 2.2 which
reflects the effects of QCT designation of a tract at the discontinuity on different measures of low income
housing supply: total number of LIHTC units, total number of rent restricted low-income units, and
total number of projects. Observations are at the tract level. Each specification controls for the QCT
eligibility dummy and the share of eligible households using a cubic polynomial that varies above and
below the QCT threshold. Specification (1) include no controls. Specifications (2) and (3) include total
population, percent of households residing in urban areas, white non-Hispanic population share, percent
of households with single mothers, share of persons age 25 or greater with less than a high school diploma,
labor force participation, and log median household income all measured in 1990 as baseline neighborhood
characteristic controls. Specification (3) includes state fixed effects.
71
Table 2.2: Robustness to Functional Form: The Effect of QCT Eligibility on LIHTC
Housing Supply.
Linear Quadratic Cubic
(1) (2) (3) (4) (5) (6)
LIHTC units 5.842*** 2.650** 4.261*** 2.398 4.810** 6.260***
(1.193) (1.242) (1.628) (1.623) (2.059) (2.052)
LIHTC low income units 4.548*** 1.763 2.751* 1.052 3.768** 5.064***
(1.041) (1.092) (1.426) (1.425) (1.852) (1.847)
LIHTC projects 0.066*** 0.058*** 0.005 -0.003 0.026 0.033
(0.014) (0.014) (0.020) (0.020) (0.026) (0.026)
Census tracts 59,459 59,449 59,449 59,459 59,449 59,449
Degree of polynomial 1 1 2 2 3 3
Population controls no yes no yes no yes
State fixed effects no yes no yes no yes
Note: Each cell represents a separate regression estimating the parameter α1 in Equation 2.3 with different
specifications of f(eligi). Observations are at the tract level. Each specification controls for the QCT
eligibility dummy, the share of eligible households, using a different polynomial that varies above and
below the QCT threshold. Specification (1) include no controls. Specifications in columns (1), (3), and
(5) include no controls and polynomials of degree one, two, and three respectively. Specifications in
columns (2), (4), and (6) add population controls discussed in the text as well as state fixed effects.
72
Table 2.3: Local Polynomial Regression Discontinuity: The Effect of QCT Eligibility on
LIHTC Housing Supply.
(1) (2) (3)
LIHTC units 2.406 1.882 1.037
(2.556) (2.864) (2.947)
LIHTC low income units 2.012 1.388 0.688
(2.193) (2.464) (2.767)
LIHTC projects 0.013 -0.037 -0.064
(0.029) (0.039) (0.043)
Census tracts 59,459 59,459 59,459
Degree of polynomial 1 2 3
Demographic controls no no no
State fixed effects no no no
Note: Each cell represents a separate regression estimating the parameter α1 in Equation 2.3 using
local polynomial regression discontinuity point estimators with robust bias-corrected confidence intervals
as outlined by Calonico et al. (2014). Each regression is performed at the census tract level without
additional covariates and using a uniform kernel. Specifications (1), (2), and (3) use local linear, quadratic,
and cubic polynomial specifications respectively.
73
Table 2.4: Reduced Form: The Effect of QCT Designation on Local Labor Market Con-
ditions.
(1) (2) (3)
Labor force participation 0.020*** 0.005* 0.006**
(0.004) (0.003) (0.003)
Unemployment rate 0.006** 0.003 0.001
(0.003) (0.002) (0.002)
Share full-time workers with low earnings -0.003** -0.002 -0.002
(0.001) (0.001) (0.001)
Census tracts 59,459 59,449 59,449
Cubic polynomial yes yes yes
Population controls no yes yes
State fixed effects no no yes
Note: Each cell represents a separate regression estimating the parameter γ1 in Equation 2.3 which reflects
the effects of QCT designation of a tract at the discontinuity on different aggregate local labor market
outcomes: labor force participation rate, unemployment rate, and the share of full-time workers ages 16
or over with annual earnings below $7,500 in 1999. Each regression is performed at the census tract level
using a uniform kernel. Specifications (1) and (3) include no controls. Specifications (2) and (4) include
total population, percent of households residing in urban areas, white non-Hispanic population share,
percent of households with single mothers, share of persons age 25 or greater with less than a high school
diploma, labor force participation, and log median household income all measured in 1990 as baseline
neighborhood characteristic controls.
74
Table 2.5: Local Polynomial Reduced Form: The Effect of QCT Designation on Local
Labor Market Conditions.
(1) (2) (3) (4)
Labor force participation 0.006 0.004 0.001 0.004
(0.004) (0.003) (0.005) (0.003)
Unemployment rate 0.003 0.004* 0.001 0.002
(0.003) (0.002) (0.004) (0.003)
Share full-time workers with low earnings 0.001 0.000 0.001 -0.000
(0.002) (0.001) (0.002) (0.002)
Census tracts 59,399 59,391 59,399 59,391
Degree of local polynomial 1 1 2 2
Population controls no yes no yes
Note: Each cell represents a separate local polynomial regression estimating the parameter γ1 in Equation
2.3 which reflects the effects of QCT designation of a tract at the discontinuity on different aggregate
local labor market outcomes: labor force participation rate, unemployment rate, and the share of full-time
workers ages 16 or over with annual earnings below $7,500 in 1999. Observations are at the tract level.
Each specification controls for the QCT eligibility dummy and the share of eligible households using a
cubic polynomial that varies above and below the QCT threshold. Specification (1) include no controls.
Specifications (2) and (3) include total population, percent of households residing in urban areas, white
non-Hispanic population share, percent of households with single mothers, share of persons age 25 or
greater with less than a high school diploma, labor force participation, and log median household income
all measured in 1990 as baseline neighborhood characteristic controls. Specification (3) includes state
fixed effects.
75
CHAPTER 3
PRISON PROLIFERATION AND CRIMINAL SENTENCING IN TEXAS
3.1 Introduction
Over the past half century, there have been substantial changes to the American criminal
justice system, with the incarcerated population growing to account for nearly 1 percent
of the total population at its peak in 2008. In this chapter, I study the changes that took
place in the state of Texas between 1990 and 2015 as a case study. The goal is twofold:
first, to examine the siting decisions of where prisons were built to accommodate rising in-
mate populations, and second, to examine the demographic changes in inmate populations
that occurred contemporaneously. Establishing these facts is necessary for understanding
how an institution with deep reach within low-income and minority communities affects
both convicted individuals and their broader communities.
I begin by investigating the geographic proliferation of prisons to identify the charac-
teristics of communities most likely chosen as sites for prisons. Establishing the charac-
teristics of prison siting is important for several reasons. First, less populous communities
with large inmate populations may see their representation in local and state politics shift
through political districting (“prison gerrymandering”) or allocation of funds. Second, sit-
ing decisions have been partially motivated by desire to spur local economic development,
particularly in rural communities where job creation is needed. Despite consensus in the
literature that prisons are an ineffective source of employment and growth, many localities
still bargain extensively for prison construction and against prison demolition.1 Finally,
the locations of prisons dictate the proximity of inmates to their families and commu-
1For several papers discussing the impacts of prison construction on local economies, see Hooks,
Mosher, Rotolo, and Lobao (2004), Carroll (2004), Burayidi and Coulibaly (2009), Hooks, Mosher, Genter,
Rotolo, and Lobao (2010), and Glasmeier and Farrigan (2007).
76
nities. Increasingly rural prisons may decrease urban inmate contact with the outside
world, eroding family bonds and increasing recidivism.2
I find that prisons were often built in counties with higher population levels, higher
minority and foreign born population shares, and were very likely to be built in a county
that already had one prison. At the tract level, these relationships persist, with tracts
with higher white population shares and lower educational attainment often being the
recipient of new facilities in the 1990s. Interestingly, they are higher income relative to
other tracts.
Next, I turn to the changes in criminal sentencing to identify who was most affected
by changes in sentencing policy. Understanding the resulting demographic changes in
prison populations is important because as the marginal inmate shifts, so too may the
spillover effects on non-incarcerated individuals. For example, in 2004, nearly 81 percent
of incarcerated women had children, with 46 percent having lived with those children
prior to incarceration (in contrast to 66 percent and 33 percent for men respectively).3
I find evidence that the composition of sentenced individuals became significantly less
male dominated, falling from nearly 95 percent in 1990 to around 83 percent in 2015. I also
found that the black share of sentenced individuals fell from 50 percent to 30 percent by
2015, whereas the white share rose from around 28 to 38 percent and the Hispanic share
rose from 23 to 32 percent. If there were gender or racial/ethnic biases in sentencing,
one implication of increasing harshness of criminal sentencing may be to narrow those
disparities.4
The remainder of the chapter is organized as follows. In Section 2, I describe the
2For more on the relationship between family visitation and recidivism, see Bales and Mears (2008),
Duwe and Clark (2011), and Mears, Cochran, Siennick, and Bales (2012).
3Author’s calculation using the Survey of Inmates in State and Federal Correctional Facilities, 2004.
4For more on sentencing disparities by gender and race/ethnicity, see Doerner and Demuth (2014) and
Demuth and Steffensmeier (2004) respectively for e.g..
77
Texas criminal justice system in particular and the major crime trends the state has
faced over the past several decades. In Section 3, I describe the data sources used in
this paper. Section 4 discusses the choice of location for prison siting and changes in the
characteristics of communities in which these institutions have been constructed. Section 5
turns to changes in sentencing policy, how these vary geographically and demographically,
as well as how they have changed over time. Finally, in Section 6, I conclude, summarizing
my findings and identifying future lines of work.
3.2 Institutional Background
The Texas Department of Criminal Justice (TDCJ) is the largest state prison system in
the United States, imprisoning over 170,000 people at its peak in 2008 and constituting
10% of the total prison population in the United States.5 The TDCJ is responsible for
administering criminal justice to adult offenders in the state of Texas, including manage-
ment of state prisons, state jails, and private correctional facilities, community supervision
and parole. TDCJ primarily oversees felony offenders, who have committed offenses more
‘severe’ that result in a sentence lasting a year or longer. By contrast, county and munic-
ipal jails hold individuals before trial and house offenders convicted of less severe offenses
such as misdemeanors – whose sentences are typically under a year.
Several factors have been documented in the economics and criminology literature and
popular press as major drivers of prison population growth. The 1980s saw a large increase
in crime rates across the country, and Texas was no exception. Figure 3.1 illustrates
how both violent and property crime rates increased considerably, with property crime
peaking in the mid-to-late 1980s and violent crime peaking a few years later in 1991, before
5Calculated using the Prisoners in 2008 report by the Bureau of Justice Statistics. Note that these
estimates are for state and federal prisoners and do not include county jail populations.
78
plummeting dramatically over the subsequent two decades. The increase in crime rates
in the 1980s and early 1990s has been at least partially attributed to the proliferation
of crack cocaine and other drugs (see Grogger and Willis (2000) and Fryer, Heaton,
Levitt, and Murphy (2013)). However, in addition to rising crime (and sometimes in
response to), changes were made at the federal, state and local level to shift towards
more intensive policing and harsher criminal sentencing policies. All of these factors in
combination contributed to rising incarceration rates in the state of Texas, as in the rest
of the country.
The increase in incarceration rates led to widespread overcrowding and lack of safety in
the Texas penal system in the 1970s and 1980s, culminating in federal court orders in the
Ruiz v. Collins. The case involved inmates suing the state over poor prison conditions,
ultimately culminating in a “Crowding Stipulation” which capped the number of inmates
a facility could hold. To progress towards the termination of federal court oversight as a
result of the case, the state allocated additional funds for the construction of new prison
facilities. The TDCJ undertook an unprecedented expansion in its prison capacity in the
late 1980s and early 1990s. Between 1988 and 1994, the Texas Public Finance Authority
issued an average of nearly $230 million in bonds annually on behalf of the TDCJ for the
design and construction of new prison units, accounting for almost 30 percent of total
new-money bonds issued.6 In addition, the TDCJ repeatedly entered into several lease-
purchase agreements with private nonprofit corporations for private prisons. The product
of these financial investments was a tremendous expansion in prison capacity, with an
average of 9 new prisons being constructed per year between 1988 and 2005, with a peak
of 31 new prisons in 1995 alone (see Figure 3.2).
6Calculated using the Texas Bond Review Board Annual Reports for Fiscal Years 1988 through 1995.
79
3.3 Data
The primary source of data comes from administrative sentencing records provided by the
Texas Department of Criminal Justice, obtained through an open records request.7 For
1990 to 2000, data contain individual-by-month records for all felony offenders under the
jurisdiction of the TDCJ. Observations include demographic characteristics such as gen-
der, race, and date of birth, crime characteristics such as date and worst crime committed,
and sentencing characteristics such as sentence length, prison unit assigned, and whether
an individual was released on parole. For years between 2001 and 2015, observations are
annual, but contain the same characteristics.
In my final sentencing data, I restrict observations only to those with an observed date
of crime (in practice, this primarily drops very old inmates who were likely convicted and
incarcerated much earlier). I also omit individuals for whom no prison was observed. This
removes both inmates who were deceased but still included in the data, and individuals
who were released on parole but were still under the jurisdiction of TDCJ. I also restrict
observations to first year each inmate was observed.8
Next, I combine sentencing data with county-level data on arrests between 1990 and
2014 (with data missing for the year 1993), which come from the Uniform Crime Reporting
Program Data: County-Level Detailed Arrest and Offense Data as provided by the Inter-
university Consortium for Political and Social Research (ICPSR). County-level arrest
estimates are included separately for adults and juveniles, with separate observations for
different kinds of crimes – particularly violent or property felonies, drug-related crimes,
7I am grateful for the funding provided by the Cornell Economics Department Small Grants in Labor
Economics which allowed me access to this data. I am also grateful to Robert Harrelson at the Texas
Department of Criminal Justice for facilitating access and answering a constant barrage of questions
about the data.
8In later versions, I hope to incorporate each unique crime spell for each individual, rather than
inmate.
80
or misdemeanors. These estimates will help distinguish changes in the composition of
incarcerations that arise from changes in sentencing policy as opposed to changes in crime
rates.
These data are paired with records from the 2005 Census of State and Federal Adult
Correctional Facilities, which include state jail and federal and state prison characteristics
including year of construction, location, security level, gender of population, and whether
the institution was privately operated. I supplement these measures with capacity es-
timates obtained through Texas Bond Review Board Annual Reports for Fiscal Years
1988 and 1995, which outline the funding for new prison construction projects and their
contributions to overall state prison capacity. I also add year of closure for facilities that
were shuttered in the later years due to declining prison populations. These estimates
were obtained through manual searches in the local Texas news.
Finally, to understand the characteristics of communities that host prisons, I use
aggregate Decennial Census estimates at the county and tract level provided through
the National Historical Geographic Information System. In particular, these summary
files give me several different measures. They include population characteristics like the
racial/ethnic composition of regions in addition to the share of immigrants – a character-
istic particularly important in a state like Texas where immigration rates from Central
and South America were high. These data also include socioeconomic characteristics such
as median income, poverty rate, and educational attainment.
3.4 Prison Siting
In the 1990s, Texas saw an unprecedented construction effort to expand the stock of
prisons and number of beds. This expansion was particularly notable due to its reach,
81
with new facilities opening not just in higher crime, urban areas but increasingly all
throughout the state (see Figure 3.3). In this section, I examine more closely the location
decisions made in this period, the characteristics of those communities, and explore some
of the implications of these decisions on broader community well-being.
In practice, prison siting in Texas was decided as follows:
A staff site-designation committee evaluates proposals based upon correctional
needs, as determined by [the TDCJ Institutional Division], including logisti-
cal concerns, operational considerations, regional needs, staff recruitment and
retention, site suitability, inmate catchment areas, community resources, and
public support in the community for a prison site. [...] Communities are en-
couraged to be creative in offering incentives in addition to donating land for
a site.9
First, I examine the characteristics of prison system expansion, looking at two different
levels of geography: counties and tracts.10 To understand these relationships, Table 3.1
reports coefficients of a linear probability model of 1990 county-level characteristics on the
probability that a prison was built in the county in the period between 1990 and 2000. At
the base of the table, we can see the sample means and that around 21 percent of counties
saw a new prison built in the 1990s – 6.3 percent a maximum security prison, 13 percent
a medium security prison, and 17.3 percent a low or minimum security prison. Looking
at the first column, we can see that the strongest predictor for prison construction was
a larger population. Point estimates also suggest they were built in counties with higher
minority share, lower foreign born population share with lower log median household
9Report to the 73rd Texas Legislature. Senate Interim Committee on Criminal Justice. November
1992.
10Census tracts, as defined by the U.S. Census Bureau, are small, relatively permanent statistical
subdivisions of a county containing populations between around 1,200 and 8,000.
82
income and slightly higher poverty rates and population shares with less than 9th grade
completed.
There is also some heterogeneity of prison type based on these characteristics – lower
income, higher population, but lower minority population share counties were more likely
to receive minimum security facilities. Community-based re-entry or “halfway house”
facilities are included in the measure of minimum security prisons, which are often located
in more urban areas where residents are in closer proximity to job opportunities. Moreover,
maximum and medium security facilities were very likely to be built in counties that
already had a prison.
Because counties are larger geographic units that may contain a variety of different
neighborhoods, I turn to analyzing these data at the tract level. As shown in Table
3.2, tracts which received a prison had slightly higher population, lower minority and
foreign-born population share, lower labor force participation rates, and worse educational
attainment. Interestingly, log median income of the tracts was higher in these tracts.
This suggests that, at least in Texas, prison siting was not directly trying to address local
economic development issues. Overall, this seems to suggest prisons are being situated in
whiter, less educated regions, but with higher income.
3.5 Sentencing Variation
Next, I turn to the individual sentencing data provided by the TDCJ to analyze how
sentencing practices changed over the sample period. Unfortunately, there are some un-
explained level shifts in the sample populations, particularly beginning in the years 2000
and then 2005 (see Figure 3.4). These changes have been difficult to account for with
changes in policy and may reflect changes in the data collection process rather than
83
sentencing. However, as shown in the previous section, a large number of new prisons
opened in the year 1995, likely explaining the spike in the sentenced population in that
year. Many of these were likely inmates who had previously been held in the county jail
system (not overseen by TDCJ) and thus were newly observed in the data in this year.
A common measure of “harshness” of sentencing in the literature has been to use the
share of incarcerations (or sentences) that were for drug-related offenses (for example, see
Charles and Luoh (2010)). A more lenient community – holding crime rates fixed – may be
less inclined to sentence someone convicted of a drug offense to prison. This suggests that
as communities become harsher in their sentencing, we should see their share of prison
sentences for drug-related offenses going up (again, holding fixed the shares of types of
crime being committed). Moreover, this if harshness differed across demographic groups
(say, more lenient sentencing for female or white defendants), as a community becomes
harsher, it will see the share of those previously advantaged individuals increasing.
Indeed, we can see that over the sample period there is an increase in the share of
inmates convicted of drug-related offenses (see Figure 3.5). In the bottom panel, we see
the share of sentences that were for drug offenses increasing in the second half of the 1990s
and through the beginning of the 2000s. In the top panel, we can see these numbers were
increasing, not simply because sentences for violent offenses were declining. In particular,
we can see that the share of sentences that were for drugs increased most sharply for white
offenders (see Figure 3.6). In addition, the racial composition of sentenced individuals
changes markedly over this period, beginning at almost 50 percent black, dropping to
below 40 percent in 1995, and then declining more or less continuously through to 2015
(see Figure 3.7).
One caution in interpreting the growing share of drug-related white offenders is that
this may reflect the lag in drug dispersion to less urban parts of the state – that is, it may
84
reflect changes in drug use rather than these changes in sentencing, particularly since we
do see increases in drug-related arrests for both black and white individuals (see Figure
3.8). However, as we can see in the figure, the trend in the share of arrests for drug offenses
for both black versus white individuals are quite similar, which is at least suggestive that
the increase in the share of white drug sentences is a result of increasing harshness of
sentencing rather than increases in drug arrests for whites relative to other racial/ethnic
groups.11
Next, I turn to changing gender composition of sentenced individuals. In the top panel
of Figure 3.9, we can see a monotonic decline in the share of sentenced inmates who were
male, falling from a staggering 95 percent in 1990 to about 83 percent in 2005. Like
the changes in racial composition, this shift could be explained either by changing gender
composition of crimes being committed, or by changes in the treatment of female offenders
by the criminal justice system. In the lower panel of Figure 3.9, we can see a similar stark
decline in the male share of arrests – from slightly more than 84 percent in 1990 to below
76 percent in 2014. The decline in the male share of arrests is smaller compared the the
decline in the male share of sentences (12 percentage points vs. 8 percentage points),
which provides some suggestive evidence that higher rates of women were sentenced to
prison conditional on arrest, however the mapping between arrests and sentences may not
be perfect.
3.6 Conclusion
In this chapter, I have documented several features of a unique dataset on criminal sen-
tencing from the Texas Department of Criminal Justice. The goal was twofold. First,
11Unfortunately, the UCR data does not report arrest numbers for Hispanic individuals – they may or
may not be included in the white series.
85
to document the spatial allocation of prisons and state jails that were constructed in the
1990s and early 2000s – in particular, testing quantitatively the anecdotal characteriza-
tions of prisons being increasingly in rural locations. I find some evidence of that, though
the measures of rurality could be improved using population density measures and explicit
urban-suburban-rural classification. A next step would also be to examine the probability
that an inmate is located in a prison outside of their county of arrest.
Second, to explore changes in inmate populations and sentencing that arose concur-
rently and as a result of the expansion of prison capacity in various counties in Texas.
To this end, most notably I find an increasing fraction of inmates who were convicted
of drug offenses, who are female, and who are white. These all suggest a trend towards
harsher punishment for convicted felons, particularly among populations who may have
previously received more lenient sentencing. These changes have different implications for
spillover effects on communities – particularly the higher fraction of female prisoners.
Given that there is less geographic segregation by gender compared to by race and
ethnicity, in future work I would like to exploit geographic differences in prison capacity
to examine the impact of relaxing capacity constraints on arrest and conviction behavior,
particularly of women. I would also like to explore the implications for prison growth
of the changes in the share of sentences for drug-related offenses. Moreover, it would be
good to link the individual-level sentencing data to other sources of administrative data
– particularly ones linked to other family members – in order to examine the spillover
effects of parental incarceration on children’s outcomes.
86
3.7 Figures
Figure 3.1: Violent and Property Crime Rates in Texas, 1970-2014.
Note: Crime rates for Texas are reported per 100,000 population and are obtained from state level crime
estimates from the FBI Uniform Crime Reporting statistics.
87
Figure 3.2: Public and Private Prisons Constructed in Texas, 1980-2005.
Note: “Prisons” includes federal prisons, state prisons, and state jails. Data come from the Census of
State and Federal Adult Correctional Facilities, 2005.
88
Figure 3.3: Prisons Constructed in Texas, 1991-1999.
Note: “Prisons” includes federal prisons, state prisons, and state jails. Data come from the Census of
State and Federal Adult Correctional Facilities 2005 geocoded with to 1990 Census tracts. Each tract
consists of approximately 2,500-8,000 people – darker regions indicate more tracts and hence higher
population density.
89
Figure 3.4: New Sentences for Recent Crimes by Year
Note: This figure plots the number of newly observed felony convictions for a crime committed within
the past 3 years. Data come from sentencing microdata obtained from the Texas Department of Criminal
Justice.
90
Figure 3.5: New Drug Sentences and Drug Share of Sentences by Year
Note: This figure plots the number of newly observed felony drug convictions for a crime committed
within the past 3 years. Data come from sentencing microdata obtained from the Texas Department of
Criminal Justice.
91
Figure 3.6: Share of Drug Sentences in Texas, by Race.
Note: This figure plots the share of new drug sentences that were to white, black, and Hispanic inmates.
Data come from sentencing microdata obtained from the Texas Department of Criminal Justice.
92
Figure 3.7: Share of Sentences in Texas, by Race.
Note: This figure plots the share of new sentences that were to white, black, and Hispanic inmates. Data
come from sentencing microdata obtained from the Texas Department of Criminal Justice.
93
Figure 3.8: Share of Arrests for Drug Related Offenses in Texas, by Race.
Note: Calculated using public-use Uniform Crime Reporting Program Data: Arrests by Age, Sex, and
Race, Summarized Yearly as provided by the Inter-university Consortium for Political and Social Re-
search. Arrest counts are restricted to those in Texas. Data for 2003 are omitted because Harris County,
the most populous county in Texas, did not report arrests.
94
Figure 3.9: Male Share of Sentences.
Note: This figure plots the share of new sentences that were male inmates and the share of adult arrests
that were male. Sentencing data come from sentencing microdata obtained from the Texas Department
of Criminal Justice. Arrest data come from public-use Uniform Crime Reporting Program Data: Arrests
by Age, Sex, and Race, Summarized Yearly as provided by the Inter-university Consortium for Political
and Social Research. Arrest counts are restricted to those in Texas.
95
3.8 Tables
Table 3.1: Predictors of 1990s Prison Construction by Security Level: Counties
Any prison? Max security? Med security? Min security?
Share in group quarters -0.269 1.842** 0.132 -0.495
(0.937) (0.753) (0.658) (1.043)
Prison stock in 1990 0.094 0.094*** 0.163*** 0.116*
(0.061) (0.035) (0.031) (0.07)
Log population 0.097*** 0.013 0.023 0.109***
(0.018) (0.011) (0.014 ) (0.016)
White share -0.854** -0.382* -0.826** 0.693**
(0.365) (0.199) (0.373) (0.272)
Black share -0.845* -0.028 -0.808* 0.23
(0.490) (0.342) (0.449) (0.376)
Foreign born share -1.524*** -0.627** -1.141** 0.738
(0.550) (0.302) (0.541) (0.548 )
Log median income -0.278** -0.089 -0.019 -0.308***
(0.136) (0.094) (0.124 ) (0.114)
Labor force participation 0.366 0.146 -0.366 0.845*
(0.525) (0.349) (0.423) (0.462 )
Share ≥ BA -0.038 -0.403 -0.345 0.163
( 0.532) (0.247) (0.553) (0.573)
Mean of dependent variable 0.205 0.063 0.13 0.173
Observations 254 254 254 254
R-squared 0.183 0.318 0.197 0.272
Robust standard errors in parentheses.
*** p<0.01, ** p<0.05, * p<0.1
Note: Each column represents a separate regression of 1990 county characteristics on the probability
that the county saw any prison/state jail construction between 1990 and 2000 (non-inclusive), and then
separately for each prison security level – maximum security, medium security, and low/minimum security.
Prison construction data come from the Census of State and Federal Adult Correctional Facilities, 2005.
Population estimates come from National Historical Geographic Information System Decennial Census
estimates.
96
Table 3.2: Predictors of 1990s Prison Construction by Security Level: Tracts
Any prison? Max security? Med security? Min security?
Share in group quarters 0.079 0.049* -0.004 0.023
(0.055) (0.026) (0.029) (0.061)
Prison stock in 1990 0.039 0.224*** 0.260*** 0.284***
(0.038) (0.035) (0.037) (0.082)
Log population 0.006** 0.003** 0.002 0.002
(0.003) (0.001) (0.002) (0.003)
White population share -0.029 -0.014 -0.030 0.067**
(0.028) (0.015) (0.027) (0.026)
Black population share -0.053* -0.017 -0.043* 0.050**
(0.028) (0.015) (0.026) (0.022)
Foreign born population share -0.095*** -0.047*** -0.066** 0.064*
(0.032) (0.018) (0.031) (0.033)
Log median income 0.012 0.003 0.003 0.007
(0.009) (0.004) (0.005) (0.011)
Labor force participation rate -0.079*** -0.030*** -0.035** -0.026
(0.025) (0.011) (0.014) (0.025)
Population share ≥ BA -0.053*** -0.007 -0.015 -0.057***
(0.015) (0.008) (0.010) (0.017)
Mean of dependent variable 0.015 0.005 0.010 0.015
Observations 3,959 3,959 3,959 3,959
R-squared 0.024 0.252 0.174 0.147
Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
Note: Each column represents a separate regression of 1990 tract characteristics on the probability
that the tract saw any prison/state jail construction between 1990 and 2000 (non-inclusive), and then
separately for each prison security level – maximum security, medium security, and low/minimum security.
Prison construction data come from the Census of State and Federal Adult Correctional Facilities, 2005.
Population estimates come from National Historical Geographic Information System Decennial Census
estimates.
97
REFERENCES
Agan, A., & Makowsky, M. (2018). The Minimum Wage, EITC, and Criminal Recidi-
vism.
Aizer, A., & Doyle, J. (2015). Juvenile Incarceration, Human Capital, and Future
Crime: Evidence from Randomly Assigned Judges. Quarterly Journal of Economics ,
130 (2), 1–46. doi: 10.3386/w19102
Aliprantis, D., & Richter, F. (2012). Local Average Neighborhood Effects from Moving to
Opportunity.
Alsan, M., & Yang, C. (2018). Fear and the Safety Net: Evidence from Secure Commu-
nities.
Bales, W., & Mears, D. (2008). Inmate Social Ties and the Transition to Society.
Journal of Research in Crime and Delinquency , 45 (3), 287–321. doi: 10.1177/
0022427808317574
Barbarino, A., & Mastrobuoni, G. (2014). The Incapacitation Effect of Incarceration:
Evidence from Several Italian Collective Pardons. American Economic Journal:
Economic Policy , 6 (1), 1–37.
Baum-Snow, N., & Marion, J. (2009). The effects of low income housing tax credit
developments on neighborhoods. Journal of Public Economics , 93 (5-6), 654–666.
doi: 10.1016/j.jpubeco.2009.01.001
Becker, G. (1968). Crime and Punishment: An Economic Approach. Journal of Political
Economy , 76 (2), 169–217. doi: jstor.org/stable/1830482
Becker, G. (1974). A Theory of Marriage. In Economics of the family: Marriage,
children, and human capital (pp. 299–351). Retrieved from http://www.nber.org/
chapters/c2970.pdf doi: 10.2307/2780254
Bhuller, M., Dahl, G., Løken, K., & Mogstad, M. (2016). Incarceration, Recidivism and
Employment.
98
Blank, R. (2002). Evaluating Welfare Reform in the United States. Journal of Economic
Literature, 40 (4), 1105–1166.
Boylan, R., & Mocan, N. (2014). Intended and unintended consequences of prison reform.
Journal of Law, Economics, and Organization, 30 (3), 558–586. doi: 10.1093/jleo/
ewt006
Brayne, S. (2014). Surveillance and System Avoidance: Criminal Justice Contact and
Institutional Attachment. American Sociological Review , 79 (3), 367–391. doi: 10
.1177/0003122414530398
Buonanno, P., & Raphael, S. (2013). Incarceration and incapacitation: Evidence from
the 2006 Italian collective pardon. American Economic Review , 103 (6), 2437–2465.
doi: 10.1257/aer.103.6.2437
Burayidi, M., & Coulibaly, M. (2009). Image Busters: How Prison Location Distorts the
Profiles of Rural Host Communities and What Can Be Done About It. Economic
Development Quarterly , 23 (2), 141–149.
Calonico, S., Cattaneo, M., & Titiunik, R. (2014). Robust Nonparametric Confidence
Intervals for Regression-Discontinuity Designs. Econometrica, 82 (6), 2295–2326.
doi: 10.3982/ECTA11757
Carlson, D., Haveman, R., Kaplan, T., & Wolfe, B. (2012a). Long-term earnings and
employment effects of housing voucher receipt. Journal of Urban Economics , 71 (1),
128–150. Retrieved from http://dx.doi.org/10.1016/j.jue.2011.07.001 doi:
10.1016/j.jue.2011.07.001
Carlson, D., Haveman, R., Kaplan, T., & Wolfe, B. (2012b). Long-term effects of public
low-income housing vouchers on neighborhood quality and household composition.
Journal of Housing Economics , 21 (2), 101–120. Retrieved from http://dx.doi
.org/10.1016/j.jhe.2012.04.004 doi: 10.1016/j.jhe.2012.04.004
Carroll, L. (2004). Prison Siting, Rural Development, Racism, and Justice Reinvestment.
99
Criminology & Public Policy , 3 (3), 481–488.
Chalfin, A., & McCrary, J. (2017). Criminal Deterrence: A Review of the Literature.
Journal of Economic Literature, 55 (1), 5–48. doi: 10.1257/aer.101.5.2157
Charles, K. K., & Luoh, M. C. (2010). Male Incarceration, the Marriage Market, and
Female Outcomes. Review of Economics and Statistics , 92 (3), 614–627.
Chetty, R., Hendren, N., & Katz, L. (2015). The Effects of Exposure to Better Neighbor-
hoods on Children: New Evidence from the Moving to Opportunity Experiment.
Cho, R. M. (2009). Impact of maternal imprisonment on children’s probability of grade
retention. Journal of Urban Economics , 65 (1), 11–23. doi: 10.1016/j.jue.2008.09
.004
Chyn, E., Hyman, J., & Kapustin, M. (2019). Housing Voucher Take-Up and Labor
Market Impacts. Journal of Policy Analysis and Management , 38 (1), 65–98. doi:
10.1002/pam.22104
Craigie, T.-A., Myers Jr., S., & Darity, W. (2018). Racial Differences in the Effect of Mar-
riageable Males on Female Family Headship. Journal of Demographic Economics ,
84 , 231–256. doi: 10.1017/dem.2018.3
Currie, J., & Yelowitz, A. (1998). Public Housing and Labor Supply.
Currie, J., & Yelowitz, A. (2000). Are public housing projects good for kids? Journal of
Public Economics , 75 (1), 99–124. doi: 10.1016/S0047-2727(99)00065-1
DeFina, R., & Hannon, L. (2010). The Impact of Adult Incarceration on Child Poverty:
A County-Level Analysis, 1995-2007. The Prison Journal , 90 (4), 377–396. doi:
10.1177/0032885510382085
DeFina, R., & Hannon, L. (2013). The Impact of Mass Incarceration on Poverty. Crime
and Delinquency , 59 (4), 562–586. doi: 10.1177/0011128708328864
Demuth, S., & Steffensmeier, D. (2004). Ethnicity effects on sentence outcomes in large
urban courts: Comparisons among white, black, and Hispanic defendants. Social
100
Science Quarterly , 85 (4), 994–1011. doi: 10.1111/j.0038-4941.2004.00255.x
Dobbie, W., Grönqvist, H., Niknami, S., Palme, M., & Priks, M. (2018). The Intergen-
erational Effects of Parental Incarceration. Retrieved from http://www.nber.org/
papers/w24186
Doerner, J., & Demuth, S. (2014). Gender and Sentencing in the Federal Courts: Are
Women Treated More Leniently? Criminal Justice Policy Review , 25 (2), 242–269.
doi: 10.1177/0887403412466877
Drago, F., & Galbiati, R. (2012). Indirect effects of a policy altering criminal behavior:
Evidence from the Italian prison experiment. American Economic Journal: Applied
Economics , 4 (2), 199–218. doi: 10.1257/app.4.2.199
Duwe, G., & Clark, V. (2011). Blessed Be the Social Tie That Binds: The Effects of
Prison Visitation on Offender Recidivism Grant. Criminal Justice Policy Review ,
24 (3), 271–296. doi: 10.1177/0887403411429724
Ellen, I., Horn, K., & O’Regan, K. (2016). Poverty concentration and the Low Income
Housing Tax Credit: Effects of siting and tenant composition. Journal of Housing
Economics , 34 , 49–59. Retrieved from http://dx.doi.org/10.1016/j.jhe.2016
.08.001 doi: 10.1016/j.jhe.2016.08.001
Finlay, K., & Neumark, D. (2010). Is Marriage Always Good for Children? Evidence from
Families Affected by Incarceration. Journal of Human Resources , 45 (4), 1046–1088.
Fishback, P., Johnson, R., & Kantor, S. (2010). Striking at the Roots of Crime: The
Impact of Welfare Spending on Crime During the Great Depression. Journal of Law
and Economics , 53 (4), 715–740.
Foley, C. F. (2011). Welfare Payments and Crime. Review of Economics and Statistics ,
93 (1), 97–112.
Freedman, M. (2012). Teaching new markets old tricks: The effects of subsidized invest-
ment on low-income neighborhoods. Journal of Public Economics , 96 (11-12), 1000–
101
1014. Retrieved from http://dx.doi.org/10.1016/j.jpubeco.2012.07.006 doi:
10.1016/j.jpubeco.2012.07.006
Freedman, M. (2013). Targeted Business Incentives and Local Labor Markets. Journal
of Human Resources , 48 (2), 311–344. doi: 10.1353/jhr.2013.0008
Freedman, M., & McGavock, T. (2015). Low-Income Housing Development, Poverty
Concentration, and Neighborhood Inequality. Journal of Policy Analysis and Man-
agement , 34 (4), 805–834. doi: 10.1002/pam
Freedman, M., & Owens, E. (2011). Low-income housing development and crime. Journal
of Urban Economics , 70 (2-3), 115–131. Retrieved from http://dx.doi.org/10
.1016/j.jue.2011.04.001 doi: 10.1016/j.jue.2011.04.001
Fryer, R., Heaton, P., Levitt, S., & Murphy, K. (2013). Measuring crack cocaine and its
impact (Vol. 51) (No. 3). doi: 10.1111/j.1465-7295.2012.00506.x
Geller, A., Garfinkel, I., Cooper, C., & Mincy, R. (2009). Parental incarceration and
child well-being: Implications for urban families. Social Science Quarterly , 90 (5),
1186–1202. doi: 10.1111/j.1540-6237.2009.00653.x
Geller, A., Garfinkel, I., & Western, B. (2011). Paternal incarceration and support for
children in Fragile Families. Demography , 48 (1), 25–47. doi: 10.1007/s13524-010
-0009-9
Gelman, A., & Imbens, G. (2017). Why high-order polynomials should not be used
in regression discontinuity designs. Journal of Business & Economic Statistics ,
0015 (May), 0–0. doi: 10.1080/07350015.2017.1366909
Glasmeier, A., & Farrigan, T. (2007). The economic impacts of the prison development
boom on persistently poor rural places. International Regional Science Review ,
30 (3), 274–299. doi: 10.1177/01660017607301608
Grogger, J. (1995). The effect of arrests on the employment and earnings of young men.
Quarterly Journal of Economics , 110 (1), 51–71.
102
Grogger, J., & Willis, M. (2000). The emergence of crack cocaine and the rise in urban
crime rates. Review of Economics and Statistics , 82 (4), 519–529. doi: 10.1162/
003465300558957
Ham, J., Swenson, C., Imrohorǧlu, A., & Song, H. (2011). Government programs can
improve local labor markets: Evidence from State Enterprise Zones, Federal Em-
powerment Zones and Federal Enterprise Community. Journal of Public Economics ,
95 (7-8), 779–797. doi: 10.1016/j.jpubeco.2010.11.027
Haskins, A. (2014). Unintended Consequences: Effects of Paternal Incarceration on Child
School Readiness and Later Special Education Placement. Sociological Science,
1 (April), 141–158. doi: 10.15195/v1.a11
Haskins, A. (2015). Paternal incarceration and child-reported behavioral functioning at
age 9. Social Science Research, 52 , 18–33. doi: 10.1016/j.ssresearch.2015.01.001
Hooks, G., Mosher, C., Genter, S., Rotolo, T., & Lobao, L. (2010). Revisiting the
impact of prison building on job growth: Education, incarceration, and county-level
employment, 1976-2004. Social Science Quarterly , 91 (1), 228–244. doi: 10.1111/
j.1540-6237.2010.00690.x
Hooks, G., Mosher, C., Rotolo, T., & Lobao, L. (2004). The Prison Industry: Carceral
Expansion and Employment in U.S. Counties, 1969-1994. Social Science Quarterly ,
85 (1), 37–57. doi: 10.1111/j.0038-4941.2004.08501004.x
Horn, K., & O’Regan, K. (2011). The low income housing tax credit and racial segregation.
Housing Policy Debate, 21 (3), 443–473. doi: 10.1080/10511482.2011.591536
Jacob, B., & Ludwig, J. (2012). The effects of housing assistance on labor supply:
Evidence from a voucher lottery. American Economic Review , 102 (1), 272–304.
doi: 10.1257/aer.102.1.272
Kling, J. (2006). Incarceration length, employment, and earnings. American Economic
Review , 96 (3), 863–876. doi: 10.1257/aer.96.3.863
103
Lee, D., & Lemieux, T. (2010). Regression Discontinuity Designs in Economics. Journal
of Economic Literature, 48 (June), 281–355. doi: 10.1257/jel.48.2.281
Liebertz, S., & Bunch, J. (2017). Examining the Externalities of Welfare Reform: TANF
and Crime. Justice Quarterly , 8825 (July), 1–28. Retrieved from https://www
.tandfonline.com/doi/full/10.1080/07418825.2017.1323113 doi: 10.1080/
07418825.2017.1323113
McCrary, J. (2008). Manipulation of the running variable in the regression discontinuity
design: A density test. Journal of Econometrics , 142 (2), 698–714. doi: 10.1016/
j.jeconom.2007.05.005
Mears, D., Cochran, J., Siennick, S., & Bales, W. (2012). Prison Visitation and Recidi-
vism. Justice Quarterly , 29 (6), 888–918. doi: 10.1080/07418825.2011.583932
Mechoulan, S. (2011). The External Effects of Black Male Incarceration on Black Females.
Journal of Labor Economics , 29 (1), 1–35. doi: 10.1086/656370
Mills, G., Gubits, D., Orr, L., Long, D., Feins, J., Kaul, B., & Wood, M. (2006). Effects
of Housing Vouchers on Welfare Families. U.S. Department of Housing and Urban
Development , 173.
Moffitt, R. (1983). An Economic Model of Welfare Stigma. American Economic Review ,
73 (5), 1023–1035. Retrieved from http://www.jstor.org/stable/1814669
Mueller-Smith, M. (2015). The Criminal and Labor Market Impacts of Incarceration.
Retrieved from http://www.columbia.edu/{~}mgm2146/incar.pdf
Neal, D., & Rick, A. (2014). The Prison Boom and the Lack of Black Progress after Smith
and Welch. doi: 10.3386/w20283
Neal, D., & Rick, A. (2016). The Prison Boom and Sentencing Policy. Journal of Legal
Studies , 45 (1), 1–41.
Neumark, D., & Kolko, J. (2010). Do enterprise zones create jobs? Evidence from
California’s enterprise zone program. Journal of Urban Economics , 68 (1), 1–19.
104
Retrieved from http://dx.doi.org/10.1016/j.jue.2010.01.002 doi: 10.1016/
j.jue.2010.01.002
Olsen, E., & Barton, D. (1983). The benefits and costs of public housing in New York City.
Journal of Public Economics , 20 (3), 299–332. doi: 10.1016/0047-2727(83)90028-2
Ruggles, S., Flood, S., Goeken, R., Grover, J., Meyer, E., Pacas, J., & Sobek, M. (n.d.).
IPUMS USA: Version 9.0 . doi: 10.4225/13/511C71F8612C3
Sanbonmatsu, L., Kling, J., Duncan, G., & Brooks-Gunn, J. (2006). Neighborhoods
and Academic Achievement : Results from the Moving to Opportunity Experiment.
Journal of Human Resources , 41 (4), 649–691. doi: 10.3368/jhr.XLI.4.649
Shroder, M. (2002). Locational Constraint, Housing Counseling, and Successful Lease-
up in a Randomized Housing Voucher Experiment. Journal of Urban Economics ,
51 (2), 315–338. doi: 10.1006/juec.2001.2247
Stock, J., & Yogo, M. (2002). Testing for Weak Instruments in Linear IV Regression.
Sugie, N. (2012). Punishment and Welfare: Paternal Incarceration and Families’ Receipt
of Public Assistance. Social Forces , 90 (4), 1403–1427.
Turney, K., & Haskins, A. (2014). Falling Behind? Children’s Early Grade Retention
after Paternal Incarceration. Sociology of Education, 87 (4), 241–258. doi: 10.1177/
0038040714547086
United States Department of Justice. Office of Justice Programs. Bureau of Justice Statis-
tics. (n.d.). National Judicial Reporting Program, 1990-2000. Inter-university Con-
sortium for Political and Social Research. doi: 10.3886/ICPSR06038.v2
Wakefield, S., & Wildeman, C. (2011). Mass imprisonment and racial disparities in
childhood behavioral problems. Criminology & Public Policy , 10 (3), 793–817. doi:
10.1111/j.1745-9133.2011.00740.x
Wildeman, C., & Turney, K. (2014). Positive, Negative, or Null? The Effects of Maternal
Incarceration on Children’s Behavioral Problems. Demography , 51 (3), 1041–1068.
105
doi: 10.1007/s13524-014-0291-z
106