Doyle 2015

JUVENILE INCARCERATION, HUMAN CAPITAL,
AND FUTURE CRIME: EVIDENCE FROM RANDOMLY

ASSIGNED JUDGES*
Downloaded from https://academic.oup.com/qje/article/130/2/759/2330376 by Serials Department user on 09 October 2020

Anna Aizer and Joseph J. Doyle, Jr.
Over 130,000 juveniles are detained in the United States each year with
70,000 in detention on any given day, yet little is known about whether such a
penalty deters future crime or interrupts social and human capital formation in
a way that increases the likelihood of later criminal behavior. This article uses
the incarceration tendency of randomly assigned judges as an instrumental
variable to estimate causal effects of juvenile incarceration on high school com-
pletion and adult recidivism. Estimates based on over 35,000 juvenile offenders
over a 10-year period from a large urban county in the United States suggest
that juvenile incarceration results in substantially lower high school completion
rates and higher adult incarceration rates, including for violent crimes. In an
attempt to understand the large effects, we found that incarceration for this
population could be very disruptive, greatly reducing the likelihood of ever
returning to school and, for those who do return, significantly increasing the
likelihood of being classified as having an emotional or behavioral disorder.
JEL Codes: K140, I210.
I. Introduction
The United States has the highest incarceration rate of any
OECD country—with rates triple that of the next highest country
(Walmsley 2013). The high rate of incarceration in the United
States cannot be explained by higher rates of crime. Since 1990,
U.S. crime rates have fallen each year, while incarceration rates
have doubled to the point where over 2.2 million adults were in-
carcerated and an additional 4.8 million were under supervision
*We thank David Autor, Julie Biehl, Janet Currie, Pedro Dal Bo, Alison
Flaum, Lawrence Grazian, Lawrence Katz, Jens Ludwig, Derek Neal, Steven
Raphael, Roberto Rigobon, Thomas Stoker, Tavneet Suri, Heidi Williams, and
seminar participants at Aarhus University, Harvard University, Institute for
Research on Poverty, Midwest Economics Association, MIT, NBER Childrens/
Labor Studies Summer Institute, Tulane University, and the University of
Maryland. We acknowledge the Chapin Hall Center for Children at the
University of Chicago for the creation of the Integrated Database on Child and
Family Programs in Illinois (IDB) that was used in this study. All findings, inter-
pretations and conclusions based on the use of the IDB are solely our responsibil-
ity and do not necessarily represent the views of the Chapin Hall Center for
Children.
! The Author(s) 2015. Published by Oxford University Press, on behalf of President
and Fellows of Harvard College. All rights reserved. For Permissions, please email:
[email protected]
The Quarterly Journal of Economics (2015), 759–804. doi:10.1093/qje/qjv003.
Advance Access publication on February 2, 2015.
759
760 QUARTERLY JOURNAL OF ECONOMICS
of correctional systems in 2011 (Glaze and Parks 2012).1 Thus,

what distinguishes the United States is the punitiveness of its

criminal justice policies: the ratio of those incarcerated to those
convicted is 70% higher in the United States than the next high-
est country (Civitas 2012). Such punitive policies are extremely
costly: federal, state, and local expenditures on corrections cur-
rently exceed $82 billion annually, with the direct expenditures
on the wider justice system totaling over $250 billion (Kennelman
2012).
The high rate of incarceration in the United States also ex-
tends to juveniles. In 2010, the stock of juvenile detainees stood at
70,792, a rate of 2.3 per 1,000 aged 10–19 (Office of Juvenile
Justice and Delinquency Prevention 2011). Including those
under correctional supervision, the United States has a juvenile
corrections rate that is five times higher than the next highest
country, South Africa (Hazel 2008). Despite the high rate of ju-
venile incarceration in the United States, little is known about its
impact on juveniles. In a life-cycle context, incarceration during
adolescence may interrupt human and social capital accumula-
tion at a critical time, leading to reduced future wages in the legal
sector and greater criminal activity. More generally, interven-
tions during childhood are thought to have greater impacts com-
pared to interventions for young adults due to propagation effects
(see, for example, Cunha, Heckman, and Lochner 2006), and
criminal activity is a particularly important context to consider
such effects due to the negative externalities associated with it.2
This article aims to estimate causal effects of juvenile incarcera-
tion on human capital accumulation, as measured by high school
completion, and recidivism as an adult.
The existing research on the impacts of incarceration on
future outcomes has focused largely on adults and on such out-
comes as employment, earnings, and recidivism. The main chal-
lenge inherent in estimating the causal impact of incarceration
(for both adults and juveniles) is to control or otherwise account
for the influence of individual characteristics that may jointly
influence incarceration and future human capital accumulation,
1. See Raphael and Stoll (2013) and Neal and Rick (2014) for detailed discus-
sions of why incarceration rates have increased so much over this period.
2. When considering the determinants of criminal activity dominated by
young adults, large effects of juvenile interventions are plausible. See, for example,
Currie and Tekin (2006).
JUVENILE INCARCERATION, HUMAN CAPITAL, FUTURE CRIME 761
criminal activity, and labor market outcomes. These characteris-

tics include greater socioeconomic disadvantage, lower levels of

cognitive achievement, and less self-control.
Previous work on recidivism conducted by criminologists
yields mixed results.3 Meanwhile, the literature on labor
market outcomes generally suggests that incarceration has a
small causal impact on the labor market earnings and employ-
ment of adult men (Western, Kling, and Weiman 2001).4 These
studies attempt to address the potential endogeneity of incarcer-
ation by controlling for a limited set of observable characteristics
(Freeman 1992; Western and Beckett 1999), or they use panel
data sets that enable one to compare earnings before and after
a spell of incarceration (Lott, 1992a 1992b; Waldfogel 1994;
Grogger 1995). The fixed effect approach, however, cannot be
used to study the impact of juvenile incarceration, as juveniles
have not yet entered the labor market. Moreover, this approach
assumes that the timing of incarceration is exogenous and is not
correlated with changing life circumstances that might also affect
labor market outcomes. A shock to labor market productivity, for
example, could lead to criminal behavior rather than the
opposite.
Another approach, first employed by Kling (2006), is to in-
strument for sentence length using an index of each judge’s sen-
tencing severity. Kling (2006) finds that incarceration has small
positive effects on employment that fade over time. Di Tella and
Schargrodsky (2013) and Green and Winik (2010) have also used
this strategy to estimate the impact of incarceration on recidivism
with differing results. This approach implicitly controls for all
unobservables (fixed and changing) that might bias estimates
3. Some work finds that incarceration increases recidivism (Spohn and

Holleran 2002; Bernburg, Krohn, and Rivera 2006), others find that it has no
effect (Gottfredson 1999; Smith and Akers 1993), and still other work finds that it
reduces recidivism (Murray and Cox 1979; Brennan and Mednick 1994). Kerley
et al. (2004) find that incarceration is negatively correlated with income, especially
for those incarcerated earlier in life.
4. Rehabilitation programs could improve labor market outcomes (Landerso
2012), although prisoners in the United States have access to few rehabilitative
services. For example, according to the General Accounting Office (2012), 31,000
prisoners are enrolled in drug rehabilitation programs, while another 51,000
remain on waiting lists.
because judges are randomly assigned to cases and is the ap-

proach that we take in this article.5

Much less is known about the consequences of incarcerating
juveniles on future outcomes. The handful of studies that exam-
ine the effect of juvenile criminal activity on education and labor
market outcomes generally find a negative relationship. Most of
the existing studies attempt to identify the causal link by control-
ling for observed individual characteristics (De Li 1999; Tanner,
Davies, and O’Grady 1999; Sweeten 2006). More sophisticated
studies of this type also control for interactions with the criminal
justice system as well as unobserved household fixed character-
istics (Hjalmarsson 2008).
A second complicating factor is that effects for juveniles on
the margin of juvenile incarceration may differ from the average
juvenile, and the former group is most likely to be affected by
policy changes. Perhaps the most convincing evidence to date
comes from a regression-discontinuity design using sentencing
rules to identify the impact of juvenile incarceration on recidi-
vism in Washington state (Hjalmarsson 2009). At the margins
where the sentencing becomes more severe, juveniles just above
the thresholds were found to be less likely to recidivate as a
youth.
Our estimation strategy addresses these complicating fac-
tors. We exploit plausibly exogenous variation in juvenile deten-
tion stemming from the random assignment of cases to judges
who vary in their sentencing. To illustrate, consider two juveniles
randomly assigned to two different judges with differing incar-
ceration tendencies. With random assignment, differences in
incarceration between juveniles are attributed to the effect of
the judge and not individual characteristics of the juvenile or
the case, as are differences in outcomes. With this strategy we
address the issue of negative selection into juvenile incarceration
and estimate effects for those at the margin of incarceration
where the judge assignment matters for the incarceration
5. A somewhat related paper by Drago, Galbiati, and Vertova (2009) exploits a

natural experiment induced by an Italian clemency bill that increased sentence
length associated with future crime for some former prisoners but not others. The
authors find that the increase in expected sentence length exerted a strong deter-
rent effect on future crime, but that the deterrent effect decreased with length of
time previously incarcerated.
decision.6 Unlike previous work, we use this strategy in a context

of juvenile offending where human capital accumulation may still

be in its formative stages, and thus the long-term effects may well
be greatest.
To carry out this strategy, we employ a unique source of linked
administrative data for over 35,000 juveniles over 10 years who
came before a juvenile court in Chicago, Illinois. These data
were linked to both public school data for the same city and
adult incarceration data for the same state to investigate effects
of juvenile incarceration on high school completion and adult
imprisonment.
We find that assignment to a judge with a high incarceration
rate in other cases leads to a significantly lower likelihood of high
school completion and a significantly higher likelihood of incar-
ceration as an adult, including incarceration for violent crimes.
Under the stronger assumptions necessary to use this rate as an
instrumental variable, juvenile incarceration is estimated to
decrease high school graduation by 13 percentage points and
increase adult incarceration by 23 percentage points. In compar-
ison, in OLS regressions with minimal controls, those incarcer-
ated as a juvenile are 39 percentage points less likely to graduate
from high school and are 41 percentage points more likely to have
entered adult prison by age 25 compared with other public school
students from the same neighborhood. Though the instrumental
variable (IV) results are considerably smaller than the OLS re-
sults with minimal controls, the differences remain large and
suggest substantial negative effects of juvenile incarceration on
long-term outcomes.
The main IV estimates and subgroup analyses suggest that
marginal cases are at particularly low (high) risk of high school
completion (adult incarceration) as a result of juvenile custody.
Indeed, the effect sizes are larger for juveniles whose observable
characteristics suggest that they are less likely to be incarcerated
as a juvenile. The results are also consistent with the idea that
the timing of incarceration matters: the strongest results are for
juveniles aged 15 and 16—a critical period of adolescence when
incarceration is most likely to end one’s high school education.
6. Chang and Schoar (2008) and Dobbie and Song (2013) employ a similar
strategy using judges assigned to bankruptcy cases. Maestas, Mullen and Strand
(2013) use disability examiner propensities to approve disability claims, and Doyle
(2008) uses case worker propensities to place children in foster care.
Finally, we explore the potential mechanisms behind the es-

timated negative effects. We find that although incarceration of

these juveniles is intended to be short in duration (one to two
months), it can be very disruptive. Once incarcerated, juveniles
are unlikely to ever return to school. However, conditional on
returning to school, they are not more likely to be transferred
to an alternative (potentially inferior) school, nor are they more
likely to be classified as special education students. Interestingly,
they are more likely to be classified for special education services
due to behavioral/emotional disorders rather than a cognitive
disability.
Our results have important implications for policies related
to juvenile incarceration, such as the adoption of alternatives to
incarceration in juvenile courts across the United States. They
also have implications for recent changes in education policy that
have placed more police officers in U.S. schools. This increase has
lead to an increase in juveniles being arrested, often for less se-
rious crimes.7 Our results suggest that more research needs to be
done evaluating whether and to what extent this has led to an
increase in juvenile incarceration to better inform policy makers’
decision to expand such a program.
The rest of the article is organized as follows. In Section II we
provide background information on the juvenile justice system
and judge assignment in our context; in Section III we describe
the data. Section IV describes the empirical strategy, Section V
presents the results, and Section VI offers interpretation and
conclusions.
II. Background
II.A. The Juvenile Justice System and Judge Assignment
In Chicago, juvenile offenders of minor crimes are often dealt
with directly by the police. Only after a number of smaller infrac-
tions or a major infraction will a child enter the juvenile court
system.8
7. In 1975, 1% of U.S. schools had police, increasing to 22% in 1997, and 40% by
2007 (Na and Gottfredson 2011).
8. Every juvenile arrest is reviewed twice before proceeding to juvenile court:
first by the police and a second time by the prosecutor’s office. At each review the
juvenile’s case can be disposed. Only those cases not dismissed by the police or the
prosecutor proceed to juvenile court.
When juveniles are charged with a crime in juvenile court,

they are assigned to a calendar which corresponds to the youth’s

neighborhood of residence. Calendars generally have one or two
judges who usually preside over cases assigned to them.9
Furthermore, approximately one-fifth of hearings are presided
over by judges who cover the calendar when the main judge(s)
are not available, known as ‘‘swing judges.’’
Within a calendar, the judge assignment is a function of the
sequence with which cases happen to enter into the system and
the judge availability that is set in advance. In particular, there
does not appear to be scope for influencing the first judge seen. It
is at the first court hearing, for example, that juveniles meet their
public defenders (who are also assigned based on day of hearing)
and learn who the judge will be. Conversations with court admin-
istrators confirm that these assignments are effectively random
and that there is no way to influence the judge assigned to the
case. As a partial check on this important assumption of random
assignment, we test the relationship between observable charac-
teristics and judge assignment.
One exception to calendar assignment based on residence of
the juvenile is youths charged with a weapons offense. Over our
time period, these youths can be assigned to a separate calendar
that oversees such offenses, but assignment to a judge within the
‘‘weapons’’ calendar is still based on the sequence of court cases
being heard.10 We account for this differential treatment of weap-
ons charges in our analysis, as described in the section on empir-
ical strategy.
In terms of sentencing, nearly all cases, 96%, that come
before the court are found guilty (typically by plea) of the charges
(Peters et al. 2002). As a result, a judge’s main influence on the
case is whether the juvenile is placed on probation or detained
and then placed on probation. For the first cases we consider here,
custody is nearly always in the Cook County Juvenile Temporary
Detention Center, which is available for children aged 10–16—
the ages applicable for juvenile offenses in Illinois.11 These
9. The district attorney is also assigned by calendar so that within a calendar

the characteristics of the district attorney are constant across the judges.
10. We attempted to learn more about the weapons court, but since it no longer
exists, administrators could only confirm its previous existence and not provide
more detail.
11. Juveniles may also be sentenced to a juvenile facility run by the Illinois
Department of Corrections where typical stays are between six months and two
sentences are indeterminate in length, but typically last one to

two months including pretrial detention. We do not observe

length of time incarcerated in our data. As a result, our analysis
considers the effects of a typical stay in incarceration (approxi-
mately 42 days in our data), but not whether and to what extent
length of time incarcerated matters. However, we describe ana-
lyses using a proxy for length of stay in the robustness discussion.
The alternative for juveniles in our sample who receive a
sentence of incarceration followed by probation is simply proba-
tion. The conditions for successful probation are universal and
include attending school (which we consider directly), not associ-
ating with known criminals, and not using illegal drugs. Once the
juvenile is on probation, the judge no longer has any contact with
the juvenile.12
It is important to note that the juvenile incarceration rate in
this state is similar to the average for the United States as a
whole.13 This increases the likelihood that the results apply
more broadly compared to a situation where the state was an
outlier in terms of incarceration rates.
II.B. Mechanisms: How Juvenile Incarceration Can Affect

Outcomes
Juvenile incarceration can affect high school completion and
future criminal activity through two potential channels: chang-
ing the skills or actions of the individual juvenile (a behavioral
channel) or changing the ways institutions regard and treat him
(a deviant labeling channel). With respect to the former, incar-
ceration can negatively affect child mental health, leading to be-
havioral problems in school and at home (Kashani et al. 1980;
years, which we also regard as incarcerated in our treatment variable. Only 0.6% of
the cases in our analysis sample are found in the Department of Corrections facility
within a year of the first hearing, however.
12. One issue is that judges could sentence juveniles to electronic monitoring or
home curfews (considered alternatives to detention). However, these alternatives
were not introduced until 1995 and were not widely used until much later. Our
sample is composed primarily of youth who came before the juvenile court prior to
adoption of these alternatives.
13. Juvenile incarceration rates per 100,000 range from 53 to 440 across the 50
U.S. states with an average 225 (Office of Juvenile Justice and Delinquency
Prevention 2011). In Illinois, the rate (178) is similar to the average for the
United States, suggesting that the state is not an outlier in its juvenile incarcera-
tion tendencies.
Forrest et al. 2000). Incarceration can also encourage the accu-

mulation of ‘‘criminal capital’’ (Bayer, Hjalmarsson, and Pozen

2009) and hinder the accumulation of social capital
(Granovetter 1995). Disrupting school attendance, even if only
for one month, can also have the effect of increasing the cost of
going to school (if ‘‘catching up’’ on lost schoolwork is costly), lead-
ing to dropout. Alternatively, incarceration could have a positive
effect on human capital accumulation and future crime by either
reducing the uncertainty regarding the cost of jail so that juve-
niles who are detained adjust (upward) the cost of spending time
in detention, or by reducing truancy since the detention facility
includes a school administered by Chicago Public Schools.
With respect to the deviant labeling channel, schools, for ex-
ample, may be unwilling to allow the juvenile to re-enroll in
school once released, forcing the juvenile to enroll in another,
potentially inferior school. Even if re-enrolled in their original
school, schools could treat juveniles differently, such as classify-
ing them as special education students. The criminal justice
system might also regard them differently—police may be more
likely to suspect and arrest them and, conditional on conviction,
judges may be more likely to sentence them more harshly, assum-
ing their juvenile records haven’t been expunged.14 Finally,
employers may be less likely to hire them (see Bernburg and
Krohn 2003), thereby increasing the likelihood of future criminal
activity.
In our empirical work, we begin by estimating the overall
effect of incarceration as a juvenile on high school completion
and adult incarceration. We follow this with an exploration of
the potential mechanisms underlying the estimated effect. To
do so we begin with an examination of whether juveniles who
are incarcerated are less likely to return to school. Among those
who do return to school, we examine whether they are more likely
to transfer schools on release, consistent with schools being
14. Juvenile records can be expunged, but this is not automatic and must be
requested. Moreover, certain conditions must be met, namely, lack of future crim-
inal conviction for five years. We spoke with juvenile defense attorneys in Cook
County to learn more about the role of juvenile detention in adult court sentencing
decisions. While adult court judges do learn about previous convictions and may
learn about previous incarceration in a Department of Corrections facility (0.6% of
our cases), they are highly unlikely to learn about detention in the Juvenile
Temporary Detention Facility (99.4% of our cases), suggesting that this is an un-
likely channel.
unwilling to re-enroll juveniles who have been incarcerated. We

follow this with an exploration of whether students who are in-

carcerated and return to school (any school) are more likely to be
classified as a special education student and the nature of the
disability.15 Finally we examine the impact of juvenile incarcer-
ation on different types of crimes. If the effects remain for types of
crime that would result in adult incarceration regardless of the
characteristics of the offender (e.g., homicide), that would be con-
sistent with a change in criminal activity on the part of the juve-
nile, not simply a change in the way the criminal justice system
treats the offender.
III. Data Description

III.A. Data Sources
The data come from three primary sources: Chicago Public
Schools (CPS) Student Database (1990–2006), the Juvenile Court
of Cook County Delinquency Database (1990–2006), and the
Illinois Department of Corrections Adult Admissions and Exits
Database (1993–2008). The data were linked using identifiers in-
cluding name, date of birth, and address information by the
Chapin Hall Center for Children, a child welfare research insti-
tute—and a leader in administrative data linkage—located at the
University of Chicago (Goerge, Van Voorhis, and Lee 1994).
The CPS data come from a system that characterizes each
child by his or her age, race, sex, birth year, measures of special
education needs, as well as the U.S. census tract of residence. We
linked the tract information to 2000 U.S. census data describing
the fraction of families in poverty. We aggregated each student’s
residence to one of 76 long-standing neighborhoods in Chicago, 67
of which are included in our analysis data set.16 Results control-
ling for the tract itself are reported in the robustness section.
The raw Juvenile Court data are at the hearing level.
These data include the date, a judge identifier, the offense, and
the disposition: probation or detention followed by probation.
Unfortunately, the length of time in a juvenile facility is not
15. There are no data on suspensions or other disciplinary actions during most
of the period of time our sample was school age.
16. On average, a community comprises 14 census tracts. We use the definitions
of community as defined by the University of Chicago, which can be found here:
http://www.lib.uchicago.edu/e/collections/maps/ssrc/.
part of the disposition; rather, the sentences tend to be indeter-

minate subject to future hearings.

The Illinois Department of Corrections data describe each
adult prisoner’s spell and allow us to observe whether these ju-
veniles are found in adult prison in Illinois later in life.
Furthermore, the data list the offense for which the individuals
are incarcerated, and we test the effects of juvenile incarceration
on adult incarceration for different types of offenses.
III.B. Sample Construction

Online Appendix Table AI describes the sample construc-
tion.17 We begin with the sample of students in the eighth
grade in the CPS system. One of our main outcomes of interest
is adult incarceration by age 25, and to measure this outcome
without censoring, we restrict the sample to those who are at
least 25 by 2008—the last year of our incarceration data, corre-
sponding to the cohort born between 1971 and 1983. This also
ensures that we do not have censoring with respect to the high
school graduation outcome. Two percent of the data were ex-
cluded due to missing U.S. census tract information or a small
number of recorded tracts that were not linked to a Chicago com-
munity, resulting in 440,797 children.
Of these, 41,764 (9.5%) of these students came before the
juvenile court system during our timeframe (1990–2006). We
focus on the juvenile’s first case in our data. We excluded a
small number of observations where the case was transferred to
an adult court, where the recorded age was miscoded (i.e., not
between 10 and 16), as well as cases assigned to judges with
fewer than 10 cases. Finally, the baseline regressions employ
fixed effects defined at the community year weapons offense
level (for reasons explained in the empirical strategy later), and
we drop cases where these cells have fewer than 10 observations.
This results in 37,692 observations in the juvenile court data.
III.C. Sample Description

Table I reports sample means for the entire CPS sample and
the juvenile court sample. For the latter, we further divide the
sample into those incarcerated as a juvenile and those not
17. All appendix material is in the Online Appendix.

TABLE I
SAMPLE MEANS

(1) (2) (3)
Full CPS Juvenile court sample
sample
Not
incarcerated Incarcerated
Juvenile characteristics
Incarcerated as a juvenile 0.021 0.00 1.00
Male 0.51 0.81 0.93
African American 0.55 0.75 0.77
Hispanic 0.27 0.17 0.15
White 0.14 0.07 0.07
Other 0.04 0.01 0.01
Special education in eighth grade 0.12 0.22 0.31
Birth year 1977 1978 1978
U.S. census tract poverty rate 0.221 0.28 0.29
Age at offense N/A 14.81 14.27
Charges
Aggravated assault N/A 0.13 0.11
Burglary 0.11 0.13
Drug law violation 0.20 0.20
Larceny theft 0.05 0.03
Car theft 0.11 0.11
Robbery 0.06 0.07
Simple assault 0.10 0.06
Vandalism 0.05 0.04
Weapons offense 0.11 0.17
Other offense 0.09 0.08
Outcomes
Graduated high school 0.40 0.12 0.03
Incarcerated as an adult 0.064 0.28 0.49
by age 25
Sample size 440,797 29,141 8,551
Notes. This table reports summary statistics for the Chicago Public School Student Database (1990–
2006) for students in a CPS school in eighth grade and at least 25 years old in 2008; columns (1) and (2)
restrict this sample to those linked to the Juvenile Court of Cook County Delinquency Database, including
cases from 1990–2000, as described in the text.
incarcerated. The only characteristic along which all three groups

appear similar is birth year, with most of the mass in the
1974–1982 birth cohorts. The samples differ considerably along
all other dimensions. The juvenile court sample is more likely
to be male and African American, more likely to be special edu-
cation students in eighth grade, and live in higher poverty
neighborhoods.18 They are also less likely to graduate high school

and more likely to be incarcerated by age 25. The graduation rate

for the full sample is only 40%, defining transfers as not graduat-
ing from high school.19 These comparisons suggest that limiting
the analysis sample to the juvenile court sample and compar-
ing outcomes for those incarcerated with those convicted but
not incarcerated is likely to reduce the bias associated with neg-
ative selection into incarceration on underlying characteristics.
However, it will not likely eliminate the bias—the samples still
differ on key characteristics that are correlated with the outcomes.
Those incarcerated are more likely to be male, slightly more (less)
likely to be African American (Hispanic), and much more likely to
be special education students in eighth grade. Those incarcerated
are also slightly younger at the time of the first offense—a predic-
tor of adult incarceration. With an average birth year of 1978 and
an average age at the first offense of 14–15 years old, the typical
case occurs in 1992–1993 in this analysis sample.
One drawback of the data is that they include only school
completion (incarceration) outcomes in the same city (state) as
the juvenile court. If individuals move away, we do not observe
their high school completion or their recidivism. Regarding high
school completion, among juveniles charged with a crime, 3.4%
transfer to private school and 10% transfer out of the district,
suggesting that we can accurately measure high school comple-
tion for the vast majority of juveniles. For the main specification,
we code this 13.4% of the sample as nongraduates. Another 18%
of the sample transfer from the CPS to an adult correctional
facility without completing high school. These individuals are
also coded as non–high school graduates.20 In the robustness sec-
tion we consider these transfers directly. Regarding our measure
of adult recidivism, data from the 2000 census show that among
those born in Illinois between 1970 and 1982, by 2000 (when they
range in age from 18 to 30), three quarters remain in Illinois,
and the rate of migration is lower for those with less education.
18. The poverty rate was filled in with the average of the contiguous census tract
poverty rates for 6 observations in the court sample and 23 observations in full CPS
sample when that information was suppressed by the census.
19. In Illinois during this period the school-leaving age was 16. This age has
been raised to 17.
20. These individuals can earn a GED in prison, but we do not have that infor-
mation. Even if they did complete a GED, a GED confers much lower wages than
does a high school diploma.
We anticipate little bias to be introduced by this form of sample

selection.

IV. Empirical Framework
IV.A. Setup
For juvenile i, consider a model that relates an outcome such
as adult recidivism, Yi, to an indicator that the juvenile was in-
carcerated at some point during his youth, JIi:
ð1Þ Yi ¼ 0 þ 1 JIi þ 2 Xi þ i ;
where Xi is a vector of control variables and i is the error term.
Any assessment of the impact of juvenile incarceration on
high school completion and adult incarceration must address
the problem posed by the positive correlation between juvenile
incarceration and factors such as severity of the crime, criminal
history, and characteristics of the juvenile that are also likely to
be correlated with the outcomes.
In our analysis, we take several steps to address this.
Specifically, we present several different specifications that incre-
mentally control for confounding factors so that we can observe
the extent to which omitted variables may be driving the ob-
served correlations between juvenile incarceration and the out-
comes. Initially, we compare juveniles incarcerated with other
children in the public school system from the same neighborhood.
We then present specifications that (i) add controls for multiple
demographic characteristics including race, sex, birth year, share
in poverty in the census tract of residence, and an indicator for
special education status in eighth grade; (ii) employ propensity
score techniques using these same geographic and demographic
controls in an attempt to further control for omitted variables;
and (iii) limit the analysis to all juveniles charged with a crime
and brought before the juvenile court, though not necessarily in-
carcerated, further controlling the age at the time of the offense
(instead of birth year), the type of crime (10 categories) and a risk
assessment index, which is a checklist of criteria that is applied
by the Department of Probation to rate each juvenile for specific
detention-related risks.21
21. The scale ranges from 1 to 15 with a higher number indicating greater risk
and therefore stronger recommendation for detention. We calculated the index
Despite the inclusion of an increasingly comprehensive set of

controls, there may still be unobservable characteristics of either

the crime or the juvenile that are correlated with both the prob-
ability of juvenile incarceration and future outcomes. In the case
of high school completion, it’s most likely that these unobservable
characteristics are negatively correlated with juvenile incarcera-
tion, biasing OLS estimates of the impact of JI downward, and in
the case of adult incarceration, it’s most likely that the unobser-
vable characteristics are positively correlated with JI, which
would bias OLS estimates of the impact of JI upward.
In addition, the effects of juvenile incarceration are likely to
be heterogeneous, and we could augment the model to allow for a
random coefficient on juvenile incarceration, which would allow
the effects to vary by juvenile. A concern in estimating such
models is a correlated random coefficient (Bjorklund and Moffitt
1987), where the placement into custody may be related to the
effect on adult incarceration. That is, judges choose the sentence,
and if they tailor sentences with the idea of deterring future crim-
inal activity, then a selection bias could understate the causal
effect of juvenile incarceration for cases on the margin of commit-
ment: those cases most likely affected by policy.
Our main empirical strategy uses a measure of the tendency
of a randomly assigned judge to order a juvenile be placed
in custody, Z, as an instrument for juvenile incarceration.
Essentially, we compare high school completion and adult incar-
ceration rates for juveniles assigned to judges that have different
propensities to incarcerate, and interpret any difference as a
causal effect of the change in incarceration associated with the
difference in these propensities. These can be considered mar-
ginal cases where the judges may disagree about the custody
decision, a margin of particular policy relevance. In the next sub-
section, we describe how we calculate the instrument in greater
detail.
IV.B. IV Calculation
For each juvenile we assign an instrument that corresponds
to the ‘‘incarceration propensity’’ of the initial judge in the
from the charge information. In the models with the charge category indicators, this
index serves to further control for the severity of the charge among those with
‘‘other offenses.’’
juvenile’s first case. The instrument, which is defined for each

juvenile i assigned to judge j(i) is simply a leave-out mean:

nX !
jðiÞ 1
1 ~k :
ZjðiÞ ¼ JI
njðiÞ 1 k6¼i
Here, nj(i) is the total number of cases seen by judge j; k in-

dexes the juvenile case seen by judge j where JI~ is equal to 1 if the
juvenile was incarcerated during the juvenile’s first case. Thus
the instrument is the judge’s incarceration rate among first cases
based on all the judge’s other cases. Algebraically, this is the
judge fixed effect in a model of custody in the initial case esti-
mated in a ‘‘leave-out’’ regression estimated over all years. This
measure ties the decision making of the judge in the first case
more directly to the first cases we consider here. The resulting
two-stage least squares estimator is a jackknife instrumental var-
iables estimator (JIVE), which is recommended for models when
the number of instruments (the judge fixed effects) is likely to
increase with sample size (Stock, Wright, and Yogo 2002;
Kolesar et al. 2011).
In both the first and second stages of the IV regressions, we
also include a vector of community weapons-offense year
fixed effects. Recall that judge assignment is based on community
and, during part of our time period, whether there was a weapons
charge. Including this fixed effect thus effectively limits the com-
parison to juveniles at risk of being assigned to the same set
of judges. With the inclusion of these controls, we can interpret
the within-cell variation in the instrument, Zj(i), as variation in
the propensity of a randomly assigned judge to incarcerate a ju-
venile relative to the other juvenile cases seen from the same
neighborhood and with either a weapon or nonweapon offense
in the same year. Note that the IV calculation is not conditional
on characteristics of the juvenile or the crime to allow a direct
examination of the sensitivity of the results with and without
controls.
IV.C. Judge Variation

Our analysis data set includes 62 judges. The average
number of initial cases per judge is 607. More than one judge
can hear each juvenile’s case over time, and the instrument is
based on the incarceration propensity of the first judge assigned
for the juvenile’s first offense.22 If the initial judge is missing in the
data as it is in 17.8 percent of the cases, we assign the juvenile to

the second judge of record. While the potential for another, later
judge to make the incarceration decision may lead to a weaker
estimated relationship between the first judge’s propensity to
incarcerate (the instrument) and an individual juvenile’s incar-
ceration status, the focus on the first judge has the advantage of
not capturing any (potential) nonrandom changing of judges.
The initial case incarceration propensity has a mean of 0.097
with a standard deviation of 0.039. Results will be shown with
alternative measures of the instrument as checks on robustness
as well. Variation in the instrument can also be seen in Figure I,
where we present the distribution of the instrument defined two
ways. First, as the leave-out mean of the probability of incarcer-
ation for each judge (denoted ‘‘raw’’ in the figure). Second, as the
residual from a regression in which we include controls shown in
Table I, including indicators for each year of age at the time of the
offense (rather than birth year) and community weapons of-
fense year indicators. The residualized measure represents
the variation in the instrument that we use for identification
and suggests substantial variation even with a full set of controls.
In particular, the raw measure ranges from approximately 4% to
21%, and the residualized measure still shows substantial varia-
tion: ranging from 6% to 18%.
This variation comes from two sources: variation among
‘‘regular’’ (i.e., nonswing) judges assigned to the same calendar
(roughly 80% of cases are seen by a regular judge) and variation
from swing judges who oversee the remaining 20% of the cases.23
V. Results
V.A. Instrument Validity
Although we cannot directly test the exclusion restriction, we
can provide evidence consistent with the condition being met.
22. Thirty-five percent of the initial cases have the same initial and final judge
across all of the hearings. Over the course of the criminal proceedings, which often
involve multiple hearings, the judge may change either temporarily or
permanently.
23. We define regular judges as those who see at least 75% of their cases in the
given calendar year and swing judges as those who see fewer than 75% of their
cases in a given calendar year.
Residualized Raw
12000

10000
8000
6000
4000
2000
0
0 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 0.11 0.12 0.13 0.14 0.15 0.16 0.17 0.18 0.19 0.2 0.21 0.22 0.23 0.24
FIGURE I
Distribution of Z: Judge Incarceration Rate
Notes: These histograms display the distribution of the leave-one-out mean
incarceration rate for the first judge in the first case in the linked Chicago
Public School–Juvenile Court of Cook County data including cases from
1990–2000 as described in the text. The residualized measure was calculated
from a regression model with full controls listed in Table I, including indicators
for each year of age at the time of the offense and community weapons
offense year indicators.
First, we have confirmed with court personnel that judges are

assigned in a way that leads to a ‘‘natural randomization’’ of
cases to judges: cases are assigned to calendars based on the ju-
venile’s residence and within calendars, judges cannot influence
which cases they hear. Second, we can partially test this empir-
ically in the data by examining whether the characteristics of
juveniles and their cases differ by judge. We do this by testing
whether the characteristics of juveniles differ based on whether
they are assigned to a judge with a high, medium, or low propen-
sity to incarcerate (defined by bottom, middle, or top tercile of the
distribution of propensity to incarcerate) relative to other judges
in the same community weapons offense year cell. The results
(Table II) show that judges with high, medium, and low propen-
sities to incarcerate are assigned juveniles that are extremely
similar in terms of their gender, race, and special education
needs; poverty rate of his census tract; and age at the time of
TABLE II
INSTRUMENT VERSUS JUVENILE CHARACTERISTICS

Z distribution
Middle vs. Top vs.

Bottom Middle Top bottom bottom
tercile tercile tercile p-value p-value
Z: first judge’s leave-out mean 0.062 0.094 0.147 (.000) (.000)
incarceration rate in first cases
Juvenile characteristics
Male 0.827 0.830 0.833 (.561) (.311)
African American 0.724 0.737 0.742 (.096) (.249)
Hispanic 0.189 0.176 0.172 (.061) (.272)
White 0.078 0.079 0.078 (.833) (.957)
Other race/ethnicity 0.009 0.008 0.007 (.352) (.345)
Special education 0.241 0.237 0.252 (.549) (.130)
U.S. census tract poverty rate 0.264 0.265 0.265 (.572) (.696)
Age at offense 14.8 14.8 14.8 (.437) (.434)
P(Juvenile incarceration j X) 0.219 0.221 0.220 (.251) (.516)

Observations 37,692
Notes. This table reports summary statistics for the linked Chicago Public School–Juvenile
Court of Cook County data including cases from 1990 to 2000 as described in the text. p-values reported
in parentheses were calculated from separate regression models of each characteristic on indicators
that the judge’s incarceration rate (Z) was in the middle or top tercile along with commu-
nity weapon year fixed effects using standard errors clustered at the community level. P(Juvenile
incarceration j X) is a predicted propensity for ever being incarcerated as a juvenile using a probit
model that employs the juvenile characteristics listed above, including indicators for each year of age at
the time of the offense.
the offense, despite significant differences in the incarceration

rate used to define the categories.24 We also calculate a single
measure—the propensity to be incarcerated based on the observ-
able juvenile characteristics and report the average propensity
for lenient, moderate, and strict judges. The propensity is the
24. An F-test of joint significance for whether these control variables predict
that the judge is in the top tercile yields a p-value of .15. When we regress our
(continuous) instrument on all of the controls, however, the controls are jointly
significant at the 1% level. That said, the coefficient values themselves are very
small: the two variables that are individually significant include special education
status and the age = 10 indicator. Special education status is associated with an
increase in the instrument of 0.0017 (compared to a mean of 0.097). The age = 10
indicator is associated with a 0.0088 reduction in the instrument (compared to the
age = 16 indicator, the excluded category).
same (0.22) regardless of judge type.25 Table II reports results for

exogenous variables, whereas the results that follow will be

shown with and without controls for the potentially endogenous
control variables determined by the court (i.e., charges). Despite
not finding a relationship between observable juvenile character-
istics and the judge’s propensity to incarcerate, we offer a final
piece of evidence, which is to present and compare results when
we control for case characteristics and models when we do not.26
We interpret the IV results as local average treatment ef-
fects: average effects for cases where the judge assignment mat-
ters for the incarceration decision. This requires a monotonicity
assumption: assignment to a strict judge need not increase the
likelihood of incarceration for each type of offender. This assump-
tion is stronger in this setting (Imbens and Angrist 1994). For
example, a judge might treat cases involving drugs relatively
harshly, but theft/larceny relatively leniently. We consider this
concern more fully in our robustness checks.
V.B. First Stage: Judge Assignment and Juvenile Incarceration

To consider the first-stage relationship between initial-judge
assignment and whether the juvenile is ever incarcerated as a
juvenile (JI), we estimate the following equation for juvenile i
assigned to judge j(i) in community weapon-offense year cell
c(i) using a linear probability model:
JIi ¼ 0 þ 1 ZjðiÞ þ 2 Xi þ cðiÞ þ i :
The vector Xi represents demographic controls and court
measures described above, as well as an indicator that the
judge identifier at the first hearing is missing). Similar results
are found for both the first-stage and the instrumental variable
results when probit models are used, which is unsurprising given
that the outcome variables are relatively far from zero. Zj(i) refers
to the judge’s incarceration rate among juveniles’ initial cases.
25. The propensity is predictive of incarceration, however. The predicted pro-

pensity is 0.272 for juveniles who were incarcerated and 0.214 for those who were
not incarcerated.
26. Another concern would be that judges may affect juveniles in other ways
besides the likelihood of juvenile incarceration. Again, conversations with court
personnel suggest that this is not the case. Moreover, judges who are more likely
to incarcerate are not more likely to incarcerate for a longer period of time, condi-
tional on any incarceration.
TABLE III
FIRST STAGE

(1) (2) (3)
Dependent variable: juvenile
incarcerations OLS
First judge’s leave-out mean incarceration 1.103 1.082 1.060
rate among first cases (0.102) (0.095) (0.097)
Demographic controls No Yes Yes
Court controls No No Yes
Observations 37,692
Mean of dependent variable 0.227
Notes. This table reports the first-stage relationship between juvenile incarceration and the instru-
ment: the judge’s incarceration rate using the linked Chicago Public School–Juvenile Court of Cook
County data including cases from 1990–2000 as described in the text. All models include commu-
nity weapons-offense year-of-offense fixed effects. Demographic controls include indicators for four
age-at-offense categories, four race/ethnicity categories, sex, special education status, and the 2000 U.S.
census tract family poverty rate. Court controls include nine offense categories, indictors for seven risk-
assessment index categories, and whether the first judge assigned was missing. Standard errors are
reported in the parentheses and are clustered at the community level.
The mean initial judge custody rate is 0.09, whereas the mean of
the dependent variable in this first-stage model—an indicator
that the juvenile was ever incarcerated—is 0.23. All standard
errors are clustered at the community level.
The results of the first stage (Table III) show that the judge’s
incarceration rate is highly predictive of whether an individual
will ever be incarcerated as a juvenile. Including additional con-
trols in columns (2) and (3) does not change the estimated effect of
being assigned to a strict judge in one’s first court appearance,
consistent with the randomness of judge assignment. Column (3),
which includes the full set of controls, reports a coefficient of 1.06.
The coefficient is not statistically significantly different from 1,
meaning that if a juvenile is assigned to a judge that is 10% more
likely to incarcerate other juveniles in their initial case, he is 10%
more likely to be incarcerated as well.27 In particular, the esti-
mate suggests that a 2 standard deviation increase in the judge
27. A coefficient greater than 1 is possible because the incarceration rate (Zj(i))
applies to whether the juvenile was incarcerated in his first case, whereas the en-
dogenous variable for which we instrument is whether the juvenile was ever incar-
cerated as a youth—for his first case or any subsequent cases.
incarceration rate would imply an increase in the likelihood of

juvenile incarceration of 8.5 percentage points—or 37% of the

mean rate of juvenile incarceration. All first-stage estimates are
precise, with t-statistics around 11.
V.C. Juvenile Incarceration and High School Completion

We estimate the impact of incarceration at any time as a
juvenile on the probability of graduating from high school accord-
ing to the equation below that echoes equation(1) above:
Yi ¼ 0 þ 1 JIi þ 2 Xi þ cðiÞ þ i ;
where Yi is an indicator for whether juvenile i in commu-

nity weapons offense year cell c(i) graduated from high
school, and JIi is an indicator for whether juvenile i was ever
incarcerated as a juvenile. We present both OLS regression re-
sults and results in which we instrument for JIi using the judge
incarceration rate, Zj(i). As with the first stage, we present results
both with and without controls (Xi). When we report results for
the full CPS sample, the year of offense and weapons offense
components of the fixed effects do not apply to those not part of
the juvenile justice system. As a result, those models include com-
munity fixed effects and birth cohort indicators instead.
Table IV reports the results. The table is organized such that
with each column we further control for potential omitted vari-
ables so that we can learn about the source(s) and size of any bias.
In the first three columns, the sample includes all children in the
CPS. Therefore in the first three specifications we are comparing
the high school completion rates of children incarcerated as juve-
niles to a control group from the same community that includes
two groups: those without any juvenile court involvement and
those with juvenile court involvement but not incarcerated as
juveniles. In the first column, which includes only community
fixed effects as controls, we observe a strong negative relation-
ship: children incarcerated as juveniles are 39 percentage points
less likely to complete high school than are other children from
their neighborhood. In column (2) we include the following demo-
graphic controls: sex, race/ethnicity, share below poverty in cen-
sus tract, year of birth fixed effects, and an indicator for special
education status in eighth grade. When we do, the coefficient es-
timate falls by almost a fourth from 0.39 to 0.29, which is still
TABLE IV
JUVENILE INCARCERATION AND HIGH SCHOOL GRADUATION
Dependent variable: graduated high school

(1) (2) (3) (4) (5) (6) (7)
Full CPS sample Juvenile court sample
Inverse propensity
OLS OLS score weighting OLS OLS 2SLS 2SLS
Juvenile incarceration 0.389 0.292 0.391 0.088 0.073 0.108 0.125
(0.0066) (0.0065) (0.0055) (0.0043) (0.0041) (0.044) (0.043)
Demographic controls No Yes Yes No Yes No Yes
Court controls N/A N/A N/A No Yes No Yes
Observations 440,797 440,797 420,033 37,692
Mean of dependent variable 0.428 0.428 0.433 0.099
Notes. This table reports the relationship between juvenile incarceration and graduation from Chicago Public Schools. Columns (1)–(3) include all students in Chicago Public
Schools in eighth grade during 1990–2006 and at least age 25 by 2008. Columns (1) and (2) include community fixed effects, while column (2) also includes indicators for race, sex,
special education status, each year of birth, and the 2000 U.S. census tract family poverty rate. Column (3) used the same controls and community indicators to calculate the
propensity score using a probit model, estimated on a subsample where probit estimation is possible (where there is variation in juvenile incarceration within cells). Columns (4)–(7)
use the linked Chicago Public School–Juvenile Court of Cook County data including cases from 1990–2000 as described in the text. These models include community weapons-
offense year-of-offense fixed effects. Demographic controls include those listed for column (2). Court controls include nine offense categories, indictors for seven risk-assessment
index categories, and whether the first judge assigned was missing. Standard errors are reported in the parentheses and are clustered at the community level. The propensity score
standard errors were calculated using 200 bootstrap replications.
JUVENILE INCARCERATION, HUMAN CAPITAL, FUTURE CRIME
781

very large given an average rate of high school completion among

this sample of 43%.28

We also present propensity score estimates to determine
whether this method can further limit the amount of omitted
variable bias. We predict the probability of juvenile incarceration
using a probit regression with the demographic characteristics
listed above as well as community indicators and estimate the
relationship between juvenile incarceration and high school com-
pletion using inverse-propensity score weighting. The result
(column (3)) is an estimate of the impact of incarceration on
high school completion that is the same as the result obtained
when we excluded most of the controls, suggesting that this
method does not effectively reduce omitted variable bias in this
particular context.29
In the next two columns (columns (4) and (5)), we limit our
sample to children with a criminal case in juvenile court. By using
this subsample, we limit our comparison or control group to ju-
veniles charged with a crime in court but not incarcerated. We
argue that this sample restriction is likely to further reduce po-
tential omitted variable bias, as this subsample is much more
disadvantaged than the general CPS sample and therefore
more similar to the sample of juveniles who are incarcerated
(see Table I). Moreover, this limits the control group to those at
risk of incarceration. Our OLS estimate in column (4), which in-
cludes only community weapons offense year of offense fixed
effects, supports this: the coefficient on juvenile incarceration
falls to 0.088 when we restrict the sample in this way, although
this is still large compared to the mean graduation rate in the
sample of 9.9%. Adding additional controls for the demographic
characteristics listed above and the characteristics of the case
(type of charge, etc.) in column (5) reduces the OLS estimate
only slightly to 0.073. This suggests that either we have ade-
quately addressed most of the potential bias from omitted vari-
ables with our sample selection and set of controls, or that the
28. As already noted, those that do not graduate include those who have trans-
ferred out of Chicago Public Schools, and it’s possible that they may have graduated
from another school, though we do not observe this. We investigate sensitivity to
removing those that transfer or recoding them as graduates as robustness checks.
29. The propensity score estimates are based on a slightly smaller sample due to
the fact that we were unable to calculate a propensity based on a probit regression
for a small subset of the sample for whom the probit perfectly predicted failure/
success.
only way to improve on these estimates is to employ an identifi-

cation strategy that exploits plausibly exogenous variation in ju-

venile incarceration.
Our final set of estimates does just that by instrumenting for
juvenile incarceration using the propensity of an individual’s ran-
domly assigned judge to incarcerate. The IV point estimates,
0.108 (column (6)) excluding controls and 0.125 including con-
trols (column (7)), are much smaller than the OLS estimates
based on the entire sample of children (columns (1) and (2)), but
larger than the OLS estimates based on the subsample of children
with a juvenile court case (columns (4) and (5)). The standard
errors are larger as well; however, the estimates are not statisti-
cally significantly different from the OLS estimates based on the
juvenile court sample.30
Reduced-form estimates (Online Appendix Table AII) are
very similar to the IV estimates, consistent with the strong rela-
tionship between the propensity of the assigned judge to incar-
cerate and one’s own incarceration, as captured by the first-stage
coefficient of 1.06. These results imply that within the range of
incarceration rates in our data described in Figure I, moving from
the least strict to most strict judge—an increase of 12 percentage
points—increases the probability of high school dropout by 1.6
percentage points, or 16% of the mean.
To the extent that cases are randomly assigned and the main
effect of the judge on juveniles is whether they are incarcerated,
we can interpret the IV estimates to suggest that juveniles on the
margin of incarceration—compliers where the judge assignment
induces a change in the incarceration decision—are 12.5 percent-
age points less likely to complete high school: essentially all stu-
dents on this margin who enter juvenile incarceration do not
graduate.31 Taken at face value, the OLS and IV point estimates
suggest that the children on this margin may experience larger
effects of juvenile incarceration on high school completion than
the average incarcerated juvenile. That is, many juveniles may
30. We cluster at the level of the community in all specifications. Clustering at

the level of the judge (of which there are 62) yields a standard error of 0.065 in the
2SLS with full controls, column (7).
31. With 23% ever placed in custody, we can calculate the weighted average of
those placed in detention and those not placed that results in the overall mean
graduation rate of 9.9%: 0.23(X – 0.125) + 0.77(X) = 0.099, implying that X = 12.7%
graduate among those not placed in detention and X – 0.125 = 0.2% graduate among
those placed in detention.
experience little causal effect of juvenile incarceration on their

high school completion—those with minor offenses are at lower

risk of not completing high school, or those charged with very
serious crimes and certain incarceration may be at such a disad-
vantage at school that high school completion is already
extremely unlikely—whereas the marginal cases may be particu-
larly affected by incarceration. We explore heterogeneity in the
treatment effects across different types of cases to explore this
possibility later.
Moreover, the treatment of interest is binary: an indicator if
the juvenile were ever incarcerated. The IV estimate extrapolates
the change in the propensity to be incarcerated to a change in the
indicator for incarceration from 0 to 1. This extrapolation can lead
to large point estimates, as well as larger standard errors. It is
worth reiterating that the range of variation in the instrument
(and subsequently in the propensity to be incarcerated) is only 12
percentage points, so any relationship between the instrument
and the unobserved propensity to graduate high school will be
magnified. In the end, we regard the point estimate as evidence
of large effects of juvenile incarceration on high school completion
for marginal cases but recognize that the larger standard errors
suggests caution in the interpretation, especially in comparison
to the magnitude of the OLS estimates.
Our finding of a strong negative impact of juvenile incarcer-
ation on this measure of human capital accumulation suggests
that we may find negative effects on adult recidivism as well,
which we explore in the next section.
V.D. From Juvenile Incarceration to Adult Incarceration

We analyze the impact of juvenile incarceration on the prob-
ability of adult incarceration in the state of Illinois using the same
empirical specifications as before. We define adult incarceration
by whether an individual was present at any point by the age of
25 in an adult correctional facility anywhere in the state.
Moreover, since we observe the types of crimes for which individ-
uals are assigned to adult correctional facilities, we can define
adult recidivism by type or severity of the adult crime.
Table V reports results for any adult incarceration, regard-
less of crime type. The adult imprisonment rate, defined this
way, is 6.7% in the larger CPS sample. The OLS results show a
strong relationship between juvenile incarceration and adult
TABLE V
JUVENILE INCARCERATION AND ADULT CRIME
Dependent variable: entered adult prison by age 25

(1) (2) (3) (4) (5) (6) (7)
Full CPS sample Juvenile court sample
Inverse propensity
OLS OLS score weighting OLS OLS 2SLS 2SLS
Juvenile incarceration 0.407 0.350 0.219 0.200 0.155 0.260 0.234
(0.0082) (0.0064) (0.013) (0.0072) (0.0073) (0.073) (0.076)
Demographic controls No Yes Yes No Yes No Yes
Court controls N/A N/A N/A No Yes No Yes
Observations 440797 440797 420033 37692
Mean of dependent variable 0.067 0.067 0.057 0.327
Notes. This table reports the relationship between juvenile incarceration and imprisonment in an adult facility by the age of 25. Columns (1)–(3) include all students in Chicago
Public Schools in eighth grade during 1990–2006 and at least age 25 by 2008. Columns (1) and (2) include community fixed effects, while column (2) also includes indicators for race,
sex, special education status, each year of birth, and the 2000 U.S. census tract family poverty rate. Column (3) used the same controls and community indicators to calculate the
propensity score using a probit model, estimated on a subsample where probit estimation is possible (where there is variation in juvenile incarceration within cells). Columns (4)–(7)
use the linked Chicago Public School–Juvenile Court of Cook County–Illinois Department of Corrections data including juvenile cases from 1990–2000 as described in the text.
These models include community weapons-offense year-of-offense fixed effects. Demographic controls include those listed for column (2). Court controls include nine offense
categories, indictors for seven risk-assessment index categories, and whether the first judge assigned was missing. Standard errors are reported in the parentheses and are
clustered at the community level. The propensity score standard errors were calculated using 200 bootstrap replications.
JUVENILE INCARCERATION, HUMAN CAPITAL, FUTURE CRIME
785

incarceration: those who were in juvenile detention are 41 per-

centage points more likely than other children residing in the

same community to be found in an adult correctional facility by
age 25 (column (1)). Adding demographic controls reduces this
relationship to 35 percentage points (column (2)), and inverse
propensity score weighting reduces the estimated effect further
still to 22 percentage points (column (3)).
When we limit the control group to those who came before the
juvenile court but were not committed and include controls for
demographic characteristics and the type and severity of the ju-
venile crime (column (5)), the estimated effect falls to 16 percent-
age points. Note that the average adult incarceration rate for this
group is considerably higher (32.7%), so that the estimate repre-
sents an increase in adult recidivism associated with juvenile in-
carceration of 49 % compared to the mean.
The IV point estimates with and without controls in columns
(6) and (7) (0.26 and 0.23, respectively) are similar to each other
but slightly larger than the most restrictive OLS estimates
for adult recidivism.32 However, the loss of precision in the
IV estimates means that they are not statistically significantly
different from these OLS estimates, and both can be characterized
as large.33 Moreover, the reduced-form estimates suggest that in
practice, the direct effect of the judge assignment is more moder-
ate in size. Moving from the least to the most strict judge increases
the probability of incarceration as an adult by 3 percentage points,
or 9% of the mean (Online Appendix Table AII).
Overall, these estimates suggest that of the two potential
effects of juvenile incarceration on future criminal activity
(deterrence of future criminal activity versus reductions in
32. Although the point estimate declines somewhat with the addition of con-
trols, the difference is not statistically significant. Furthermore, if the decline sug-
gested that ‘‘strict’’ judges hear ‘‘tougher’’ cases, then we would expect a similar
change in magnitude when considering high school completion. Instead, the mag-
nitude increased when we added controls to the model for high school completion.
Together, this suggests that any differences in the types of juveniles who go before
stricter judges are not systematically related to the outcomes.
33. We cluster at the level of the community in all regressions. When we cluster
at the level of the judge in the 2SLS regression with full controls, column (7) of
Table V, the standard error increases to 0.084, still highly significant.
human capital accumulation, social capital and networks, or

other factors such as deviant labeling), the latter dominates.34

We also estimate the impact of juvenile incarceration on
adult recidivism by crime type, given that some types of crime
generate larger welfare costs. Specifically, we estimate the
impact of juvenile incarceration on adult recidivism for four
types of crime: homicide, violent crime, property crime, and
drug crimes. These categories are not exclusive, and an individ-
ual might have been incarcerated for more than one type of crime
by age 25.35 For each crime type, we present three sets of results:
OLS based on the full CPS, OLS based on the juvenile subsample,
and IV based on the juvenile subsample. The results (Table VI)
show that in the OLS for the full CPS sample, those who are
incarcerated as juveniles are much more likely to have recidi-
vated for each of the four types of crime. Limiting the sample to
those with a juvenile court case reduces the estimates consider-
ably though they are still large: those incarcerated are 2.1 per-
centage points more likely to be incarcerated for a homicide as an
adult (mean = 4%), 6.1 percentage points more likely to be incar-
cerated for violent crime (mean = 12%), 4.7 percentage points
more likely to be incarcerated for property crime (mean = 6%),
and 7.8 percentage points more likely to be incarcerated for a
drug offense (mean = 18%).
The IV estimates are larger, increasing to 3.5 percentage
points for homicide (though not statistically significant), 15 for
a violent crime, 14 for property crime, and 10 for drug-related
crimes. It is important to note that even though the point esti-
mates more than double in some cases, the standard errors also
increase substantially compared to the OLS standard errors. The
results broken down by type suggest that not only are children
incarcerated as juveniles more likely to recidivate as adults, but
the recidivism is for types of crime that are both serious and
costly.
A potential explanation is that judges may be more likely to
incarcerate adults if they have been incarcerated as juveniles, as
34. We considered employment and earnings as well, although it is more diffi-

cult to link juvenile cases to wage report data in Illinois that do not include the date
of birth. While we find negative point estimates of the effects of juvenile incarcer-
ation on employment, the standard errors are not precise.
35. We present additional results disaggregating the crime types to specific
crimes and again find substantial increases across a wide range of offenses
(Online Appendix Table AVIII).
TABLE VI 788
JUVENILE INCARCERATION AND ADULT CRIME TYPE
(1) (2) (3) (4) (5) (6)

Dependent variable: entered adult prison by age 25 for crime type
Homicide Violent
OLS OLS 2SLS OLS OLS 2SLS

Juvenile incarceration 0.051 0.021 0.035 0.138 0.061 0.149
(0.0031) (0.0030) (0.030) (0.0046) (0.0050) (0.041)
Sample Full CPS Juvenile court Juvenile court Full CPS Juvenile court Juvenile court
Mean of dep. var.: JI = 0 0.008 0.043 0.043 0.024 0.121 0.121
Observations 440,797 37,692 37,692 440,797 37,692 37,692
Property Drug

(0.0040) (0.0038) (0.044) (0.011) (0.0068) (0.052)
Sample Full CPS Juvenile Court Juvenile Court Full CPS Juvenile Court Juvenile Court
QUARTERLY JOURNAL OF ECONOMICS
Mean of dep. var. 0.013 0.060 0.060 0.034 0.176 0.176

Observations 440,797 37,692 37,692 440,797 37,692 37,692
Notes. This table reports the relationship between juvenile incarceration and imprisonment in an adult facility by the age of 25 for particular types of offenses. The outcomes
are not mutually exclusive. All models include full controls listed in Table V. Full CPS models include community fixed effects. Juvenile court models include community weapons
offense year-of-offense fixed effects. Standard errors are reported in the parentheses and are clustered at the community level.

juvenile records are not routinely expunged in Cook County.

However, the large estimated effect for adult recidivism for ho-

micide and violent crimes—crimes for which incarceration is
nearly certain, regardless of juvenile incarceration, suggest that
this is not driving the results. In addition, nearly all of the cases
that come before the juvenile court result in a conviction that may
affect later sentencing in an adult court; we are considering the
additional effect of juvenile incarceration.36
These results are considerably larger than effects generally
estimated for adults, which would be consistent with juvenile in-
carceration occurring during a particularly sensitive time in the
life cycle when human and social capital are forming. They are
consistent with estimates in Hjalmarsson (2008) based on survey
data and including very detailed characteristics of the juveniles,
including interactions with the criminal justice system. They are
not consistent with Hjalmarsson (2009), who uses a sentencing
index with a cutoff score for incarceration in a regression discon-
tinuity design to identify the impact of juvenile incarceration on
juvenile recidivism and finds that juvenile incarceration reduces
juvenile recidivism. One possible explanation for the different
results is that we consider different margins (cases where
judges may disagree about the incarceration decision versus
cases near the cutoff ). Another is that Hjalmarsson (2009) con-
sidered juvenile recidivism, which does not include serious of-
fenses committed as a juvenile and transferred to adult courts,
whereas this article considers adult recidivism and we show that
many juveniles are transferring out of high school and into an
adult correctional facility.37
V.E. Heterogeneous Treatment Effects across Observable

Characteristics
In this section we explore potential heterogeneity in the
treatment effects. We present OLS and 2SLS estimates stratified
by observable child characteristics (Table VII). Differences in the
IV results are suggestive of differential impacts of incarceration
36. As already noted, our discussions with court officials suggest that stays in
the Temporary Detention Facility are unlikely to be featured in adult proceedings
above and beyond the juvenile conviction.
37. That said, when we consider juvenile recidivism our context, we find that
our findings are more robust to timeframe and potential censoring: incarceration
within one year of the first hearing is associated with a greater likelihood of reap-
pearing before the juvenile court in a subsequent case.
TABLE VII
EFFECTS OF JUVENILE INCARCERATION BY CASE AND CHILD TYPES

(1) (2) (3) (4) (5) (6)
Entered Entered adult
Graduated adult prison prison by age 25
Dependent variable: high school by Age 25 for violent offense
OLS 2SLS OLS 2SLS OLS 2SLS
Juvenile offense: violent

(0.006) (0.071) (0.010) (0.109) (0.008) (0.080)
Mean of dependent variable 0.118 0.118 0.295 0.295 0.121 0.121
Observations 15,561
Juvenile offense: nonviolent

(0.005) (0.042) (0.010) (0.108) (0.006) (0.058)
Observations 22,131
Propensity of juvenile incarceration <= median

(0.007) (0.075) (0.013) (0.092) (0.0079) (0.063)
Observations 18,846
Propensity of juvenile incarceration > median

(0.005) (0.054) (0.008) (0.114) (0.0062) (0.058)
Observations 18,846
Age = 13 or 14
(0.006) (0.075) (0.012) (0.140) (0.009) (0.079)
Observations 11,404
Age = 15 or 16
(0.005) (0.056) (0.010) (0.098) (0.006) (0.064)
Observations 23,734
Special education
(0.005) (0.055) (0.012) (0.125) (0.009) (0.098)
Observations 8,999
Not special education

(0.005) (0.055) (0.008) (0.103) (0.005) (0.061)
Observations 28,693
Notes. This table reports results for different subgroups of juveniles. All models include commu-
nity weapons offense year-of-offense fixed effects and full controls as listed in Table V. Standard
errors are reported in the parentheses and are clustered at the community level.
on the propensity to complete high school and adult recidivism.

Given the data requirements of the approach, however, differ-

ences across subgroups are rarely statistically significantly dif-
ferent and should be regarded as suggestive only.
When we characterize juveniles by type of their first offense
(violent versus nonviolent), the OLS estimates of the impact of
juvenile incarceration on high school completion are similar for
the two types, but when we instrument, the negative impact of
incarceration increases in magnitude for the nonviolent for whom
the IV estimate is roughly double the estimate based on the whole
sample (Table VII, top two panels). In contrast, the IV estimate of
high school completion for juveniles accused of a violent crime are
much smaller in magnitude and insignificant. One interpretation
of these results is that the effects of juvenile incarceration on high
school completion are larger for those at the margin of incarcer-
ation in contrast to those most sure to be incarcerated. This is
consistent with results in which we split the sample based on
juveniles’ predicted probability of juvenile incarceration (esti-
mated by a probit with the full set of controls). The negative
effect of incarceration on high school graduation is much greater
for those with a lower propensity of juvenile incarceration in both
the OLS and IV settings (Table VII, third and fourth panels).
With respect to the adult incarceration effects, we also find
larger effects of juvenile incarceration on recidivism (for any
crime and for a violent crime) for those with a lower propensity
of juvenile incarceration, similar to the high school completion
results. However, juveniles charged with a violent first crime
are more likely to recidivate for any crime, and much more
likely to recidivate for a violent crime relative to those charged
with a nonviolent crime, contrary to the high school completion
results.
The impact of incarceration on high school completion and
adult recidivism also varies with juvenile characteristics such as
age (Table VII, fifth and sixth panels).38 The overall effects are
largely coming from juveniles aged 15–16, perhaps because the
38. We stratify by sex and race as well. Of the 37,692 juveniles in the sample,
fewer than 6,000 are girls and the results for girls, while large in magnitude with
respect to high school completion in particular, are very imprecise. With respect to
race, the main results are similar to those found for African Americans, the point
estimates for high school graduation are larger in magnitude for white and
Hispanic juveniles. For adult incarceration, the point estimate is particularly
large (and imprecise) for Hispanic juveniles (Online Appendix Table AV).
incarceration occurs during a point in the life cycle when drop-

ping out of school is possible. Meanwhile, the impact of incarcer-

ation is qualitatively similar for those with and without special
education needs.39
That stronger estimated effects of juvenile incarceration on
high school completion for some groups are not necessarily ac-
companied by stronger effects on adult incarceration suggests
that the impact of juvenile incarceration on adult incarceration
is not working entirely through the negative impact on high
school completion. This is not surprising, as we expect incarcer-
ation to affect a juvenile in many ways, including effects on social
capital and networks or ‘‘deviant labeling,’’ in addition to any
effect on high school graduation. Still, to gauge the potential mag-
nitude of the high school completion channel, consider that
Lochner and Moretti (2004) found that among African
Americans, high school completion results in an 8 percentage
point decline in the likelihood of being in jail as an adult (the
point estimates for whites are lower and less precise, but not
significantly different from the estimates for blacks). Based on
this, we calculate that of the 20 percentage point increase in
adult incarceration, only 5% comes from the 13 percentage
point decrease in high school completion.40
One caveat is that Lochner and Moretti (2004) base their
analysis on the 1960, 1970, and 1980 censuses. Since then, the
labor market return to high school completion has increased sig-
nificantly. Between 1980 and 2000, Deschenes (2006) estimates
that the causal return to a year of single year of schooling in-
creased by as much as 40%. As such, it is likely that the causal
impact of education on crime has likewise increased over this
period, which would result in a larger role for high school com-
pletion in explaining the impact of juvenile incarceration on adult
crime. In any event, the results suggest that for juveniles on the
margin of incarceration, such detention appears to negatively
affect the human and social capital formation in more ways
39. In Online Appendix Table AVII we present results of regressions for addi-
tional subsamples defined by gender and race. The results show that the main
results stem from male offenders, whereas the results for female offenders (a
much smaller subset of the data) are noisier.
40. There is a related literature on the relationship between school days and
criminal activity among juveniles (see Jacob and Lefgren 2003; Luallen 2006) in
which they explore the incapacitating effect of school attendance on juvenile crim-
inal activity.
than we can measure through high school completion and adult

incarceration.

In summary, the results suggest that across different groups
of children, juvenile incarceration is associated with lower high
school completion and higher adult recidivism. In general, the
high school completion results show large differences across
type, with greater effects coming from those less likely to be in-
carcerated as a juvenile. The adult recidivism results do not dis-
play the same pattern, but show large effects across types.
V.F. Exploring Potential Mechanisms

To further examine the potential mechanisms behind our
results (behavioral changes or deviant labeling), we consider a
number of additional analyses.
First, we examine the extent to which juveniles incarcerated
for even a relatively short period of time (one to two months) ever
return to school. We find that 62% of all children with a spell in
the school located in the detention facility never return to a CPS
school. Table VIII presents the results of more formal OLS and IV
analyses that show that juvenile incarceration significantly redu-
ces the likelihood that the student is observed in school one year
after the initial hearing (0.215 compared to a mean of 0.67).
This can be interpreted as either a change in the behavior of
the juvenile for whom catching up represents a significant barrier
to return, or an act on the part of the school actively discouraging
a previously incarcerated student from returning.
To explore this further, we examine whether students who do
return to a CPS school are more likely to transfer to another
school, which we interpret as evidence that the schools are treat-
ing incarcerated juveniles differently. In our data, only 28% of
juveniles who leave the school in the detention facility returned
to the same CPS school in which they were enrolled prior to in-
carceration, while 10% transferred to another school within the
CPS. We also considered whether incarceration increases the
likelihood of transferring to another CPS school after the initial
hearing (outside of the school located in the detention facility or in
the Cook County Jail). This analysis requires us to limit our
sample to the 18,195 juveniles whom we observe in high school
prior to juvenile incarceration, reducing our power considerably.
When we do, in the OLS regressions with a full set of controls we
find that juvenile incarceration is positively but only slightly
TABLE VIII
INTERMEDIATE SCHOOLING OUTCOMES: HIGH SCHOOL TRANSFERS

(1) (2) (3) (4) (5) (6)
Ever present Transferred to Ultimate
in CPS school another CPS transfer:
at least 1 year high school adult
after Initial in years after correctional
Dependent variable: hearing hearing facility

(0.0063) (0.069) (0.010) (0.243) (0.006) (0.060)
Mean of dependent variable 0.666 0.242 0.175
Observations 37,692 18,195 37,692
Notes. This table reports the relationship between juvenile incarceration and intermediate schooling
outcomes. ‘‘Ultimate transfer’’ means this is the last known whereabouts by CPS; the transferred high
school sample is restricted to students who were in high school at the time of the first juvenile court
hearing; ‘‘Ever present in a CPS school at least 1 year after initial hearing’’ excludes schools within the
detention facility or the Cook County Jail. All models include community weapons offense year-of-of-
fense fixed effects and full controls as listed in Table V. Standard errors are reported in the parentheses
and are clustered at the community level.
related to high school transfer in the years after the initial hear-
ing (0.055 relative to a mean of 0.242), but when we instrument
for juvenile incarceration, the effect is negative although very
imprecise (Table VIII).41 Given the imprecision it is difficult to
make strong claims, but we did not find evidence consistent with
juvenile incarceration leading to high school transfers. When we
consider transfers to an adult correctional facility, juvenile incar-
ceration is found to lead to significant increases in this outcome.
In a second analysis we examine whether juveniles who
spend time incarcerated are more likely to be classified as a spe-
cial education student on release. For this analysis we must limit
our sample to those for whom we can observe special education
status in the year after the initial hearing (79% of our sample,
n = 29,794). There is no relationship between juvenile incarcera-
tion and whether one is designated special education (Table IX),
but interestingly, incarceration is associated with a change in the
source of the disability. Those incarcerated are more likely to be
41. We also examined transfers to ‘‘alternative high schools’’ outside of the crim-
inal justice system, but only 1.5% are found to do so, and we do not find a statistically
significant relationship between juvenile incarceration and this (relatively rare)
outcome.
TABLE IX
INTERMEDIATE SCHOOLING OUTCOMES: SPECIAL EDUCATION STATUS

(1) (2) (3) (4) (5) (6)
Special education type observed in
years after initial hearing
Emotional/
behavioral Learning
Dependent variable: Any Special Education disorder disability

(0.004) (0.037) (0.003) (0.043) (0.004) (0.039)
Mean of dependent 0.193 0.082 0.085

variable
Observations 29,794
Notes. This table reports the relationship between juvenile incarceration and intermediate schooling
outcomes. It is restricted to students where the special education status is observed in years following the
hearing. ‘‘Any Special Education’’ includes categories not included in the two main categories reported
separately. All models include community weapons offense year-of-offense fixed effects and full controls
as listed in Table V. Standard errors are reported in the parentheses and are clustered at the community
level.
characterized as having an emotional or behavioral disorder and

equally less likely to be characterized as having a learning dis-
ability. The estimated effects are larger (in absolute magnitude)
in the IV regressions, though standard errors also increase (the
estimates remain significant at conventional levels). Again, the
reclassification could be due to changes in behavior on the part of
the juvenile or a change in the way that the school system labels
students who have been incarcerated.42
Another test involves looking at the impact of juvenile incar-
ceration on crimes for which it is reasonable to assume that being
labeled as having spent time incarcerated as a juvenile should
have little impact on arrest or incarceration: homicide and other
violent crimes. We found (Table VI) that juvenile incarceration
42. As noted previously, a limitation of the database is that data on suspensions

or disciplinary infractions, as well as truancy, do not extend back to our time period.
Nevertheless, outcomes based on these measures would also reflect actions by the
juvenile and school. Test score data are available, but assessing the impact on test
scores is complicated by the fact that so many of the juveniles never return to school,
and many of these juveniles managed not to take an exam in years when they are in
school. Those with a panel of test scores constitute a very selected sample of
juveniles.
does result in greater likelihood of being incarcerated as an adult

for these serious crimes, consistent with changes in criminal be-

havior on release.
In sum, these results are consistent with both changes in
juvenile behavior and changes in institutions’ treatment of juve-
niles explaining the large effects of juvenile incarceration on high
school completion and recidivism.
V.G. Additional Tests of Robustness

A concern when using judge fixed effects as instruments is
that the monotonicity assumption may fail.43 For example, some
judges could be particularly strict for only a subset of offenses,
such as violent crimes, and these judges could be relatively le-
nient for, say, property crimes. To investigate this possibility, we
categorized the juvenile offenses into four mutually exclusive
groups: violent, property, drug, and other. We found that judges
who are strict for violent crimes tend to be strict for other offense
types as well.44 Even so, we recalculate the instrument for each
judge offense type, thereby relaxing the monotonicity assump-
tion. The results (Online Appendix Table AIV) show similar
impacts for high school graduation and adult incarceration. We
take this as strong evidence that this potential failure of the
monotonicity assumption is not driving the main results. We
also find similar results when we allow the incarceration rate to
vary within judge but across cases with African American and
non–African American defendants.
43. Juvenile incarceration is monotonically increasing in the leave-out mean of

the judge’s incarceration rate, which provides some evidence that the monotonicity
assumption may be satisfied. Furthermore, we investigated whether treatment
effects differed across judges in an effort to estimate marginal treatment effects
(Heckman and Vytlacil 2005; Doyle 2008). We found that these estimates were too
imprecise to explore variation across judges. The point estimates suggested that the
negative effects for high school completion and adult incarceration were more likely
to stem from judges with higher incarceration rates, where the marginal juveniles
are likely to be cases with unobservable characteristics associated with a lower
likelihood of incarceration (compared to the margins relevant among the more le-
nient judges).
44. The relationship is not one to one, however, which is why it is useful to
estimate effects using the recalculated instrument. In particular, in a regression
of the judge’s violent crime incarceration rate on the judge’s property crime incar-
ceration rate within the usual fixed-effect cells, we find a coefficient of 0.84 (std.
err. = 0.10), for drug crimes the coefficient is 0.68 (std. err. = 0.11), and for other
crimes the coefficient is 0.64 (std. err. = 0.09).
As a second robustness check, we allow the fixed effects

within which juveniles are compared to vary. Specifically, we in-

clude fixed effects defined at the level of the community, commu-
nity year, community weapon, census tract, tract year,
tract weapon, and finally tract weapon year (Online
Appendix Table AV). The results are remarkably stable across
these different types of fixed effects.
In a third set of robustness checks, we change how we treat
transfers in our high school graduation analysis. Previously we
define graduation only if the records in the public school data
indicate graduation with certainty and those who transferred
(16% of the sample) were considered nongraduates. In Online
Appendix Table AVI, Panel A, we simply remove from the
sample all transfers for whom high school completion is not
known. The results are similar though slightly larger than the
main set of estimates. In Panel B, we keep the full sample but
define as the outcome an indicator equal to 1 if the student is one
of these 16% of transfers. The estimates are positive but not sig-
nificant. Last, we explore whether juvenile incarceration affects
transfers to a private school or to another school outside of
Chicago, which may result in better educational outcomes. We
find that juvenile incarceration reduces the likelihood of trans-
ferring in this way, although the estimates are not statistically
significant.45
We also consider the robustness of the results to changes in
calculation of our instrument. We drop cases with a missing judge
at the first hearing, trim the instrument of extreme values, and
estimate the instrument using a probit (Online Appendix Table
AVII), and the results are unchanged. We also explore the extent
to which the effects are driven by swing judges, who preside over
one fifth of our cases. When we exclude swing judges, the results
are unchanged.
V.H. Length of Stay

Last, we do not observe length of stay in the detention facil-
ity, but we do observe a proxy: the length of time spent in the high
school located in the facility. This is just a proxy because students
could drop out of high school, and they do not appear in our data
45. In 2SLS models with full controls, the estimated effect of juvenile incarcer-
ation on transfer to a private school is 0.019 (std. err. = 0.026); for transfer outside
of Chicago is 0.070 (std. err. = 0.036)
during the summer. Online Appendix Table AVIII shows that

length of stay is not related to our instrument: ‘‘strict’’ judges in

terms of incarceration rates are not associated with longer stays,
conditional on entering detention. The sign of the estimate sug-
gests shorter stays for those assigned to strict judges, consistent
with these judges incarcerating less serious offenders on average,
but the magnitude is small.46 Furthermore, the leave-out mean of
the judge’s length of stay is not predictive of a juvenile’s length of
stay when we control for the judge’s incarceration rate. Not sur-
prisingly, then, when we estimate the impact of length of stay
(including zeros) on our outcomes, with the instrument being
the judge’s leave-out mean length of stay in the year after initial
hearings, we find estimates that are similar to moving from
0 days to the mean number of days in the facility (Online
Appendix Table AVIII).
We considered one final specification—dividing cases into
short versus long stays (defined by the median in our sample)
and then estimating the effect of a short or long stay (compared
to no stay) on these outcomes. For this specification, our instru-
mental variables are the judge’s leave-out means of short and
long stays. While the point estimates for the recidivism outcome
suggest larger effects for longer stays, statistically the estimated
effects of the two stay lengths are the same due to large standard
errors. We conclude that we do not have the power to detect
whether length of stay matters in these data, and we interpret
our main results as the impacts of a typical stay in incarceration
(averaging 42 days according to our proxy) on future outcomes.
Future work exploring the impact of length of incarceration as a
juvenile on future outcomes is needed.
VI. Conclusions
Juvenile incarceration is expensive, with expenditures on ju-
venile corrections totaling $6 billion annually in the United
States, and the average annual (direct) cost of a incarcerating a
juvenile topping $88,000 (Mendel 2011). If juvenile incarceration
either enhanced human capital accumulation or deterred future
crime and incarceration, a trade-off could be considered.
46. Although imprecise, the point estimate suggests that a 2 standard deviation
increase in our instrument (0.08) is associated with 2 fewer days in detention,
compared to a mean of 100 days.
Rather, we find that for juveniles on the margin of incarceration,

such detention leads to both a decrease in high school completion

and an increase in adult incarceration. In exploring the mecha-
nisms behind these effects, we find that once incarcerated, a ju-
venile is unlikely to ever return to school, suggesting that even
relatively short periods of incarceration can be very disruptive
and have severe long-term consequences for this population.
Moreover, for those who do return to school, they are more
likely to be classified as having a disability due to a social or be-
havioral disorder, likely reducing the probability of graduation
even among those who do return to school and possibly increasing
the probability of future criminal behavior.
Our results imply that it may be welfare enhancing to use
alternatives to juvenile incarceration. Illinois has more recently
adopted an array of such policies, including electronic monitoring
and well-enforced curfews that serve as substitutes for juvenile
incarceration. These substitutes have been growing in popularity
across the United States. Our results suggest that their contin-
ued expansion have the potential to increase high school gradu-
ation rates and reduce the likelihood of adult crime. In addition to
reducing juvenile incarceration, policies that address the low
rates at which juveniles return to school on release by providing
additional support and resources for these at-risk juveniles may
also be effective in reducing the negative impact of incarceration
on human capital accumulation and other outcomes.
In contrast to the increasing adoption of alternatives to in-
carceration that reduce juvenile incarceration, many states have
adopted policies of increasing police presence in schools, which
has led to an increase in juvenile arrests for relatively mild in-
fractions. If this leads to an increase in juvenile detention, which
seems likely, then the continued expansion of this policy has the
potential to reduce high school graduation rates for those directly
affected.
To consider the full set of costs and benefits of policies affect-
ing juvenile arrest and incarceration, one must also consider the
potential reduction in crime due to the incapacitation effect of
incarceration as well as the deterrent effects of strict punishment
on the criminal activity of other youths. Regarding incapacita-
tion, to the extent that alternatives such as strict curfews or elec-
tronic monitoring also serve to incapacitate, this should be less of
a concern. Regarding deterrence, evidence suggests that juve-
niles’ criminal propensity is particularly inelastic with respect
to penalties (Lee and McCrary 2005), which implies that this may
be of second-order importance compared to the large decrease in

high school completion and increase in adult incarceration found
here. If this is the case, then the results suggest that a continued
move toward less restrictive juvenile sentencing would increase
human capital accumulation and lower the propensity of these
juveniles to become incarcerated as adults without an increase in
juvenile crime.
Brown University and NBER

MIT and NBER
Supplementary Material
An Online Appendix for this article can be found at QJE
online (qje.oxfordjournal.org).
References
Bayer, Patrick, Randi Hjalmarsson, and David Pozen, ‘‘Building Criminal Capital
Behind Bars: Peer Effects in Juvenile Corrections,’’ Quarterly Journal of
Economics, 124 (2009), 105–147.
Becker, Gary, ‘‘Crime and Punishment: An Economic Approach,’’ Journal of
Political Economy, 76 (1968), 169–217.
Bernburg, Jon Gunnar, and Marvin D. Krohn, ‘‘Labeling, Life Chances, and Adult
Crime: The Direct and Indirect Effects of Official Intervention in Adolescence
on Crime in Early Adulthood,’’ Criminology, 41 (2003), 1287–1318.
Bernburg, Jon Gunnar, Marvin D. Krohn, and Craig J. Rivera, ‘‘Official Labeling,
Criminal Embeddedness and Subsequent Delinquency: A Longitudinal Test
of Labeling Theory,’’ Journal of Research in Crime and Delinquency, 43
(2006), 67–88.
Bjorklund, Anders, and Robert Moffitt, ‘‘The Estimation of Wage Gains and
Welfare Gains in Self-Selection Models,’’ The Review of Economics and
Statistics, 69 (1987), 42–49.
Brennan, Patricia A., and Sarnoff A. Mednick, ‘‘Learning Theory Approach to the
Deterrence of Criminal Recidivism,’’ Journal of Abnormal Psychology, 103
(1994), 430–440.
Cameron, Stephen, and James J. Heckman, ‘‘Determinants of Young Male
Schooling and Training Choices,’’ NBER Working Paper No. 4327, 1993.
Chang, Tom, and Antoinette Schoar, ‘‘Judge Specific Differences in Chapter 11
and Firm Outcomes,’’ MIT Working Paper, 2008.
Civitas, ‘‘Comparisons of Crime in OECD Countries,’’ (2012), available at http://
www.civitas.org.uk/crime/crime_stats_oecdjan2012.pdf.
Cook, Philip, and Songman Kang, ‘‘Birthdays, Schooling and Crime: New
Evidence on the Crime-Dropout Nexus,’’ NBER Working Paper No. 18791,
2013.
Cunha, Flavio, James J. Heckman, and Lance Lochner, ‘‘Interpreting the
Evidence on Life Cycle Skill Formation,’’ in Handbook of the Economics of
Education, Eric A. Hanushek, and Finis Welch, eds. (Amsterdam: Elsevier,
2006).
Currie, Janet, and Erdal Tekin, ‘‘Does Child Abuse Lead to Crime?,’’ NBER
Working Paper No. 12171, 2006.
De Li, Spencer, ‘‘Legal Sanctions and Youths’ Status Achievement: A
Longitudinal Study,’’ Justice Quarterly, 16 (1999), 377–401.
Deschenes, Olivier, ‘‘Unobserved Ability, Comparative Advantage, and the Rising

Return to Ability in the US,’’ Working Paper, 2006.
Di Tella, Rafael, and Ernesto Schargrodsky, ‘‘Criminal Recidivism after Prison
and Electronic Monitoring,’’ Journal of Political Economy, 121 (2013), 28–73.

Dobbie, William, and Jae Song, ‘‘Debt Relief and Debtor Outcomes: Measuring
The Effects of Consumer Bankruptcy Protection,’’ Mimeo, Harvard
University, 2013.
Doyle, Joseph, ‘‘Child Protection and Adult Crime,’’ Journal of Political Economy,
116 (2008), 746–770.
Drago, Francesco, Roberto Galbiati, and Pietro Vertova, ‘‘The Deterrent Effects of
Prison: Evidence from a Natural Experiment,’’ Journal of Political Economy,
117 (2009), 254–270.
Forrest, Christopher B., Ellen Tambor, Anne W. Riley, Margaret E. Ensminger,
and Barbara Starfield, ‘‘The Health Profile of Incarcerated Male Youths,’’
Pediatrics, 105 (2000), 286–291.
Freeman, Richard B., ‘‘Crime and the Employment of Disadvantaged Youths,’’ in
Urban Labor Markets and Job Opportunities, George E. Peterson, and
Wayne Vroman, eds. (Washington, DC: Urban Institutes Press, 1992).
General Accounting Office, Growing Inmate Crowding Negatively Affects Inmates,
Staff, and Infrastructure (Washington, DC: GAO, 2012).
Glaeser, Edward, and Bruce Sacerdote, ‘‘Why Is There More Crime in Cities?,’’
Journal of Political Economy, 107 (1999), S225–S258.
Glaze, Lauren E., and Erika Parks, ‘‘Correctional Populations in the United
States, 2011,’’ Bureau of Justice Statistics Bulletin, NCJ 239972, 2012.
Goerge, Robert M., John Van Voorhis, and Bong Joo Lee, ‘‘Illinois’s Longitudinal
and Relational Child and Family Research Database,’’ Social Science
Computer Review, 12 (1994), 351–365.
Gottfredson, Don, Effects of Judges’ Sentencing Decisions on Criminal Careers
(Washington, DC: National Institute of Justice, 1999).
Granovetter, Mark, Getting a Job: A Study of Contacts and Careers (Chicago:
University of Chicago Press, 1995).
Green, Donald, and Daniel Winik, ‘‘Using Random Judge Assignments to
Estimate the Effects of Incarceration and Probation on Recidivism among
Drug Offenders,’’ Criminology, 48 (2010), 10–23.
Grogger, Jeffrey, ‘‘The Effect of Arrests on the Employment and Earnings of
Young Men,’’ Quarterly Journal of Economics, 110 (1995), 51–72.
Hazel, Neal, Cross-National Comparison of Youth Justice (London: Youth Justice
Board for England and Wales, 2008).
Heckman, James J., and Edward Vytlacil, ‘‘Structural Equations, Treatment
Effects, and Econometric Policy Evaluation,’’ Econometrica, 73 (2005),
669–738.
Hjalmarsson, Randi, ‘‘Criminal Justice Involvement and High School
Completion,’’ Journal of Urban Economics, 63 (2008), 613–630.
———, ‘‘Juvenile Jails: A Path to the Straight and Narrow or Hardened
Criminality?,’’ Journal of Law and Economics, 52 (2009), 779–809.
Imbens, Guido W., and Joshua D. Angrist, ‘‘Identification and Estimation of Local
Average Treatment Effects,’’ Econometrica, 62 (1994), 467–475.
Jacob, Brian, and Lars Lefgren, ‘‘Are Idle Hands the Devil’s Workshop?
Incapacitation, Concentration, and Juvenile Crime,’’ American Economic
Review, 93 (2003), 1560–1577.
Kashani, Javad H., George W. Manning, Donald H. McKnew, Leon Cytryn, John
F. Simonds, and Phil C. Wooderson, ‘‘Depression among Incarcerated
Delinquents,’’ Psychiatry Resources, 3 (1980), 185–191.
Kennelman, Tracey, Justice Expenditure and Employment Extracts 2009
(Preliminary), Bureau of Justice Statistics, NCJ 237913, 2012.
Kerley, Kent R., Michael L. Benson, Matthew R. Lee, and Francis Cullen, ‘‘Race,
Criminal Justice Contact, and Adult Position in the Social Stratification
System,’’ Social Problems, 51 (2004), 549–568.
Kling, Jeffrey, ‘‘Incarceration Length, Employment, and Earnings,’’ American
Economic Review, 96 (2006), 863–876.
Kling, Jeffrey, and John Tyler, ‘‘Prison-Based Education and Reentry into the
Mainstream Labor Market,’’ in Barriers to Reentry? The Labor Market for
Released Prisoners in Post-Industrial America, Shawn Bushway, Michael

Stoll, and David Weiman, eds. (New York: Russell Sage Foundation, 2007).
Kolesar, Michal, Raj Chetty, John Friedman, Edward Glaeser, and Guido Imbens,
‘‘Identification and Inference with Many Invalid Instruments,’’ NBER

Landerso, Rasmus, ‘‘Does Incarceration Length Affect Labor Market Outcomes
for Violent Offenders?,’’ The Rockwool Foundation Research Unit Study
Paper No. 39, 2012.
Lee, David, and Justin McCrary, ‘‘Crime Punishment and Myopia,’’ NBER
Levitt, Steven, ‘‘Juvenile Crime and Punishment,’’ Journal of Political Economy,
106 (1998), 1156–1185.
———, ‘‘Understanding Why Crime Fell in the 1990s: Four Factors that Explain
the Decline and Six that Do Not,’’ Journal of Economic Perspectives, 18
(2004), 163–190.
Lochner, Lance, ‘‘Education, Work, and Crime: A Human Capital Approach,’’
International Economic Review, 45 (2004), 811–843.
Lochner, Lance, and Enrico Moretti, ‘‘The Effect of Education on Crime: Evidence
from Prison Inmates, Arrests and Self-Reports,’’ American Economic Review,
94 (2004), 155–189.
Lott, John, ‘‘An Attempt at Measuring the Total Monetary Penalty from Drug
Convictions: The Importance of an Individual’s Reputation,’’ Journal of
Legal Studies, 21 (1992a), 159–187.
———, ‘‘Do We Punish High Income Criminals Too Heavily?,’’ Economic Inquiry,
30 (1992b), 583–608.
Luallen, Jeremy, ‘‘School’s Out . . . Forever: A Study of Juvenile Crime, At-Risk
Youths and Teacher Strikes,’’ Journal of Urban Economics, 59 (2006),
75–103.
Machin, Stephen, Oliver Marie, and Suncica Vujic, ‘‘The Crime Reducing Effect of
Education,’’ Economic Journal, 121 (2011), 463–484.
Maestas, Nicole, Kathleen Mullen, and Alexander Strand, ‘‘Does Disability
Insurance Receipt Discourage Work? Using Examiner Assignment to
Estimate Causal Effects of SSDI Receipt,’’ American Economic Review, 103
(2013), 1797–1829.
Mendel, Richard A., No Place for Kids: The Case for Reducing Juvenile
Incarceration (2011).
Murray, Charles, and Louis Cox, Beyond Probation: Juvenile Corrections and the
Chronic Delinquent (1979).
Na, Chongmin, and Denise C. Gottfredson, ‘‘Police Officers in Schools: Effects on
School Crime and the Processing of Offending Behaviors,’’ Justice Quarterly,
30 (2011), 619–650.
Neal, Derek, and Armin Rick, ‘‘The Prison Boom and the Lack of Black Progress
after Smith and Welch,’’ NBER Working Paper No. 20283, 2014.
Office of Juvenile Justice and Delinquency Prevention, Census of Juveniles in
Residential Placement 2010 (Washington, DC: OJJDP, 2014).
Peters, Clark, Thomas Haywood, Bong Joo Lee, and Ada Skyles Case Processing
and Services to Children in the Juvenile Justice Division of the Cook County
Juvenile Court (Chicago: Chapin Hall, 2002).
Raphael, Steven, and Michael Stoll Why Are So Many Americans in Prison? (New
York: Russell Sage Foundation, 2013).
Sampson, Robert, and John Laub, ‘‘A Life Course Theory of Cumulative
Disadvantage and the Stability of Delinquency,’’ Advances in
Criminological Theory, 7 (1997), 133–161.
Smith, Linda, and Ronald Akers, ‘‘A Comparison of Recidivism of Florida’s
Community Control and Prison: A Five-Year Survival Analysis,’’ Journal of
Research in Crime and Delinquency, 30 (1993), 267–292.
Spohn, Cassia, and David Holleran, ‘‘The Effects of Imprisonment on Recidivism
Rates of Felony Offenders: A Focus on Drug Offenders,’’ Criminology, 40
(2002), 329–357.
Stock, James, Jonathan Wright, and Motochiro Yogo, ‘‘A Survey of Weak
Instruments and Weak Identification in Generalized Method of Moments,’’
Journal of Business and Economic Statistics, 20 (2002), 518–529.
Sweeten, Gary, ‘‘Who Will Graduate? Disruption of High School Education by

Arrest and Court Involvement,’’ Justice Quarterly, 23 (2006), 462–480.
Tanner, Julian, Scott Davies, and Bill O’Grady, ‘‘Whatever Happened to
Yesterday’s Rebels? Longitudinal Effects of Youth Delinquency on

Education and Employment,’’ Social Problems, 46 (1999), 250–274.
Waldfogel, Joel, ‘‘The Effect of Criminal Conviction on Income and the Trust
Reposes in the Workmen,’’ Journal of Human Resources, 29 (1994), 62–81.
Walmsley, Roy, ‘‘World Prison Population List (Tenth Edition),’’ International
Centre for Prison Studies, London, 2013.
Western, Bruce, and Katherine Beckett, ‘‘How Unregulated Is the U.S. Labor
Market? The Penal System as a Labor Market Institution,’’ American
Journal of Sociology, 104 (1999), 1030–1060.
Western, Bruce, Jeffrey Kling, and David Weiman, ‘‘The Labor Market
Consequences of Incarceration,’’ Crime and Delinquency, 47 (2001), 410–427.
Wilson, David B., Catherine Gallagher, and Doris MacKenzie, ‘‘A Meta Analysis of
Corrections-Based Education, Vocation and Work Programs for Adult
Offenders,’’ Journal of Research in Crime and Delinquency, 37 (2000),
347–368.
This page intentionally left blank

Doyle 2015

Uploaded by

Copyright:

Available Formats

Doyle 2015

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Doyle 2015

Uploaded by

Copyright:

Available Formats

JUVENILE INCARCERATION, HUMAN CAPITAL,

AND FUTURE CRIME: EVIDENCE FROM RANDOMLY

Downloaded from https://academic.oup.com/qje/article/130/2/759/2330376 by Serials Department user on 09 October 2020

of correctional systems in 2011 (Glaze and Parks 2012).1 Thus,

Downloaded from https://academic.oup.com/qje/article/130/2/759/2330376 by Serials Department user on 09 October 2020

criminal activity, and labor market outcomes. These characteris-

Downloaded from https://academic.oup.com/qje/article/130/2/759/2330376 by Serials Department user on 09 October 2020

3. Some work finds that incarceration increases recidivism (Spohn and

because judges are randomly assigned to cases and is the ap-

Downloaded from https://academic.oup.com/qje/article/130/2/759/2330376 by Serials Department user on 09 October 2020

5. A somewhat related paper by Drago, Galbiati, and Vertova (2009) exploits a

decision.6 Unlike previous work, we use this strategy in a context

Downloaded from https://academic.oup.com/qje/article/130/2/759/2330376 by Serials Department user on 09 October 2020

Finally, we explore the potential mechanisms behind the es-

Downloaded from https://academic.oup.com/qje/article/130/2/759/2330376 by Serials Department user on 09 October 2020

When juveniles are charged with a crime in juvenile court,

Downloaded from https://academic.oup.com/qje/article/130/2/759/2330376 by Serials Department user on 09 October 2020

9. The district attorney is also assigned by calendar so that within a calendar

sentences are indeterminate in length, but typically last one to

Downloaded from https://academic.oup.com/qje/article/130/2/759/2330376 by Serials Department user on 09 October 2020

II.B. Mechanisms: How Juvenile Incarceration Can Affect

Forrest et al. 2000). Incarceration can also encourage the accu-

Downloaded from https://academic.oup.com/qje/article/130/2/759/2330376 by Serials Department user on 09 October 2020

unwilling to re-enroll juveniles who have been incarcerated. We

Downloaded from https://academic.oup.com/qje/article/130/2/759/2330376 by Serials Department user on 09 October 2020

III. Data Description

part of the disposition; rather, the sentences tend to be indeter-

Downloaded from https://academic.oup.com/qje/article/130/2/759/2330376 by Serials Department user on 09 October 2020

III.B. Sample Construction

III.C. Sample Description

17. All appendix material is in the Online Appendix.

Downloaded from https://academic.oup.com/qje/article/130/2/759/2330376 by Serials Department user on 09 October 2020

Sample size 440,797 29,141 8,551

incarcerated. The only characteristic along which all three groups

neighborhoods.18 They are also less likely to graduate high school

Downloaded from https://academic.oup.com/qje/article/130/2/759/2330376 by Serials Department user on 09 October 2020

We anticipate little bias to be introduced by this form of sample

Downloaded from https://academic.oup.com/qje/article/130/2/759/2330376 by Serials Department user on 09 October 2020

Despite the inclusion of an increasingly comprehensive set of

Downloaded from https://academic.oup.com/qje/article/130/2/759/2330376 by Serials Department user on 09 October 2020

juvenile’s first case. The instrument, which is defined for each

Downloaded from https://academic.oup.com/qje/article/130/2/759/2330376 by Serials Department user on 09 October 2020

Here, nj(i) is the total number of cases seen by judge j; k in-

IV.C. Judge Variation

Downloaded from https://academic.oup.com/qje/article/130/2/759/2330376 by Serials Department user on 09 October 2020

Downloaded from https://academic.oup.com/qje/article/130/2/759/2330376 by Serials Department user on 09 October 2020

First, we have confirmed with court personnel that judges are

Downloaded from https://academic.oup.com/qje/article/130/2/759/2330376 by Serials Department user on 09 October 2020

Middle vs. Top vs.

P(Juvenile incarceration j X) 0.219 0.221 0.220 (.251) (.516)

the offense, despite significant differences in the incarceration

same (0.22) regardless of judge type.25 Table II reports results for

Downloaded from https://academic.oup.com/qje/article/130/2/759/2330376 by Serials Department user on 09 October 2020

V.B. First Stage: Judge Assignment and Juvenile Incarceration

25. The propensity is predictive of incarceration, however. The predicted pro-

Downloaded from https://academic.oup.com/qje/article/130/2/759/2330376 by Serials Department user on 09 October 2020

incarceration rate would imply an increase in the likelihood of

Downloaded from https://academic.oup.com/qje/article/130/2/759/2330376 by Serials Department user on 09 October 2020

V.C. Juvenile Incarceration and High School Completion