“Something Works” in U.S.

Jails:
Misconduct and Recidivism Effects of the IGNITE Program∗
Marcella Alsan† Arkey Barnett‡ Peter Hull§ Crystal S. Yang¶

March 29, 2024

Abstract
A longstanding and influential view in U.S. correctional policy is that “nothing works”
when it comes to rehabilitating incarcerated individuals. We revisit this hypothesis by
studying an innovative law-enforcement-led program launched in the county jail of Flint,
Michigan: Inmate Growth Naturally and Intentionally Through Education (IGNITE).
We develop an instrumental variable approach to estimate the effects of IGNITE expo-
sure, which leverages quasi-random court delays that cause individuals to spend more
time in jail both before and after the program’s launch. Holding time in jail fixed, we
find that one additional month of IGNITE exposure reduces within-jail misconduct by
49% and reduces three-month recidivism by 18%, with the recidivism effects growing
over time. Surveys of staff and community members, along with administrative test
score records and within-jail text messages, suggest that cultural change and improved
literacy and numeracy scores are key contributing mechanisms.

∗
We thank Mauricio Cáceres Bravo, Jennifer Doleac, Sara Heller, Adriana Lleras-Muney, Matthew Mizel, Michael
Mueller-Smith, Matthew Pecenco, Evan Rose, Nico Rotundo, Yotam Shem-Tov, and seminar participants at Harvard
University, Opportunity Insights, the University of Connecticut, the University of Michigan, the NBER Economics
of Crime Working Group, and the University of California Los Angeles for helpful comments. We are grateful to
Meghan Beal at the National Sheriffs’ Association, Superintendent Mickie Kujat and Principal Brad Basista of Mt.
Morris Consolidated Schools, and to the administration and incarcerated individuals of Genesee County and Saginaw
County Jails—especially Deputy Annie Bueche, Undersheriff Mike Gomez, Major Jason Gould, Lieutenant Ebony
Rasco, and Sheriff Christopher Swanson. We gratefully acknowledge funding from Arnold Ventures. Xingyou Ye,
Miguel Purroy, Meera Mody, Anne Fogarty, Jeongmook Lim, and Michelle Wu provided expert research assistance.
†
Harvard Kennedy School and NBER. Email: marcella [email protected]
‡
University of Michigan. Email: [email protected]
§
Brown University and NBER. Email: peter [email protected]
¶
Harvard Law School and NBER. Email: [email protected]
1 Introduction
The US has one of the highest incarceration rates in the world, with over two million people in
correctional facilities daily (Zeng, 2022). Over 600,000 of these individuals are held in local jails,
the vast majority of whom are unconvicted or awaiting sentencing (Sawyer and Wagner, 2023).
These numbers partly reflect high rates of recidivism: one in four individuals released from jail, for
example, are re-jailed within the same year (Jones and Sawyer, 2019). Addressing such incarceration
cycles, and reducing recidivism more broadly, remain persistent policy challenges (Doleac, 2023).
Views on the effectiveness of rehabilitation in US prisons have generally been negative and slow
to change since the influential “nothing works” doctrine, often attributed to Martinson (1974).
From a review of observational studies of prison rehabilitation programs in the 1970s, Martinson
concluded that with “isolated exception” there was no “appreciable effect on recidivism.” While
this conclusion has been challenged, with more optimistic recent findings on the efficacy of certain
rehabilitative programs for certain populations (e.g., Weisburd, Farrington and Gill 2017; Heller
et al. 2017; Arbour, Marchand and Lacroix 2023), the “nothing works” doctrine became mantra
for much of the U.S. correctional community. In the years following the Martinson report, U.S.
correctional policy largely shifted away from principles of rehabilitation to a focus on deterrence
and incapacitation (Andrews and Bonta, 2010), with rehabilitative programs receiving even less
investment within U.S. jails—where resources are scant and stays are presumed short.
Outside the U.S., however, rehabilitative programming has become a mainstay of incarceration
and is often seen as crucial for reintegrating incarcerated individuals into society. Correctional
policy in Norway, for example, bases rehabilitation efforts around the “principle of normality”: that
life inside a correctional facility should resemble life outside as closely as possible. A recent quasi-
experimental analysis finds that time spent in such facilities leads to large reductions in recidivism
and other adverse outcomes (Bhuller et al., 2020). Inspired by these principles of rehabilitation,
some U.S. cities have begun to incorporate these ideas into the design of their correctional facilities.
For example, in 2020, a medium-security state prison located just outside of Philadelphia established
a small housing unit known as “Little Scandinavia,” originally available to six individuals serving
life sentences (Strange, 2023). But whether such rehabilitative policies and philosophies can work
in other contexts—and particularly in U.S. jails—remains an open question.
This paper studies an innovative law-enforcement-led rehabilitation program, launched in Septem-
ber 2020 in the county jail of Flint, Michigan: Inmate Growth Naturally and Intentionally Through
Education (IGNITE). Nominally, IGNITE is an educational program which offers tailored course-
work and training to all jailed individuals with high takeup rates. In practice, however, IGNITE
administrators emphasize a cultural change in the jail that goes well beyond coursework and embod-
ies a rehabilitative philosophy not unlike Norway’s principle of normality. Administrators say, for
example, that IGNITE is “much more than giving people a free education. It’s about giving people
hope when they have no hope.” (Barrett and Greene, 2023). At the same time, a notable difference
with the Norwegian experience—besides the U.S. jail context—is the program’s cost: IGNITE is
largely funded with existing county resources and staff, avoiding the kinds of large spending Norway

1
and other countries have used to launch rehabilitative systems.1 This low cost and perceptions of
broad success have recently led the National Sherrifs’ Association to begin scaling-up programs
similar to IGNITE in many jails across the U.S.2
To estimate the effects of IGNITE exposure, we leverage unique administrative data and a
novel instrumental variable (IV) approach based on idiosyncratic delays in court appointments.
Court delays are common for jailed individuals and can significantly extend their time spent in
jail. In our setting, District Court delays appear conditionally as-good-as-randomly assigned and
extend time in jail by around two weeks (27%) on average, both before and after the launch
of IGNITE. We use this variation to instrument for the time a jailed individual is exposed to
IGNITE, accounting for any baseline (i.e., non-IGNITE) effects of increased jail time, via a two-
treatment IV specification. Effectively, this specification differences post- vs. pre-IGNITE IV
estimates to isolate the marginal effect of IGNITE exposure while holding fixed time in jail and
netting out any potential direct effects of delays. We formalize the key new assumption underlying
this “difference-in-IVs” approach and develop graphical diagnostics akin to standard “pre-trend”
checks in conventional difference-in-differences strategies.3 We also contrast the IV independence,
exclusion, and monotonicity assumptions in our approach with those underlying more conventional
“judge IV” designs (employed, e.g., in Bhuller et al. 2020), which appear less tenable in our context.4
We find that exposure to IGNITE dramatically reduces an individual’s propensity for both
within-jail misconduct and post-release recidivism. One additional month of exposure to IGNITE
is estimated to reduce the number of weekly major misconduct incidents by 0.16 (49%) and to
reduce three-month recidivism by 8 percentage points (18%). These effects are similar across
different demographic groups, prior offense status, and predicted exposure to the Flint water crisis.
Estimated recidivism effects grow over time—to around a 15 percentage point reduction in one-
year recidivism—and are concentrated among individuals with high predicted recidivism risk. In
economic terms, we find that one additional month of IGNITE exposure reduces the three-month
social cost of crime post-release by at least $3,000 per incarcerated individual. Over a year, the
social cost of crime reduction is at least $5,600 per person-month.
These main findings are robust to a number of potential threats to our IV strategy. Notably, we
find that instrument compliers are similar on a wide range of observable characteristics before and
after the launch of IGNITE, supporting our interpretation of difference-in-IV estimates as effects
of the program itself. Other robustness checks probe the more standard identifying assumptions of
1
For example, Norway spends around $93,000 each year per prisoner in its system (Beaumont, 2023) while Genesee
County Jail spent around $70 per incarcerated individual per day in 2021 (Finley, 2019) or around $25,500 per year.
2
As of February 2024, county jails in eleven states have adopted IGNITE and jails in seven more states have
begun plans for adoption; see NSA (2024).
3
Importantly, our IV strategy does not leverage a conventional “parallel trends” restriction on untreated potential
outcomes. Instead, we leverage the as-good-as-random assignment of court delays and an assumption that delay-
induced time-in-jail effects would have been similar after the start of IGNITE if not for the program’s launch.
4
For example, our monotonicity assumption requires that court delays weakly increase an individual’s time in
jail—a likely more plausible condition than the often-critiqued assumption of monotonic decision-making across
heterogeneous judges (Mueller-Smith, 2015; Frandsen, Lefgren and Leslie, 2023). Empirical tests of conventional
judge IV monotonicity and exclusion restrictions reject decisively in our setting.

2
as-good-as-random instrument assignment, exclusion, and monotonicity, and show that our find-
ings are not driven by changing conditions from the COVID-19 pandemic or changes in reporting
behavior. We also obtain qualitatively similar (though generally less precise) estimates from more
conventional difference-in-differences and judge IV strategies, despite these alternative strategies
relying on different (and arguably less tenable) assumptions and using less complete or fine-grained
data. An alternative difference-in-IVs strategy that uses a neighboring county as a “control” group
in the post-IGNITE era, instead of our baseline pre-post comparison in the treated county, also
yields similar recidivism effect estimates. We further obtain similar estimates from a “double”
difference-in-IVs specification, which combines cross-county and over-time comparisons to relax
our baseline identifying assumptions.
We explore two primary drivers of these large misconduct and recidivism effects. First, we show
that exposure to the formal educational programming in IGNITE likely led to substantial improve-
ments in literacy and numeracy scores among incarcerated individuals. Comparing standardized
test scores before and after enrollment in this programming, we find that individuals gained a full
grade level, on average, in both math and reading from low baseline levels. While some of these
gains may reflect heightened attention or improved test-taking, rather than human capital accu-
mulation per se, they are massive even when compared to some of the most effective educational
interventions documented in recent quasi-experimental literatures (e.g., Cohodes and Roy, 2023).
Second, we deploy surveys to several stakeholders—including Flint community members, the
formerly incarcerated, and current Genesee County custody staff—to assess the extent of cultural
change alongside formal educational programming. We find that individuals who had personally
been exposed to IGNITE or have relatives who were exposed are 23 percentage points (70%) more
likely to view law enforcement favorably, suggesting a positive spillover effect from jail experience
to perceptions about police. Furthermore, among those who were incarcerated or who had close
contacts with the incarcerated, exposure to IGNITE is associated with a higher likelihood of en-
gagement in positive social activities (though this effect is not statistically significant). A sentiment
analysis of administrative data collected from text messages sent from the incarcerated individuals
to jail staff supports the survey findings: IGNITE-exposed incarcerated individuals are more likely
to use words categorized as positive and associated with trust than those incarcerated before IG-
NITE. Custody officers who interact more with IGNITE participants are also twice as likely to view
educational programs for incarcerated individuals as worthwhile compared to officers that do not
regularly interact with IGNITE participants. Taken together, these qualitative analyses support
the view that a widespread cultural change occurred for both participants and staff.
Broadly, our findings suggest that “something works” when it comes to the rehabilitation of
incarcerated individuals in U.S. jails. In fact, we find that IGNITE generates recidivism reductions
comparable to or larger than a range of rehabilitative programs in varied settings and countries,
including Norway (e.g., Heller et al. 2017; Mastrobuoni and Terlizzese 2022; Arbour, Marchand and
Lacroix 2023; Shem-Tov, Raphael and Skog 2022; Bhuller et al. 2020). Notably, the recidivism re-
duction from IGNITE is similar to the impact of cognitive behavioral therapy (CBT) administered

3
to juvenile arrestees (Heller et al. 2017), despite IGNITE serving high-risk adults whose behavior
is arguably less malleable. The effectiveness of IGNITE thus demonstrates that rehabilitative prin-
ciples can be successfully implemented even within a county jail in one of the most disadvantaged
cities in America: Flint, Michigan, which has been described as a once prosperous city “devastated
by global economic forces, population loss, racism, disinvestment, and breakdowns in accountabil-
ity at multiple levels of government” (Leiser, Wang and Tatum III, 2022). Despite these inherent
challenges, we show that a relatively low-cost and law-enforcement-led program resulted in better
outcomes for incarcerated individuals and far-reaching improvements in public safety.
Our analysis contributes to a large literature studying the impact of various interventions on
crime and recidivism. In particular, we add to a growing body of work documenting beneficial
effects of rehabilitative programming (described above), diversion from the criminal justice system
itself (e.g., Mueller-Smith and Schnepel 2021; Augustine et al. 2022), specialized criminal courts
(e.g., Golestani, Owens and Raissian 2024), improvements in prison conditions (e.g., Tobón 2022),
and alternatives to incarceration (e.g., Di Tella and Schargrodsky 2013; Lee 2023; Henneguelle,
Monnery and Kensey 2016; Williams and Weatherburn 2022). Our paper also relates to work
documenting the impact of education on crime more broadly (e.g., Lochner and Moretti 2004;
Lavecchia, Oreopoulos and Spencer 2024), as well as work studying the effects of health care,
employment, and other programs on crime.5 In addition to conventional recidivism outcomes, we
estimate effects on within-facility misconduct—adding to a small but growing quasi-experimental
literature with access to such outcomes (Arbour, Marchand and Lacroix 2023; Bravo 2024).
Methodologically, we contribute a new IV strategy that leverages administrative delays which
extend an individual’s time exposed to an institution before and after a policy reform. Like Ab-
dulkadiroğlu et al. (2016) and Autor et al. (2017), we use a two-treatment IV model to isolate
the causal effects of interest via quasi-experimental shocks. Our approach is closest to Abdulka-
diroğlu et al. (2016) in that we combine cross-sectional shocks with variation in potential policy
exposure over time; it differs from Autor et al. (2017)’s and those in other papers studying the
effects of delays or administrative interruptions themselves (e.g., Yang 2016; Iverson 2018; Ho,
Hamilton and Roos 2000) in that we use quasi-experimental delay shocks as an instrument for the
policy’s exposure. We develop several diagnostic tools and extensions of this “difference-in-IVs”
approach, which may be fruitfully applied both within and outside of criminal justice settings. We
pair this quasi-experimental approach with a series of qualitative analyses to help explore possible
mechanisms, in the same mixed-methods spirit as Bergman et al. (Forthcoming).
The remainder of this paper is organized as follows. Section 2 details the institutional setting.
Section 3 describes data sources and the analysis sample. Section 4 develops our IV strategy.
Section 5 presents the main results and extensions. Section 6 contextualizes our findings and
explores possible mechanisms. Section 7 concludes.
5
See, e.g., Packham and Slusky (2023); Raphael and Weiman (2007); Sabol (2007); Yang (2017); Bhatt et al.
(2024); Deza, Maclean and Solomon (2022); Bondurant, Lindo and Swensen (2018); Jácome (2020); Tuttle (2019);
Deshpande and Mueller-Smith (2022); Arenberg, Neller and Stripling (2024); Darolia, Mueser and Cronin (2021);
Tyler and Kling (2006) and Bailey et al. (2020).

4
2 Institutional Setting
2.1 Genesee County Jail and Court System
IGNITE was launched September 2020 in Genesee County Jail, which holds individuals primarily
from the surrounding city of Flint, Michigan (see Appendix Figure A1). Flint is a majority-Black
city with around one-third of households living in poverty. It has experienced several major crises
in recent years, including the Flint Water Crisis and multiple cases of financial mismanagement.
Flint also consistently has one of the highest crime levels among U.S. cities, with a homicide rate
exceeding seven times the national average (Stebbins, 2021).
As with the approximately 3,000 U.S. jails across the country, Genesee County Jail primarily
holds three groups of individuals: (1) arrested individuals who are being detained before trial (rep-
resenting over 90% of the jail population), (2) convicted individuals who are awaiting sentencing,
and (3) sentenced individuals with incarceration time of less than one year. Jail populations in
Genesee County and across the U.S. are overwhelmingly male, young, and non-white.6 Incarcer-
ated individuals are also much more likely to lack a high school degree compared to the general
population; a 2014 prison study found that 72% lacked literate proficiency compared to 52% of
U.S. households (NCES, 2014). Time spent in jail has increased over time, with the national mean
length of stay rising over the last decade from 22.7 days to 32.8 days (Zeng, 2022). Court delays,
described further below, are a primary reason for longer jail spells.
Individuals’ first point of contact with Genesee County Jail occurs shortly after their arrest. On
the basis of arrest charges and other considerations, a prosecutor decides whether to file criminal
charges. At this point, the case formally enters into the court system and follows a particular
sequence of required events. The typical flow of a case through the Genesee County Court Sys-
tem is shown in Appendix Figure A2. Defendants usually start their case in the District Court,
which handles all initial arraignments, probable cause conferences, and preliminary examinations,
with cases assigned to a particular court based on location of arrest (67th District Court, 2022;
Supreme Court, 2023). Misdemeanor offenses (less serious crimes that usually carry a maximum jail
term of one year) proceed in the District Court through the trial, plea, and sentencing processes.
In more serious felony cases with sufficient evidence, the case is “bound over” (i.e., transferred) to
the Circuit Court, which handles the pretrial, trial, plea, and sentencing processes. The District
Court and Circuit Court can also reconsider an individual’s bond amount and decide whether to
release them on electronic tether.
Before IGNITE, Genesee County Jail had very limited educational programming for incarcer-
ated individuals. The jail only offered a GED class to a small group of selected individuals through
a local school providing adult education (Mt. Morris Consolidated Schools). Such limited pro-
gramming is typical of U.S. jails.7 As described by one correctional administrator in 2019: “we [in
6
In 2021, for example, 87% of those incarcerated in jails nationally were men, 52% were under the age of 35, and
51% were non-white or Hispanic (Zeng, 2022).
7
For example, according to Harlow (2003), only 60% of U.S. jails reported any educational or training program-
ming in its last census (1999) and the quality and accessibility of that programming varies substantially and often

5
Genesee County Jail] were just kind of functioning.... We were sending people to court, sending
people to prison, getting people out” (Barrett and Greene, 2023).

2.2 IGNITE
In the wake of the murder of George Floyd in May 2020, and the elevated racial tensions that
followed in Flint and other parts of the country, the Genesee County Sheriff launched IGNITE: a
new jail education program which was available to nearly all incarcerated individuals with incentives
for participation. The stated mission of this program is to reduce recidivism and end the cycle of
generational incarceration through education. Here we summarize key features of the program and
its launch; Appendix B gives further institutional details.
Since its September 2020 launch, IGNITE has relied on repurposed jail space and staff. For
instance, the day room in Genesee County Jail was transformed into a large classroom used by
different groups of incarcerated individuals at different times of the day (see Appendix Figure
A3). Staffing consisted of two full-time deputies, who oversaw the day-to-day operations of the
program, and a GED teacher from the nearby Mt. Morris Consolidated Schools who acted as a
circulating educator. The jail installed two dedicated WiFi networks for access to the internet and
initially purchased 600 tablets for IGNITE participants. Program costs were largely covered by
revenue generated from the use of commissary tablets, which participants could use in off-hours to
purchase and access games and puzzles. Consequently, the county budget for correctional services
did not substantially change with IGNITE’s launch (see Appendix Figure A4). Both before and
after IGNITE, Genesee County Jail spent around $70 per individual-day (Finley, 2019).
In addition to being available to nearly all incarcerated individuals (with the exception of
those deemed medically unstable or who were immediately released without charge), IGNITE
has three distinguishing features. First, instruction is tailored to each individual based on their
educational background and baseline testing (see Appendix Figure A5 for examples). Participants
are enrolled in class five days per week for two hours per day and are placed at intake into short-
term, medium-term, or long-term coursework depending on their predicted length of stay. Some
incarcerated individuals work on basic literacy while others work towards completing their GED and
others complete programs for college credit. Individuals work on Chromebooks, allowing for more
personalized instruction and for educators to float around the classroom, monitoring progress and
answering questions. IGNITE also offers additional technical programming, including certification
for food handling, commercial driving, masonry, and welding training. The program regularly hosts
graduation ceremonies (see Appendix Figure A3), where incarcerated individuals celebrate a new
diploma, course completion, or job certification in cap and gown alongside family and friends.
Second, participation in IGNITE is incentivized and takeup rates are high (around 90%, per
administrative data detailed below). Educational programming occurs during two dedicated hours
of instruction woven into the daily schedule (see Appendix Figure A6). During instruction time, all
depends on the discretion of jail administrators. The programming itself is generally carried out in a small classroom
of dedicated space with capacity and staffing constraints.

6
other jail activities cease. Non-participating individuals remain in their cells, while participating
individuals receive tablets to access educational programming, which they can also use to access
approved entertainment during non-IGNITE hours.
Finally, in addition to providing a wide range of education, IGNITE was intended to launch a
meaningful cultural change for both incarcerated individuals and correctional officers. Correctional
officers were asked to facilitate a learning environment by treating incarcerated individuals as
students capable of change and growth. This perspective shift was felt immediately, with one
jail administrator noting that the start of IGNITE represented a “shock to the jail culture” with
officers saying, “We’re doing what? We’re bringing in teachers? We’re providing tablets? Are
you kidding?” (Barrett and Greene, 2023). In the post-IGNITE period, incarcerated individuals
were described as not just waiting for court dates but as anticipating a productive life post-release
because of their participation in the program. Correctional staff also expressed new views of
incarcerated individuals. At a graduation ceremony, a correctional officer recalls holding the door
for the graduating individuals and shaking their hands, stating that “It really humanizes people...It
humanizes the inmate population, and it humanizes the deputy population.” (Barrett and Greene,
2023). We return to this idea of cultural change using original survey data in Section 6.3.

2.3 Court Delays

We leverage administrative delays in the Genesee County court system to estimate the effects of
IGNITE exposure. Genesee County, as in many other parts of the U.S., routinely experiences court
delays and backlogs that can cause individuals to spend many months or even years in jail waiting
for cases to be adjudicated. In addition to these routine delays, the start of policy responses to the
COVID-19 pandemic in March 2020 and the subsequent spread of different variants exacerbated
delays and resulted in major court closures which suspended trials indefinitely in both Genesee
County. Administrative court delays contribute to lengthy jail spells. Among arrested individuals,
the mean length of stay in Genesee County Jail is 1.5 months with a long right tail: the 90th
percentile of time in jail is around four months.
Court delays stem from numerous opportunities for rescheduling, as shown in Appendix Figure
A2. Red arrows show the primary court hearings that can be rescheduled in the District Court
(which primarily handles initial hearings and misdemeanors) while purple arrows denote additional
opportunities for court delays that can occur for felony cases in the Circuit Court. Because these
court hearings occur prior to a finding of guilt or innocence, delays in the timing of these events
will primarily affect individuals who are incarcerated pretrial and awaiting case disposition. In
practice, these delays are common and highly impactful for an individual’s time in jail. One jail
administrator notes: “if you think about the people that are in jail, they expect to go to court.
But it gets adjourned, they get another court date, it gets adjourned, another court date, it gets
dismissed, they have to reissue a warrant, and they never leave jail” (Diaz, 2020). Such anecdotes
align with patterns observed in our data where nearly 40% of all scheduled court dates are delayed,
mostly by the District Court.

7
 As we show below in Section 4.3, court delays appear largely idiosyncratic among individuals
assigned to the same court during similar time periods and facing similar charges. Anecdotally,
most delays are due to changes in the schedule of the judge or prosecutor assigned to the case and
rarely occur at the request of the defendant or defense attorney.8 While our data do not usually
provide a rationale for each observed delay, we provide evidence below that incarcerated individuals
send internal messages to jail administrators asking when they will next appear in court with a
greater frequency when there are court delays. This pattern is consistent with anecdotal evidence
that most delays are not caused by the defendant. We perform several robustness checks in Section
5.2 that narrow in on sources of delay which are more likely to be court-induced. We also use the
fact that delays are similarly common in neighboring Saginaw County to conduct placebo checks.

3 Data and Sample Construction

Our analysis of IGNITE leverages administrative data from several sources along with original
surveys of the local community, formerly incarcerated individuals, and correctional staff. This
section describes each data source and key variables; Appendix C gives additional details.

3.1 Data Sources and Key Variables

The Jail Management System (JMS) and Recidivism. The JMS is a comprehensive elec-
tronic database, used by Genesee County Jail and other Michigan jails in Michigan since 2015, that
tracks incarcerated individuals from booking to release. We obtained JMS data for both Genesee
County Jail and neighboring Saginaw County Jail from January 2015 to May 2023. The JMS data
include demographic information (age, name, race, home address), the arrest date, the booking
date, the release date, charges, and case disposition outcomes for each incarcerated individual.
Data from Genesee County also include the arrest location, which we use to identify the specific
District Court that handles the case. We use these data to construct our primary outcome of recidi-
vism, defined as whether an individual is rebooked in jail over a given time period.9 Our baseline
recidivism outcome is measured over three months, though we study recidivism for up to one year
post-release. Although 90% of recidivism occurs within the same county (Yang, 2017; Raphael
and Weiman, 2007; Sabol, 2007; Schnepel, 2018; Alper, Durose and Markman, 2018), we define
recidivism as rearrest in either Genesee or Saginaw Counties to allow for mobility. In practice,
results are virtually identical when we restrict to only Genesee County. We also check robustness
to alternative measures of recidivism based on an individual being recharged or reconvicted.

The District and Circuit Court Register of Actions (ROA) and Court Delays. We collect
ROAs from District and Circuit Courts in Genesee and Saginaw Counties by scraping publicly-
8
Delays can also stem from changes in courtroom availability. For example a Genesee County District Court
courtroom closed temporarily in 2023 because of a sewage leak (Jeltema, 2023).
9
Officers that make arrests in the community are different from custody officers that run IGNITE within the jail.
See Appendix C for more details.

8
available online case management systems. The ROAs represent permanent case histories of all
hearings and events during an individual’s case. These data include information on defendant
charges, activities, proceedings, and filings for the case, along with dates and times of new court
appointments, presiding judges, and notices of adjourned or rescheduled appointments. We use
these records to create a comprehensive timeline of court hearings for each incarcerated individual.
We identify court delays by changes to scheduled hearings that result in them being “removed from
the calendar.” Appendix Figure A7 shows an example ROA with such an identified delay. Our
baseline specification uses an indicator for any District Court delay as an instrument, though we
consider robustness to several other instrument specifications.

Jail Incident Reports and Misconduct. We use Jail Incident Report data from Genesee County
Jail to capture within-jail misconduct and medical events for incarcerated individuals. Misconduct
is categorized as either major or minor. Examples of major misconduct include threatening an-
other with bodily harm, introducing contraband, violence and disruption, and refusing to follow
instructions. Examples of minor misconduct include disorderly conduct, being in an unauthorized
area, possession of unauthorized items, and lying.10 Medical events include suicidality and suicide
attempts. We observe each incident date and the name of the involved individuals. Our main anal-
ysis uses major misconduct as a primary outcome of interest, as this is both more consequential for
staff and incarcerated individuals and also less likely to suffer from misreporting. We study minor
misconduct and medical outcomes as secondary outcomes.

The Kites Electronic Message System. Kites is an electronic internal messaging system be-
tween jail administrators and incarcerated individuals in Genesee County. This messaging system
is available to all incarcerated individuals via a kiosk in the jail. We observe the content of all sent
messages along with the sender’s identity, the date of the message, and any follow-up responses.
Incarcerated individuals can send messages to request services from numerous individuals, includ-
ing food service vendors, commissary services, medical personnel, and other administrative staff.
Incarcerated individuals also often send questions about their case, including inquiries about court-
initiated delays and when they will be released (see Appendix Figure A8). We use these data to
assess the reaction of incarcerated individuals to court delays. We also study message sentiment to
explore possible mechanisms.

Mt. Morris Educational Data. We obtain administrative data from Mt. Morris Consolidated
Schools, which contain date- and time-stamped course advancement and completion records from
2021 onwards. The data also include pre- and post-instruction test scores from September 2020 to
October 2023. Test scores are from Comprehensive Adult Student Assessment Systems (CASAS)
exams in math and reading. Pre-assessments are administered to all IGNITE participants in order
to place them in appropriate educational programming. Once enrolled in IGNITE, incarcerated
10
Complaints of misconduct are investigated internally. If the investigator determines charges should be filed, they
send the potential charges to a prosecutor who decides whether or not to charge the individual. If the individual is
charged and sentenced they may have their ongoing jail time lengthened, though this is rare in practice.

9
individuals take post-assessments every 40 hours of completed instruction in order to measure
their progress. In practice, we do not have post-assessments for all participants since tests are not
completed if an individual is discharged without sufficient lead time to inform staff. In addition,
electronic testing was only made available recently. In the end, only a few hundred paired pre-
and post-assessments are available for analysis. Nevertheless, we find that incarcerated individuals
with these assessments are representative of the jail population (see Appendix Table A2).

ViaPath Data. ViaPath is the internet service provider for Genesee County Jail and supplies con-
nectivity for the Chromebook-based educational and tablet-based entertainment content available
to IGNITE participants. In addition, ViaPath maintains logs of the amount of time incarcerated
individuals spend in video calls and telephone calls to individuals outside of the jail. We have access
to all ViaPath data, including individual identifiers, from February 2021 onward. We link these
records to JMS data to determine the rate of IGNITE participation from tablet use. This exercise
shows that 90% of individuals incarcerated on or after February 2021 participated in IGNITE.11

Community Survey. We conducted the Flint Community Survey in December 2023. Community
members and two ministers of local churches distributed the survey to ensure it would not be
influenced by IGNITE administrators. The survey was anonymous and asked respondents about
their own experience being incarcerated in Genesee County Jail or the experiences of a close friend or
family member. Importantly, the survey did not mention IGNITE in these questions. Respondents
received a $25 restricted-use Walmart eGift card for completing the survey; the overall response rate
was 87%. We construct our main exposure variable as an indicator for the respondent being held
in Genesee County Jail after IGNITE was introduced and find that participants are well-balanced
across this measure (see Appendix Table A3). The primary survey outcome is a respondent’s
perception of local law enforcement (this was elicited from all respondents). Secondary outcomes
elicited from those with direct or indirect jail experience include their level of hopefulness for the
future and their participation in positive activities (i.e., employment, education, or caregiving).
The recruitment flyer is shown in Panel A of Appendix Figure A9. A link to the full survey and
further details are given in Appendix D.

Custody Staff Survey. We administered The Genesee County Jail Custody Staff Survey to all
current staff in January 2024. The purpose of the survey was to assess staff views towards incar-
cerated individuals in general, rehabilitation programs and educational opportunities in particular,
and overall job satisfaction. The survey was anonymous and incentivized with a restricted-use $25
Walmart eGift card. The overall response rate was 44%. Here exposure is defined as spending
more vs. less time with IGNITE participants; Appendix Table A4 shows staff characteristics are
well-balanced across this measure. The recruitment flyer is shown in Panel B of Appendix Figure
A9. A link to the full survey and further details are given in Appendix D.
11
The ViaPath data is not timestamped, however, preventing measurement of participation in individual jail spells.

10
3.2 Main Analysis Sample
We combine the above data sources, merging on unique case or person identifiers, to construct our
main analysis sample. Appendix Figure A10 summarizes the sample construction. We start with
the universe of arrests in the Genesee County JMS data and set aside those booked before January
1, 2016, which we use as a hold-out sample to predict recidivism risk for certain analyses. We
also exclude individuals booked after May 2022, so as to have enough time to measure 12-month
recidivism for all individuals. We then merge JMS to ROAs and exclude incarcerated individual-
spells where the individual was immediately released without charge, since these individuals did not
interact with the court system and were ineligible for IGNITE. We also exclude a small portion (4%)
of remaining individuals who are not Michigan residents, since we are unlikely to accurately measure
their recidivism, individuals who are missing demographic information (1%), and individuals who
are not yet released for at least three months (1%). We link these data to Jail Incident Data,
Kites data, and Mt. Morris data using individual identifiers. The resulting sample includes 23,610
incarcerated individual-spells representing 14,794 unique individuals. When studying recidivism
outcomes, we further drop a small minority (around 6%) of individuals who leave jail via a transfer
to prison in order to avoid mechanical incapacitation effects.12
Summary statistics for this sample are shown in Appendix Table A5. The sample is 76% male,
53% Black, and the majority fall into the age range of 25-44.13 43% of individuals were booked in
the past year and 53% are charged with a felony, with an average number of charges of 1.4. The
average time in jail is 1.5 months, with a standard deviation of 4.2. Nearly 40% of the incarcerated
individual-spells in Genesee County Jail experience a court delay in District Court and 18% are
rebooked in the three months after release.
A potential concern when using administrative crime data is reporting behavior. In principle,
correctional officers could under-report within-jail misconduct to demonstrate the effectiveness of
IGNITE. In practice, this concern is lessened for our identification strategy, which uses variation
in court delays rather than simple cross-sectional or over-time comparisons. Strategic misreporting
would have to be correlated with the court delay instrument. Moreover, there is minimal concern
of strategic misreporting for our primary recidivism outcomes, as rearrest and rebooking decisions
in Genesee and Saginaw Counties are made by the local police forces—not by the Genesee County
Sheriff or jail administrators.14
12
The vast majority of incarcerated individuals are released from jail without conditions, in part because judges
usually consider time spent in jail when sentencing. Below we show that we find no effects of time in IGNITE on
different release conditions or on post-conviction outcomes, including being sentenced to state prison.
13
The Black share is high relative to that of Genesee County (20%) and comparable to that of Flint (56%).
14
While reported within jail and therefore subjective to some misreporting concerns, the shares of major misconduct
and medical incidents are constant over time (see Appendix Figure A11). We also find that the schedule outlined
in Appendix Figure A6 is roughly adhered to as measured by the floor-specific login times on Chromebooks (see
Appendix Figure A12).

11
3.3 Motivating Evidence
Figure 1 motivates further study of the impact of IGNITE on recidivism by plotting the relationship
between predicted and observed recidivism among those booked in Genesee County before and after
the start of IGNITE. Specifically, we plot the average three-month recidivism rates of individuals
booked before and after September 2020 by bins of the individuals’ predicted recidivism risk,
obtained from a logit regression on individual observables in a 2015 holdout sample (described
above). Prior to the start of IGNITE, actual recidivism rates closely track these predictions.
However, after IGNITE was launched, actual recidivism rates are significantly lower, uniformly
across all levels of predicted risk. This pattern suggests a dramatic change in recidivism outcomes
that coincides with the launch of IGNITE programming, though the purely time-series analysis is
far from conclusive. We next develop and apply a more sophisticated quasi-experimental strategy
to estimate causal effects of IGNITE exposure.

4 Empirical Strategy
4.1 Difference-in-IVs Approach
To formalize our IV strategy, consider a population of individuals booked into Genesee County
Jail either before or after the launch of IGNITE in September 2020. Let Pi ∈ {0, 1} indicate that
individual i was booked post-IGNITE, let MiJ count the number of months individual i spends in
jail, and let MiI count the number of months i is exposed to IGNITE within the jail. To start simply,
we assume that nobody booked pre-IGNITE is exposed to IGNITE: i.e., that MiI = MiJ × Pi . Our
general IV strategy, developed below, relaxes this assumption to allow individuals booked before
IGNITE to be partially exposed by virtue of their continued incarceration in September 2020.
Consider a simple causal model relating MiJ and MiI to their effects on an outcome Yi :

Yi = Yi (0) + γi MiJ + βi MiI , (1)

where Yi (0) is an untreated potential outcome, i.e., the outcome that individual i would see with
no time in jail or IGNITE. Here γi denotes the incremental effect of time in jail for individual i in
the absence of IGNITE, while βi denotes the incremental effect of IGNITE exposure of individual i
holding fixed their time in jail. We assume these potentially heterogeneous causal effects are linear
in time only for initial ease of exposition; below we discuss a more general causal model.
To estimate causal effects, we assume that individuals are as-good-as-randomly assigned to a
court delay indicator Zi ∈ {0, 1}. Here, again only for initial simplicity, we imagine Zi is uncon-
ditionally randomly assigned (i.e., without any controls) and known to have no direct effect on
outcomes, making it statistically independent of (Pi , Yi (0), γi , βi ). Court delays extend time in jail
both pre- and post-IGNITE, making Zi positively correlated with both MiJ and MiI .
Under these conditions, an IV regression of Yi on either MiJ (in the pre-IGNITE period) or MiI

12
(in the post-IGNITE period), instrumenting with Zi , identifies a weighted average of causal effects:

Cov(Zi , Yi | Pi = 0)
β P re ≡
 P re 
= E ω i γ i | P i = 0 (2)
Cov(Zi , MiJ | Pi = 0)
Cov(Zi , Yi | Pi = 1)
β P ost = E ωiP ost (γi + βi ) | Pi = 1 ,
 
≡ I
(3)
Cov(Zi , Mi | Pi = 1)

where ωiP re and ωiP ost are weights that average to one and capture the relative “complier” status of
individual i: the relative amount of time in jail individual i is induced to via court delays.15 When
delays only weakly increase time in jail pre- or post-IGNITE (a natural first-stage monotonicity
condition), both weighting schemes are convex: ωiP re ≥ 0 and ωiP ost ≥ 0.
Equation (2) shows that the pre-IGNITE IV identifies a convex weighted average of time-in-jail
effects γi while Equation (3) shows the post-IGNITE IV identifies a weighted average of γi + βi .
The latter combines marginal IGNITE exposure effects, βi , with baseline time-in-jail effects γi . To
isolate IGNITE exposure effects, we consider the following condition on jail effects over time:

E ωiP re γi | Pi = 0 = E ωiP ost γi | Pi = 1 .

   
(4)

Equation (4) restricts heterogeneity in baseline time-in-jail effects pre- and post-IGNITE, similar
in spirit to a conventional “parallel trends” restriction on untreated potential outcome changes
before and after a policy change in conventional difference-in-differences analyses. Our condition
is satisfied when, if not for the start of IGNITE, the IV estimates would not have changed in
September 2020.16 Clearly, this condition is satisfied when time-in-jail effects γi are homogenous or
otherwise uncorrelated with the IV weights ωiP re and ωiP ost . Below we show how the condition can
be probed graphically, as with “pre-trend” checks in conventional difference-in-differences; we also
relax it by incorporating additional cross-sectional comparisons with neighboring Saginaw County.
Under Equation (2), a difference-in-IVs identifies a weighted average of IGNITE exposure effects.
Specifically, differencing Equations (3) and (2), we have by Equation (4):

β ∆ ≡ β P ost − β P re = E ωiP ost βi | Pi = 1 ,

 
(5)

where again ωiP ost ≥ 0 when court delays do not reduce time in jail. β ∆ then captures a convex
average of incremental effects of additional time exposed to IGNITE, βi , holding time in jail fixed.
Appendix E.2 generalizes this result to nonlinear causal effects of MiJ and MiI , showing that under
an appropriate generalization of Equation (4), the difference-in-IVs identifies an average causal
response (ACR) function, as in Angrist and Imbens (1995). This shows that β ∆ generally captures
15
Formally, ωiP re = (MiJ (1) − MiJ (0))/E[MiJ (1) − MiJ (0) | Pi = 0] where MiJ (z) denotes individual i’s potential
time in jail when Zi = z and ωiP ost = (MiI (1) − MiI (0))/E[MiI (1) − MiI (0) | Pi = 1] where MiI (z) denotes individual
i’s potential time in IGNITE when Zi = z. See Appendix E.1 for derivations of Equations (2) and (3).
16
Note that, unlike a conventional parallel trends assumption, Equation (4) imposes no model on untreated
potential outcomes Yi (0). Our strategy to avoiding omitted variable bias from Yi (0) can instead be viewed as “design-
based” (Borusyak and Hull, Forthcoming), leveraging the as-good-as-random assignment of court delay shocks.

13
a weighted average of incremental IGNITE exposure effects at different margins of exposure time.17

4.2 IV Specification
Our main estimates come from a two-treatment IV specification that applies the above difference-
in-IVs logic while accommodating additional controls and the possibility that individuals booked
before the launch of IGNITE were nevertheless exposed to the program after September 2020. For
a given outcome Yi , we estimate:

Yi = βMiI + γMiJ + Xi′ δ + εi (6)

where again MiI and MiJ count the months individual i is exposed to IGNITE and jail, respectively.
Here Xi is a covariate vector that includes the indicator for a post-IGNITE booking Pi , along with
other controls and a constant. We instrument for the two endogenous variables, MiI and MiJ , with
Zi and Zi × Pi (controlling for Xi ), where Zi again indicates a District Court delay.
The main IV coefficient of interest, β, reduces to a difference-in-IVs when no individuals booked
pre-IGNITE are exposed to IGNITE (i.e., MiJ × Pi ), and when the controls in Xi are saturated in
Pi . In this case, the IV estimate is given by β̂ = β̂ P ost − β̂ P re where β̂ P ost and β̂ P re are estimates
from two separate IV specifications:

Yi = β P re MiJ + Xi′ δ P re + εPi re (7)

Yi = β P ost MiI + Xi′ δ P ost + εPi ost , (8)

estimated in the Pi = 0 and Pi = 1 subsamples, respectively. The general two-treatment IV

specification, Equation (6), allows for individuals with exposure to jail both pre- and post-IGNITE.
We include two types of controls in Xi in addition to Pi . First, in all IV specifications, we
include a set of design controls intended to account for any non-randomness in the assignment of
court delays, Zi . The design controls are court division fixed effects (FEs) based on the location of
arrest (since delays may be more likely for some divisions of Genesee County than others), booking
month and day-of-week FEs for the first scheduled hearing (since delay propensities may vary over
time), the number of charges and FEs for charge type (felony, misdemeanor traffic, misdemeanor
DUI, and other misdemeanor crime) to account for different probabilities of delay by the seriousness
of the charge. Second, in some IV specifications, we include auxiliary controls reflecting individual
demographics and other characteristics. These controls are not needed for identification, but may
yield precision gains by absorbing residual variation in the outcomes.
17
The appendix model continues to impose additive separability of time-in-jail and time-in-IGNITE effects, while
allowing for arbitrary effect heterogeneity for these two treatments individually (i.e. fully relaxing the linear dose-
response form of (1)). Without separability, simple differencing as in (5) may not suffice to fully isolate IGNITE
effects from non-IGNITE time-in-jail effects, although this is less of an issue for recidivism outcomes where pre-
IGNITE time-in-jail effects are small and insignificant. In practice, we find minimal effect heterogeneity across many
observables including predicted time-in-jail—suggesting the baseline linear model gives a reasonable approximation.

14
4.3 Identifying Assumptions and Tests
IV estimates of (6) capture average causal effects of IGNITE exposure and non-IGNITE time in
jail under four assumptions. The first three assumptions are standard in IV analyses and the fourth
follows Equation (4). Here we discuss each assumption and provide some initial empirical support.
Our first identifying assumption is that the court delay indicator Zi is as-good-as-randomly
assigned given the design controls in Xi . This assumption is consistent with the institutional
setting (see Section 2.3) as well as a number of empirical balance tests shown in Table 1. Panel A
of this table shows that several observable individual characteristics are uncorrelated with Zi given
the design controls, while Panel B further shows balance on the characteristics of census tracts in
which individuals reside. The characteristics in these two panels constitute our auxiliary controls.
Section 5.2 summarizes additional checks of as-good-as-random assignment.18
Alongside these balance tests, Panel C of Table 1 shows that experiencing a court delay signif-
icantly extends an individual’s time in jail—an implicit instrument relevance condition for our IV
strategy. On average, individuals spend 0.4 months (around two weeks, or 27%) longer in jail when
they experience a court delay (given the design controls). This first stage is highly significant, with
an F -statistic of around 40.5. We further explore the first-stage relationship below.
Our second and third identifying assumptions are a standard IV exclusion restriction and mono-
tonicity condition: i.e., that court delays do not affect our outcomes of interest except by extending
time in jail, and that delays only weakly increase time in jail (both pre- and post-IGNITE). These
assumptions are also consistent with the institutional setting, and we probe them empirically in
Section 5.2. In Section 5.3 we further discuss how our exclusion and monotonicity assumptions may
be more plausible than in a conventional “judge IV” strategy which leverages as-good-as-random
District Court judge assignment instead of court delays (while also showing estimates from this
alternative strategy).
The final identifying assumption follows Equation (4) and allows us to interpret estimates of β
in terms of the causal effects of additional IGNITE exposure holding time in jail fixed. Intuitively,
the assumption is satisfied when IV estimates of time-in-jail effects would not have systematically
changed in September 2020 if not for the launch of IGNITE. There are two primary threats to
this assumption. First, as in a conventional difference-in-differences approach, our identifying
assumption could be violated if another unobserved policy change or broader change in Genesee
County occurred around the start of IGNITE. Unlike with a conventional difference-in-differences
approach, however, such time-varying confounds would have to affect the effects of time in jail
rather than potential outcome levels. Below we conduct a non-parametric analysis of court delay
effects over time—akin to the standard “pre-trend” check in conventional difference-in-differences
analyses—which suggests minimal scope for such time-varying confounds in our setting.
18
Appendix Table A6 checks for differential attrition, which could introduce bias even when delays are as-good-as-
randomly assigned. Reassuringly, we find that court delays do not cause individuals to exit our baseline three-month
recidivism analysis sample at a significantly higher rate. For longer windows there is some evidence of differential
attrition but effect sizes are small. With 12-month recidivism, for example, court delays are found to make individuals
1.1 percentage points less likely to stay in the sample off a baseline follow-up rate of 98%.

15
 The second potential threat to this assumption is that the types of individuals who comply
with the court delay instrument changed before and after IGNITE. More formally, Equation (4)
could fail if β P re and β P ost put different weight on heterogeneous time-in-jail effects. Below, we
show these effects are relatively small in the pre-IGNITE period and that causal effects are generally
homogeneous across observable characteristics—reducing concerns of bias from effect heterogeneity.
More direct evidence comes from Table 2, which shows the average observable characteristics of
instrument compliers before and after IGNITE.19 We find no statistically significant differences in
these averages, suggesting compliers are broadly comparable before and after the start of IGNITE.20
We make three further points on the interpretation of our IV estimates. First, while our primary
interest is on the causal interpretation of the IGNITE exposure effect β, we note that the combined
β + γ coefficient may be causally interpretable under weaker conditions. Specifically, β + γ captures
the average effect of increased time in jail in the post-IGNITE period when baseline time-in-jail
effects are not comparable pre- and post-IGNITE (i.e., when Equation (4) fails). Correspondingly,
we report estimates of this combined effect along with estimates of β.
Second, we note that β may retain its interpretation as an average causal effect of IGNITE
exposure when the conventional IV assumptions (as-good-as-random assignment, exclusion, and
monotonicity) fail, provided the bias from such violations manifests similarly in the pre- and post-
IGNITE periods. For example, the causal interpretation of β is robust to court delays directly
affecting within-jail misconduct by increasing an individual’s frustration with the criminal justice
system (a potential exclusion restriction violation) provided such frustration effects are similar pre-
and post-IGNITE.21 In Section 5.3 we develop extensions of our baseline approach that further
weakens this exclusion restriction by incorporating additional comparisons to Saginaw County.
Third, we note that β targets the average effect of exposure to IGNITE programming and not
the effect of program participation itself. While participation rates are known to be high—around
90% on average—we do not have individual participation data that would let us study the latter.
Under a plausible monotonicity condition, a hypothetical extended IV approach would scale our
estimates by such a takeup rate.22 In this scenario, the magnitude of our effects can be viewed as
giving a lower bound on the magnitude of effects from IGNITE participation.
19
Specifically, we report estimated means of individual characteristics weighted by the same measures of compliance
status that underlie the IV estimates of jail exposure effects pre- and post-IGNITE. See Appendix E.3 for details.
20
Similarly, Appendix Figure A13 shows that the weights our baseline IV specification puts on different margins of
exposure time are relatively similar pre- and post-IGNITE. See again Appendix E.3 for details on these calculations.
Below we show effects are homogeneous across individuals with different predicted time in jail, reducing concerns
about any pre-post differences in the exposure time weights.
21
Estimates of γ or the combined β + γ coefficient would not, however, be causally interpretable in such cases.
22
The exclusion restriction in this hypothetical specification would generally rule out within-jail spillovers across
individuals who do and do not participate in IGNITE, in contrast with our preferred exposure treatment specification.

16
5 Main Findings
5.1 Misconduct and Recidivism Effects
Figure 2 plots the reduced-form variation underlying our primary IV estimates. Each point shows
the estimated effect of court delays on one of our primary outcomes—either weekly major miscon-
duct or three-month recidivism—separately by an individual’s booking month. We obtain these
estimates by regressing the outcome on the court delay instrument, adjusting for the design controls
and the auxiliary controls from Panels A and B of Table 1.
The figure shows strikingly different reduced-form effects of court delays pre- and post-IGNITE.
Before September 2020, delayed individuals on average saw at most a small increase in within-jail
weekly misconduct rates and no increase in post-release recidivism. A pre-IGNITE IV specification
would scale these reduced-form effects by the corresponding first stage to find small or no effects
on time in jail before the start of IGNITE. In contrast, court delays had sizable negative effects on
both misconduct and recidivism after the start of IGNITE. As described in Section 4.1, a difference-
in-IVs estimate contrasting these pre- and post-IGNITE estimates would therefore suggest large
negative IGNITE exposure effects.23 Importantly for this interpretation, the figure shows no clear
trends in the reduced-form effects of either outcome either before or after September 2020. It is
therefore plausible that, if not for the start of IGNITE, the time-in-jail effects on misconduct or
recidivism would have remained slightly positive or insignificant.
Table 3 reports our main reduced-form and IV estimates of misconduct and recidivism effects.
IV estimates in columns 1 and 3 are from Equation (6), with a post-IGNITE indicator and all
design controls in Xi . In columns 2 and 4 we further include the auxiliary controls; consistent with
the balance tests in Table 1, these controls do not materially change the estimates. Reduced-form
estimates come from regressing outcomes on the two instruments, the court delay indicator and its
interaction with the post-IGNITE dummy, adjusting for the controls.24
The table reports large estimated effects of IGNITE exposure on both within-jail misconduct
and post-release recidivism. On average, one additional month in IGNITE is estimated to reduce
weekly major misconduct incidents by 0.16 and three-month recidivism by 8.1 percentage points.
These represent reductions of 49% and 18%, respectively, relative to reported control complier
means.25 As in Figure 2, we find no effects of additional months in jail on recidivism pre-IGNITE.
The combined recidivism effect of months in IGNITE and jail is therefore similar to the estimated
IGNITE exposure effect. We find a larger positive effect of months in jail on misconduct pre-
IGNITE. The combined misconduct effect of post-IGNITE months-in-jail is therefore smaller than
the estimated IGNITE exposure effect (around -8.2 percentage points). We further contextualize
23
Appendix Figure A14 shows the corresponding first-stage plot. The average effect of court delays on time in jail
is roughly constant pre- and post-IGNITE, at around two weeks.
24
Appendix Table A7 shows corresponding first-stage estimates. First-stage F -statistics, computed as in Sanderson
and Windmeijer (2016), are around 80 for the time-in-jail treatment and around 56 for the time-in-IGNITE treatment.
25
Control complier means come from IV regressions of Yi · 1[MiJ < m] on 1[MiJ < m] instrumenting by Zi with
both design and auxiliary controls. Following Appendix E.3, this estimates average outcomes when individuals spend
less than m months in jail. We set m to correspond to a “control” condition of less than one week in jail.

17
our primary IGNITE exposure effect estimates in Section 6.1, below.
Figure 3 shows how estimated misconduct and recidivism effects vary over time. We plot IV
estimates of IGNITE exposure effects (obtained as in columns 2 and 4 of Table 3) for two alternative
outcomes: whether an individual experienced any major misconduct in a given week since booking
(Panel A of Figure 3) and whether an individual was rebooked in a given month since release (Panel
B of Figure 3). Panel A shows that estimated misconduct effects are relatively stable over time,
with around an 8 percentage point reduction in misconduct risk in any given week since booking.
In contrast, Panel B shows that estimated recidivism effects grow steadily over time—to around a
15 percentage point reduction in one-year recidivism post-release.
Appendix Figures A15 and A16 explore heterogeneity in our baseline misconduct and recidivism
effect estimates by individual demographics, prior offense status, high vs. low predicted lead expo-
sure from the Flint water crisis, and predicted recidivism risk.26 Specifically, we estimate versions
of Equation (6) which (i) add as treatments interactions of the months in IGNITE or months in
jail treatments with bins of observable characteristics, (ii) add to the instrument list interactions
of (Zi , Zi × Pi ) with the same bins, and (iii) add the bin dummies as controls. The figures plot
resulting estimates of bin-specific months-in-IGNITE and months-in-jail effects, which are valid
under conditional versions of our main identifying assumptions. Overall, we find roughly similar
effect estimates across demographic groups, prior offense status, and predicted lead exposure (see
Appendix Figure A15). We do, however, find meaningful heterogeneity by predicted recidivism risk
(see Appendix Figure A16): recidivism reductions are larger for riskier individuals.
Estimated effects on alternative recidivism and misconduct measures, along with other related
outcomes, are shown in Appendix Table A1. Panel A shows significant IGNITE effects on the three-
month probability of an individual being recharged or reconvicted, as well as in rates of weekly
minor misconduct. In addition, we find significant IGNITE effects on the rate of weekly serious
violent misconduct within jail, which is less susceptible to reporting bias.27 Panel B shows we find
no significant effects of IGNITE exposure on whether an individual is released on tether, released on
bail, sentenced to prison, convicted, or released to a rehabilitation center. The lack of effects here is
unsurprising since pretrial judges are not permitted to consider in-jail misconduct. Judges are also
not provided such information by custody staff prior to making pretrial, conviction, or sentencing
decisions. The large post-release recidivism effects we find in Table 3 thus do not appear to be
mediated by the channels in Panel B of Appendix Table A1 (e.g. incapacitation effects). We also
find no effects of IGNITE exposure on suicide attempts or other medical incidents within jail, in
contrast to the large major misconduct effect estimates in Table 3.
26
Lead exposure is predicted from an individual’s residential zip code. We predict recidivism risk by a logit regres-
sion on the auxiliary controls in the 2015 holdout sample. Appendix Figure A17 shows we do not find heterogeneity
by predicted time in jail, constructed analogously by OLS.
27
More serious misconduct includes threatening another with bodily harm; escaping, attempting to escape or
helping another to escape from the law; and inflicting bodily injury upon another person.

18
5.2 Robustness Checks
As discussed above, Tables 1-2 and Figure 2 show balance and trend analyses that broadly support
our IV strategy. Here we discuss a number of additional robustness checks, summarized in Table 4.
One category of potential concerns stems from the COVID-19 pandemic. Major pandemic
policy responses occurred between March 2020 (the start of lockdowns) and June 2021 (when
vaccines were first widely distributed), which overlaps with the launch of IGNITE in September
2020. Around this period, one might imagine that jailed individuals were of relatively higher
criminal risk (if lower-risk individuals were released to reduce jail populations), that misconduct
rates declined simply because individuals were more segregated within the jail (due to COVID-
19 quarantine protocols), or that the pandemic and related policies more broadly affected how
misconduct and recidivism outcomes were measured.28 For the first concern, it is reassuring that
we find, if anything, larger effects among individuals with high levels of predicted recidivism risk
(see Appendix Figure A16). For the second and third concerns, Table 4 reassuringly shows that we
obtain similar results as our baseline estimates when estimating effects on misconduct not involving
others (e.g., counterfeiting or forgery) and that we obtain similar estimates for both misconduct and
recidivism when controlling for a time trend (interacted with the court delay instrument) or when
altogether excluding March 2020 to June 2021 from the analysis sample. We further show in Section
5.3 that recidivism rates declined in Genesee County around September 2020 relative to Saginaw
County, despite both counties being subject to the same statewide COVID-19 protocols, and that
the same IV specification deployed in Saginaw County finds no placebo IGNITE effects. Together,
these checks suggest our findings are not driven by changing conditions from the pandemic.
A second category of potential concerns is violations of as-good-as-random instrument assign-
ment. One might be concerned, for example, that some District Court delays are initiated by the
incarcerated individual and thus are potentially endogenous. Reassuringly, Table 4 shows we obtain
similar or larger estimates when using alternative definitions of the instrument that are less sus-
ceptible to manipulation: i.e., when including Circuit Court delays, restricting to COVID-19 and
fiscal crisis delays, using only delays around Federal holidays, or restricting to delays occurring on
days with multiple court delays across different individuals. Panel A of Appendix Figure A18 gives
further evidence that delays are not self-initiated. We find that the probability an incarcerated
individual sent a Kites message with a communication or court-related request (using the words
talk, speak, need, can, please, court or judge) jumped in the four weeks after the COVID-19-induced
court closure on March 17, 2020. Panel B shows no such increase in an analogous event study one
year prior. Together with the balance checks in Table 1, these checks broadly support the view
that court delays are as-good-as-randomly assigned.
A third category of potential concerns focuses on the IV exclusion restriction. Even when
delays are as-good-as-randomly assigned, one might be concerned that they have direct effects on
misconduct or recidivism by, for example, increasing an individual’s frustration with the criminal
28
The total Genesee County Jail population declined by around 40% in the onset of the pandemic before recovering
to pre-pandemic levels in the winter of 2020.

19
justice system. Table 4 shows we obtain similar estimates when controlling for whether an individual
experienced multiple court delays, as one proxy for such frustration. Recall also that our baseline
IV approach allows for any direct effects of the instrument provided they are similar in the pre- and
post-IGNITE period. Therefore, any time-invariant “frustration effect” would be differenced out
and would not bias our IGNITE effect estimates. In Section 5.3 we discuss checks using alternative
differencing strategies, which are valid under different or weaker exclusion restrictions.
A final potential concern, specific to the within-jail misconduct outcome, is that IGNITE par-
ticipation or access to tablets through the program simply occupied the time that individual would
have otherwise spent engaging in misbehavior. In other words, IGNITE might have reduced within-
jail misconduct simply via an within-jail “incapacitation” effect. This could affect the interpretation
of the large reductions in misconduct we find in our baseline specification, but would not introduce
bias. Reassuringly, Table 4 shows we obtain similar misconduct effects when restricting to times of
day when there was no IGNITE programming as well as to hours when individuals did not have
access to commissary tablets for entertainment purposes.

5.3 Alternative Identification Strategies

Table 5 shows we obtain very similar recidivism effect estimates from alternative difference-in-IVs
specifications which use Saginaw County as a “control” group in the post-IGNITE period. Column
1 presents our baseline estimates from Genesee County. Column 2 estimates Equation (6) with data
from both counties in the post-IGNITE period only, instrumenting (MiI , MiJ ) with (Zi , Zi × Si ),
where Si ∈ {0, 1} indicates that individual i was booked in Saginaw County, and including the
baseline design and auxiliary controls (with Si replacing Pi ). As in our baseline analysis, we
find no recidivism effect of increased time in jail in the absence of IGNITE but a large negative
recidivism effect from IGNITE exposure.29 The -6.4 percentage point IGNITE effect is similar to
our baseline -8.1 percentage point effect, replicated in column 1. In column 3, we estimate our main
specification in Saginaw County only, obtaining a tight null estimate of the change in time-in-jail
effects before and after September 2020. This suggests there was no simultaneous regional change
in jail or rearrest policy confounding our baseline Genesee County estimates. Finally, column 4
subtracts this Saginaw County placebo check from our baseline over-time IGNITE exposure effect in
a “double” difference-in-IVs specification, showing again a large recidivism effect of -7.5 percentage
points. Notably, this specification weakens our baseline identifying assumptions by differencing out
any direct time (or county) effects that would otherwise confound the estimates in columns 1 or 2.
Appendix Table A8 shows estimates from an alternative IV strategy which, as in Bhuller et al.
(2020), instruments with leave-one-out averages of District Court judge “leniency.” We estimate
Equation (7) in both the pre-IGNITE and post-IGNITE periods, instrumenting for time in jail
MiJ with Li , defined as the average time in jail of individuals other than i who were assigned
to individual i’s District Court judge, residualized on court division fixed effects (FEs), day-of-
week FEs, and month FEs. This judge IV strategy is valid when judges are as-good-as-randomly
29
We do not observe misconduct outcomes in Saginaw County; see Appendix C.3 for details.

20
assigned given the controls, with their assignment only affecting outcomes through time in jail and
shifting all individuals’ time in jail in the same direction across judges. These exclusion restriction
and monotonicity conditions may be less plausible than our baseline identifying assumptions, since
judges make many decisions that affect an individual’s experience within and beyond jail and also
because these decisions may result in longer stays in jail for some individuals and shorter stays for
others. Correspondingly, a joint test proposed by Frandsen, Lefgren and Leslie (2023) rejects judge
IV exclusion and monotonicity decisively, in both pre- and post-IGNITE samples (see Appendix
Table A9). Nevertheless, the table shows difference-in-IVs estimates that are qualitatively similar
(though generally less precise) as our preferred IGNITE exposure effect estimates.
Appendix Figure A19 shows estimates from an alternative difference-in-differences strategy,
which compares overall trends in three-month recidivism rates from Genesee County to correspond-
ing trends from neighboring Saginaw County. We plot event study coefficients from regressing the
recidivism of individuals booked in either county on a Genesee County indicator interacted with
the individual’s booking date relative to December 2019: the period after which a nontrivial share
of individuals booked in Genesee County were exposed to IGNITE starting in September 2020 (see
Appendix Figure A20).30 Recidivism trends are similar between the two counties prior this period
but diverge sharply thereafter, with individuals booked in Genesee County seeing an average re-
duction in recidivism of around 3 percentage points (see Appendix Table A10). This translates to a
reduction of around 11% for one month of exposure to IGNITE, similar to the 18% reduction found
using our preferred IV strategy.31 While less fine-grained than our IV strategy, the event study
helps build further confidence in its core logic with flat pre-trends showing no unusual pre-IGNITE
recidivism dynamics in Genesee County.

6 Contextualization and Mechanisms

6.1 Cost of Crime Effects
Appendix Table A11 translates our main recidivism effect estimates into estimates of the effect of
IGNITE exposure on post-release social costs of crime. Specifically, we estimate Equation (6) with
an outcome that measures the total cost of future crimes at different horizons. We following the
most conservative cost of crime estimates in Miller et al. (2021).32 For our baseline three-month
recidivism horizon, we find that one additional month of IGNITE exposure decreases the social
cost of crime by around $2,957 per person. As in Figure 3, this estimated effect grows over time,
30
We do not use an individual’s time to IGNITE exposure because this is endogenous for individuals booked prior
to September 2020.
31
Individuals in Genesee County spent around 1.5 months in jail both pre- and post-IGNITE (Appendix Figure
A14), average three-month recidivism is around 19 percentage points at baseline (see Appendix Table A10), and
(3pp/1.5 months)/19pp ≈ 11%.
32
This calculation divides future crimes into DUIs, drug offenses, motor vehicle offenses, persons offenses, property
offenses, public order offenses, weapons offenses, and other offenses. Within each of these crime types, we take the
lowest social cost estimate from Miller et al. (2021) to provide the most conservative estimate; for example, we use
the cost estimate for assault instead of murder for persons offenses.

21
to a per-person-month reduction of around $5,615 over a horizon of 12 months. These estimates
are around 24% and 12% of the control complier means, respectively, and are likely to understate
true crime cost reductions both because we use conservative cost estimates and because we do not
count any costs of within-jail misconduct. The large crime cost reductions are especially notable
given stable spending in Genesee County Jail pre- and post-IGNITE (see Appendix Figure A4).

6.2 Literature Comparison

Figure 4 compares our baseline recidivism effect estimates to other quasi-experimental estimates
in the literature. Panel A compares the estimated relative effect of IGNITE exposure on one-year
recidivism to comparable relative effect sizes of other rehabilitative programs for justice-involved
individuals. When possible, we compute one-month effects of these programs by assuming linear
effects. See Appendix Section F for details on the effect size calculations.
This literature comparison shows that our main one-month IGNITE exposure effect (a 23.3% re-
duction in one-year recidivism) is comparable to or larger than the effects from other rehabilitative
programs, including CBT programming (Heller et al., 2017), open prisons (Mastrobuoni and Terl-
izzese, 2022), diversion (Mueller-Smith and Schnepel, 2021; Augustine et al., 2022), and restorative
justice conferencing (Shem-Tov, Raphael and Skog, 2022). We also compare our months-in jail-
estimates to incarceration effects found in other studies of jail and prison (Panel B) and present
estimates from programming geared towards high-risk (but not necessarily justice-involved) individ-
uals for further context (Appendix Figure A21). Together, these comparisons show that IGNITE—
deployed in a U.S. jail among an especially high-risk adult population—generates similar reductions
in recidivism as programming in other correctional environments and countries.33

6.3 Potential Mechanisms

Given the educational nature of IGNITE programming, one obvious candidate driver of the large
decrease in misconduct and recidivism we find is increased literacy and numeracy. Incarcerated
individuals tend to enter IGNITE with very low reading and math achievement as measured by
CASAS scores. Figure 5 shows that the distribution of pre-assessment math and reading scores are
centered around a 5th-6th grade equivalence and a 6th-7th grade equivalence, respectively.
A comparison of test scores before and after an individual’s enrollment in IGNITE programming
suggests substantial improvements in math and reading skill. Figure 5 shows that, on average,
individuals gained the equivalent of around one grade level in both subjects. These gains also
appear widespread: the full distribution of post-test scores is shifted to the right in both subjects,
with a marked tilt towards higher grade equivalencies in reading. While these gains may not solely
reflect human capital accumulation (for example because of the possibility of an increased routine
and reduced distraction in IGNITE improving individuals’ test-taking skill), they are nevertheless
suggestive of improved educational achievement. The magnitude of achievement gains are massive
33
Notably, the 95% confidence interval from our IV specification overlaps with that of Bhuller et al. (2020) for
incarceration in Norway.

22
even when compared to some of the most effective educational interventions documented in recent
quasi-experimental literatures (e.g., Cohodes and Roy 2023) and consistent with policymakers’
views that IGNITE’s educational programming was a broad success (Erwin, 2023).
However, institutional knowledge and our first-stage estimates both suggest that formal edu-
cational programming is not the full story behind the large misconduct and recidivism effects. As
noted in Section 2.2, IGNITE was intended to create significant cultural change within the jail,
which could potentially enhance formal education and even possibly affect individuals who did not
participate in programming. Our court delay instrument only increased exposure to IGNITE by
around two weeks on average, making it unlikely that the misconduct and recidivism reductions
effects came from added instruction time alone. To explore the possible role of a within-jail cultural
shift, we next turn to our survey analyses.
Table 6 reports the effect of IGNITE exposure on impressions of law enforcement among com-
munity members and the formerly incarcerated. IGNITE exposure is defined as an indicator for
whether the survey respondent or someone in their close social network served time in Genesee
County Jail after the program was launched in September 2020 (about one-third of those surveyed
were exposed). The primary outcome is respondents’ views of law enforcement, which was elicited
from everyone. We find that IGNITE exposure predicts a positive view: agreement with the phrase
“Law enforcement looks out for me and my community” is 23 percentage points for IGNITE-
exposed respondents which is roughly 75% of the unexposed mean. Column 2 moreover shows
this effect is driven by respondents who had a longer IGNITE exposure. These findings indicate
that IGNITE improved perceptions of procedural justice and police legitimacy, a core component
of effective policing (Tyler and Fischer, 2014). Columns 3-6 show results on post-incarceration
outcomes which were elicited from the roughly 70% of respondents that were justice-affected. Here
we find positive but insignificant effects of IGNITE on engagement in positive activities (defined as
employment, education, or caretaking), and no effect on hopefulness about the future. Appendix
Table A3 shows that IGNITE-exposed individuals are observably similar to other individuals in
the survey, supporting the interpretation of these coefficients as estimates of causal effects.
The more favorable views of law enforcement by formerly incarcerated individuals are mirrored
by more favorable views of educational programs for incarcerated individuals by custody staff
at Genesee County Jail. Staff who regularly interact with IGNITE participants are 34 percentage
points more likely to view education in jails as worthwhile (see Table 7 and Appendix Figure A22).34
In contrast, there is no overall effect on job satisfaction (see Appendix Table A12). Exposure to
IGNITE participants is well-balanced among respondents (Appendix Table A4), again supporting
a causal interpretation of these findings.
Appendix Figure A23 further supports the apparent IGNITE culture change, via a sentiment
analysis of Kites messages sent by incarcerated individuals to custody staff before and after Septem-
ber 2020. We use the NRC Word-Emotion Association Lexicon from Mohammad and Turney
34
Appendix Table A13 lists some representative answers to a free-response question asking staff for ways IGNITE
has changed their experience working in Genesee County Jail.

23
(2010), which labels each English word as being associated with up to two sentiments (negative
and positive) and up to eight emotions (anger, anticipation, disgust, fear, joy, sadness, surprise,
and trust). Panel A shows a significant increase in the share of Kites words categorized as hav-
ing only positive sentiment of around four percentage points or 14% higher than the pre-IGNITE
level. The share of words categorized as having only negative sentiment and of words categorized
as having “neutral” sentiment (either both positive and negative or neither positive nor negative)
fell after IGNITE. Panel B further shows that shares of words in Kites messages associated with
anticipation, anger, fear, and sadness fell post-IGNITE, while the share of words associated with
trust increased. Together with the survey analysis, these qualitative findings support anecdotal
evidence of a broad shift in Genesee County Jail culture for both incarcerated individuals and staff.

7 Conclusion
This study provides the first quasi-experimental evidence that educational programming in U.S.
county jails can reduce post-release recidivism and potentially mitigate the kinds of incarceration
cycles that have long stymied criminal justice policymaking. Exposure to the Genesee County
Jail IGNITE program dramatically and persistently reduces both within-jail misconduct and post-
release recidivism, with relatively similar effects by race, sex, age, and prior offense status. We
estimate that one additional month of IGNITE exposure reduces the social cost of crime by at least
$5,300 per person in the year after their release. Qualitative evidence suggests a broad cultural
change within the jail as a core mechanism, resembling the kinds of rehabilitation-oriented policies
found outside the U.S.
We expect the IV strategy we develop for estimating IGNITE exposure effects using quasi-
random court delays to prove useful in other settings where idiosyncratic delays in administrative
policy extend an individual’s time exposed to an institution before and after a policy reform. One
could imagine, for example, evaluating the effects of reforms on healthcare or other benefit pro-
grams by comparing individuals whose appointments are and are not affected by idiosyncractic
rescheduling in the pre- and post-reform period. Our empirical framework shows how such vari-
ation can be leveraged with standard IV assumptions and a novel restriction on over-time effect
homogeneity, which we demonstrate can be validated empirically.
An important question for future research is whether the large IGNITE effects we estimate can
be replicated in other U.S. jails. Indeed, the National Sheriffs Association is committed to bringing
IGNITE to all jails in the country, though this process is still in early stages. As is often the
case with scaling-up attempts, rigorous evaluation of this process will be essential (Duflo, 2004),
potentially using similar tools as those developed here. But broadly our findings suggest that—as
in distressed Flint, Michigan—something can work.

24
References
67th District Court, State of Michigan. 2022. “Administrative Order 2022-05: Assignment of
Cases.” https://www.67thdc.com/wp-content/uploads/2022/07/LAO-Assignment-of-Cases.pdf
(accessed: May 29, 2023).
7NEWS Australia. 2020. “American riots: Michigan sheriff puts down weapons and joins
protesters in march — 7NEWS.” https://www.youtube.com/watch?v=fBo9E9Be kw (accessed:
May 25, 2023).
ABC 12 News. 2024. “Genesee County Sheriff’s Office holds IGNITE graduation.”
https://www.abc12.com/video/genesee-county-sheriffs-office-holds-ignite-graduation/
video 7badfb62-21d7-59d5-8eea-a0f0b7faece0.html (accessed: May 25, 2023).
Abdulkadiroğlu, Atila, Joshua D. Angrist, Peter D. Hull, and Parag A. Pathak. 2016.
“Charters without Lotteries: Testing Takeovers in New Orleans and Boston.” American Eco-
nomic Review, 106(7): 1878–1920.
Aizer, Anna, and Joseph J. Doyle Jr. 2015. “Juvenile incarceration, human capital, and future
crime: Evidence from randomly assigned judges.” Quarterly Journal of Economics, 130(2): 759–
803.
Alper, Mariel, Matthew R. Durose, and Joshua Markman. 2018. “2018 Update on Prisoner
Recidivism: A 9-year Follow-up Period (2005-2014).” US Department of Justice, Office of Justice
Programs, Bureau of Justice Statistics.
Andrews, D. A., and J. Bonta. 2010. The Psychology of Criminal Conduct. Anderson Publishing
Co.
Angrist, Joshua D., and Guido W. Imbens. 1995. “Two-stage Least Squares Estimation of
Average Causal Effects in Models with Variable Treatment Intensity.” Journal of the American
Statistical Association, 90(430): 431–442.
Arbour, William, Steeve Marchand, and Guy Lacroix. 2023. “Prison Rehabilitation Pro-
grams and Recidivism: Evidence from Variations in Availability.” Journal of Human Resources.
Arenberg, Samuel, Seth Neller, and Sam Stripling. 2024. “The Impact of Youth Medicaid
Eligibility on Adult Incarceration.” American Economic Journal: Applied Economics, 16(1): 121–
156.
Augustine, Elsa, Johanna Lacoe, Steven Raphael, and Alissa Skog. 2022. “The impact of
felony diversion in san francisco.” Journal of Policy Analysis and Management, 41(3): 683–709.
Autor, David H., Nicole Maestas, Kathleen J. Mullen, and Alexander Strand. 2017.
“Does Delay Cause Decay? The Effect of Administrative Decision Time on the Labor Force
Participation and Earnings of Disability Applicants.” NBER Working Paper No. 20840.
Bailey, Martha J., Hilary W. Hoynes, Maya Rossin-Slater, and Reed Walker. 2020.
“Is the Social Safety Net a Long-term Investment? Large-scale Evidence from the Food Stamps
Program.” NBER Working Paper No. 26942.
Barrett, Katherine, and Richard Greene. 2023. “‘IGNITE’ing an educational fire
in U.S. jails.” Route Fifty, August 14. https://www.route-fifty.com/management/2023/08/
igniteing-educational-fire-us-jails/389402/.
Beaumont, Leland R. 2023. “A case for Norway’s Rehabilitation Ori-

25
 ented Prison System.” The Fulcrum, November 14. https://thefulcrum.us/
a-case-for-norways-rehabilitation-oriented-prison-system.
Bergman, Peter, Raj Chetty, Stefanie DeLuca, Nathaniel Hendren, Lawrence F. Katz,
and Christopher Palmer. Forthcoming. “Creating Moves to Opportunity: Experimental Ev-
idence on Barriers to Neighborhood Choice.” American Economic Review.
Bhatt, Monica P., Sara B. Heller, Max Kapustin, Marianne Bertrand, and Christopher
Blattman. 2024. “Predicting and Preventing Gun Violence: An Experimental Evaluation of
READI Chicago.” Quarterly Journal of Economics, 139(1): 1–56.
Bhuller, Manudeep, Gordon B. Dahl, Katrine V. Løken, and Magne Mogstad. 2020.
“Incarceration, Recidivism, and Employment.” Journal of Political Economy, 128(4): 1269–1324.
Blattman, Christopher, Julian C. Jamison, and Margaret Sheridan. 2017. “Reducing
Crime and Violence: Experimental Evidence from Cognitive Behavioral Therapy in Liberia.”
American Economic Review, 107(4): 1165–1206.
Bondurant, Samuel R., Jason M. Lindo, and Isaac D. Swensen. 2018. “Substance Abuse
Treatment Centers and Local Crime.” Journal of Urban Economics, 104: 124–133.
Borusyak, Kirill, and Peter Hull. Forthcoming. “Negative Weights are No Concern in Design-
based Specifications.” AEA Papers and Proceedings.
Bravo, Mauricio Cáceres. 2024. “Prison Effects on Inmate Misconduct: Evidence from Trans-
fers.” Working Paper.
Bushway, Shawn D, and Emily G Owens. 2013. “Framing punishment: Incarceration, recom-
mended sentences, and recidivism.” Journal of Law and Economics, 56(2): 301–331.
Cohodes, Sarah, and Susha Roy. 2023. “Thirty years of charter schools: What does lottery-
based research tell us?” Blueprint Labs Discussion Paper #2023.18.
Darolia, Rajeev, Peter Mueser, and Jacob Cronin. 2021. “Labor market returns to a prison
GED.” Economics of Education Review, 82: 102093.
Data Driven Detroit. 2022. “LeadBloodLevels 2017 byTract 20181129.”
Data retrieved from DATA.GOV, https://catalog.data.gov/no/dataset/
leadbloodlevels-2017-bytract-20181129-dec8f.
Deshpande, Manasi, and Michael Mueller-Smith. 2022. “Does Welfare Prevent Crime?
The Criminal Justice Outcomes of Youth Removed from SSI.” Quarterly Journal of Economics,
137(4): 2263–2307.
Deza, Monica, Johanna Catherine Maclean, and Keisha Solomon. 2022. “Local access to
mental healthcare and crime.” Journal of Urban Economics, 129: 103410.
Diaz, Amy. 2020. “More than 90% of inmates at the Genesee County
Jail are awaiting trial.” Flint Beat, October 2020. https://flintbeat.com/
more-than-90-of-inmates-at-the-genesee-county-jail-are-awaiting-trial/.
Di Tella, Rafael, and Ernesto Schargrodsky. 2013. “Criminal Recidivism after Prison and
Electronic Monitoring.” Journal of Political Economy, 121(1): 28–73.
Dobbie, Will, Jacob Goldin, and Crystal S. Yang. 2018. “The Effects of Pre-trial Detention
on Conviction, Future Crime, and Employment: Evidence from Randomly Assigned Judges.”

26
 American Economic Review, 108(2): 201–240.
Doleac, Jennifer L. 2023. “Encouraging Desistance from Crime.” Journal of Economic Literature,
61(2): 383–427.
Drago, Francesco, Roberto Galbiati, and Pietro Vertova. 2009. “The Deterrent Effects of
Prison: Evidence from a Natural Experiment.” Journal of Political Economy, 117(2): 257–280.
Duflo, Esther. 2004. “Scaling Up and Evaluation.” Annual World Bank Conference on Develop-
ment Economics, 341–369.
Erwin, Alyssa. 2023. “Celebrating 3 years of I.G.N.I.T.E with the launch
of new program R.I.S.E.” ABC12 News, September 8. https://www.
abc12.com/news/celebrating-3-years-of-i-g-n-i-t-e-with-the-launch-of-new/article
31ee47d6-4e74-11ee-a65b-976a4a5eeace.html.
Estelle, Sarah M., and David C. Phillips. 2018. “Smart sentencing guidelines: The effect of
marginal policy changes on recidivism.” Journal of Public Economics, 164: 270–293.
Finley, Nolan. 2019. “Finley: Genesee County inmates wait more
than four years for trial.” The Detroit News, December 11. https:
//www.detroitnews.com/story/opinion/columnists/nolan-finley/2019/12/11/
finley-genesee-county-jail-contends-with-complex-overcrowding-delays/4385554002/.
Frandsen, Brigham, Lars Lefgren, and Emily Leslie. 2023. “Judging Judge Fixed Effects.”
American Economic Review, 113(1): 253–77.
Genesee County Controller’s Office. 2019. “Adopted Budget, Fiscal Year 2018/2019.” Genesee
County, Michigan. https://www.geneseecountymi.gov/Document Center/Department/Fiscal%
20Services/Budget/2018-2019%20Adopted%20Budget.pdf (accessed: May 25, 2023).
Genesee County Controller’s Office. 2022. “Adopted Budget, Fiscal Year 2021/2022.” Gene-
see County, Michigan. https://www.geneseecountymi.gov/2021-2022%20Adopted%20Budget.pdf
(accessed: May 25, 2023).
Golestani, Aria, Emily Owens, and Kerri Raissian. 2024. “Specialization in Criminal Courts:
Decision Making, Recidivism, and Re-victimization in Domestic Violence Courts in Tennessee.”
Journal of Human Resources.
Harlow, Caroline Wolf. 2003. “Education and Correctional Populations.” Bureau of Justice
Statistics Special Report, 12.
Heller, Sara B., Anuj K. Shah, Jonathan Guryan, Jens Ludwig, Sendhil Mullainathan,
and Harold A. Pollack. 2017. “Thinking, Fast and Slow? Some Field Experiments to Eeduce
Crime and Dropout in Chicago.” Quarterly Journal of Economics, 132(1): 1–54.
Henneguelle, Anaı̈s, Benjamin Monnery, and Annie Kensey. 2016. “Better at Home than
in Prison? The Effects of Electronic Monitoring on Recidivism in France.” Journal of Law and
Economics, 59(3): 629–667.
Hjalmarsson, Randi, and Matthew J. Lindquist. 2022. “The Health Effects of Prison.”
American Economic Journal: Applied Economics, 14(4): 234–270.
Ho, Vivian, Barton H Hamilton, and Leslie L Roos. 2000. “Multiple approaches to assessing
the effects of delays for hip fracture patients in the United States and Canada.” Health Services
Research, 34(7): 1499–1518.

27
Humphries, John Eric, Aurelie Ouss, Kamelia Stavreva, Winnie van Dijk, and
Megan T Stevenson. 2023. “Conviction, Incarceration, and Recidivism: Understanding the
Revolving Door.” http://dx.doi.org/10.2139/ssrn.4507597.
Iverson, Jana M. 2018. “Early Motor and Communicative Development in Infants with an Older
Sibling with Autism Spectrum Disorder.” Journal of Speech, Language, and Hearing Research,
61(11): 2673–2684.
Jácome, Elisa. 2020. “Mental Health and Criminal Involvement: Evidence from Losing Medicaid
Eligibility.” Job Market Paper, Princeton University.
Jeltema, Ryan. 2023. “Sewage leak from ceiling closes Flint courtroom.” ABC12 News, Decem-
ber 14. https://www.abc12.com/news/health/sewage-leak-from-ceiling-closes-flint-courtroom/
article 79f47e00-9a83-11ee-92b8-6bd43a77d81c.html.
Jones, Alexi, and Wendy Sawyer. 2019. “Arrest, Release, Repeat: How police and jails are
misused to respond to social problems.” Prison Policy Initiative. https://www.prisonpolicy.org/
reports/repeatarrests.html.
Kuziemko, Ilyana. 2013. “How Should Inmates be Released from Prison? An Assessment of
Parole Versus Fixed-sentence Regimes.” Quarterly Journal of Economics, 128(1): 371–424.
Lavecchia, Adam M., Philip Oreopoulos, and Noah Spencer. 2024. “The Impact of Com-
prehensive Student Support on Crime: Evidence from the Pathways to Education Program.”
NBER Working Paper No. 32045.
Lee, Logan M. 2023. “Halfway home? residential housing and reincarceration.” American Eco-
nomic Journal: Applied Economics, 15(3): 117–149.
Leiser, Stephanie, Shu Wang, and James L. Tatum III. 2022. “A Twenty-year Review of
Flint’s Financial Condition.” Center for Local, State, and Urban Policy. https://closup.umich.
edu/research/twenty-year-review-flints-financial-condition.
Leslie, Emily, and Nolan G Pope. 2017. “The unintended impact of pretrial detention on
case outcomes: Evidence from New York City arraignments.” Journal of Law and Economics,
60(3): 529–557.
Lochner, Lance, and Enrico Moretti. 2004. “The effect of education on crime: Evidence from
prison inmates, arrests, and self-reports.” American economic review, 94(1): 155–189.
Lotti, Giulia. 2022. “Tough on Young Offenders: Harmful or Helpful?” Journal of Human Re-
sources, 57(4): 1276–1310.
Martinson, Robert. 1974. “What Works? Questions and Answers about Prison Reform.” The
Public Interest, 35: 22–54.
Mastrobuoni, Giovanni, and Daniele Terlizzese. 2022. “Leave the Door Open? Prison Con-
ditions and Recidivism.” American Economic Journal: Applied Economics, 14(4): 200–233.
May, Jake. 2021. “12 more inmates graduate from Genesee County Jail’s
IGNITE program.” MLive Flint, June 9. https://www.mlive.com/galleries/
TFMHNGKMPZBXJIYVNHUDNAL4ZQ/.
Michigan Judiciary. 2024. “MiCOURT Case Search.” https://micourt.courts.michigan.gov/
case-search/ (accessed: on February 22, 2024).

28
Miller, Ted R., Mark A. Cohen, David I. Swedler, Bina Ali, and Delia V. Hendrie.
2021. “Incidence and Costs of Personal and Property Crimes in the USA, 2017.” Journal of
Benefit-Cost Analysis, 12(1): 24–54.
Mohammad, Saif, and Peter Turney. 2010. “Emotions Evoked by Common Words and Phrases:
Using Mechanical Turk to Create an Emotion Lexicon.” Proceedings of the NAACL HLT 2010
Workshop on Computational Approaches to Analysis and Generation of Emotion in Text, 26–34.
Mueller-Smith, Michael. 2015. “The Criminal and Labor Market Impacts of Incarceration.”
Working Paper.
Mueller-Smith, Michael, and Kevin T. Schnepel. 2021. “Diversion in the Criminal Justice
System.” Review of Economic Studies, 88(2): 883–936.
NCES, National Center for Education Statistics. 2014. “U.S. Program for the International
Assessment of Adult Competencies (PIAAC), U.S. National Supplement: Prison Study 2014.”
NPI, National Policing Institute, and National Sheriffs’ Association NSA. 2023. “The
Inmate Growth Naturally and Intentionally Through Education (I.G.N.I.T.E.) Program: Genesee
County (Michigan) Sheriff’s Office. Compendium: Community Policing and Procedural Justice
in Jails Part 5.2.” Washington, DC: Office of Community Oriented Policing Services. https://
portal.cops.usdoj.gov/resourcecenter/content.ashx/cops-w0975-pub.pdf (accessed: February 22,
2024).
NSA, National Sheriffs’ Association. 2024. “NSA IGNITE® | NATIONAL SHERIFFS’ AS-
SOCIATION.” https://www.sheriffs.org/ignite (accessed: February 22, 2024).
Packham, Analisa, and David Slusky. 2023. “Accessing the Safety Net: How Medicaid Affects
Health and Recidivism.” NBER Working Paper No. 31971.
Raphael, Steven, and David F. Weiman. 2007. “The Impact of Local Labor Market Conditions
on the Likelihood that Parolees are Returned to Custody.” In Barriers to Reentry? The Labor
Market for Released Prisoners in Post-industrial America. 304–332. Russell Sage Foundation.
Roach, Michael A., and Max M. Schanzenbach. 2015. “The Effect of Prison Sentence Length
on Recidivism: Evidence from Random Judicial Assignment.” Northwestern Law & Econ Re-
search Paper, No. 16-08.
Rose, Evan K., and Yotam Shem-Tov. 2021. “How Does Incarceration Affect Reoffending?
Estimating the Dose-response Function.” Journal of Political Economy, 129(12): 3302–3356.
Sabol, William J. 2007. “Local Labor Market Conditions and Post-prison Employment: Evidence
from Ohio.” In Barriers to Reentry? The Labor Market for Released Prisoners in Post-Industrial
America. 257–303. Russell Sage Foundation.
Sanderson, Eleanor, and Frank Windmeijer. 2016. “A Weak Instrument F-test in Linear IV
Models with Multiple Endogenous Variables.” Journal of Econometrics, 190(2): 212–221.
Sawyer, Wendy, and Peter Wagner. 2023. “Mass Incarceration: The Whole Pie 2023.” Prison
Policy Initiative, March 14. https://www.prisonpolicy.org/reports/pie2023.html.
Schnepel, Kevin T. 2018. “Good Jobs and Recidivism.” Economic Journal, 128(608): 447–469.
Shem-Tov, Yotam, Steven Raphael, and Alissa Skog. 2022. “Can restorative justice con-
ferencing reduce recidivism? Evidence from the make-it-right program.” NBER Working Paper
No. 29150.

29
Stebbins, Samuel. 2021. “Flint, MI Reported One of the Highest Mur-
der Rates in the US.” 247wallst. Available at: https://247wallst.com/city/
flint-mi-reported-one-of-the-highest-murder-rates-in-the-us/ (Accessed: 2023-05-29).
Strange, Joshua. 2023. “LITTLE SCANDINAVIA: Norwegian prison model takes shape at
Pennsylvania facility.” San Quentin News, January 2023. https://sanquentinnews.com/little-
scandinavia-norwegian-prison-model-takes-shape-at-pennsylvania-facility/.
Supreme Court, Michigan. 2023. “MICHIGAN COURT RULES OF 1985
Chapter 6. Criminal Procedure Updated May 10, 2023.” https://www.courts.
michigan.gov/siteassets/rules-instructions-administrative-orders/michigan-court-rules/
court-rules-book-ch-6-responsive-html5.zip/index.html#t=Court Rules Book Ch 6%
2FCourt Rules Chapter 6%2FCourt Rules Chapter 6.htm (accessed: May 29, 2023).
Tobón, Santiago. 2022. “Do Better Prisons Reduce Recidivism? Evidence from a Prison Con-
struction Program.” Review of Economics and Statistics, 104(6): 1256–1272.
Tuttle, Cody. 2019. “Snapping Back: Food Stamp Bans and Criminal Recidivism.” American
Economic Journal: Economic Policy, 11(2): 301–327.
Tyler, John H, and Jeffrey R Kling. 2006. “Prison-Based Education and Re-Entry into the
Mainstream Labor Market.” NBER Working Paper No. 12114.
Tyler, Tom, and Craig Fischer. 2014. “Legitimacy and procedural justice: A new element
of police leadership.” Police Executive Research Forum. https://www.policeforum.org/assets/
docs/Free Online Documents/Leadership/legitimacy%20and%20procedural%20justice%20-%
20a%20new%20element%20of%20police%20leadership.pdf.
U.S. Census Bureau. 2016. “2016 American Community Survey 5-Year Estimates.” https://
www.census.gov/data/developers/data-sets/acs-5year.html.
Weisburd, David, David P. Farrington, and Charlotte Gill. 2017. “What Works in Crime
Prevention and Rehabilitation: An Assessment of Systematic Reviews.” Criminology & Public
Policy, 16(2): 415–449.
Williams, Jenny, and Don Weatherburn. 2022. “Can Electronic Monitoring Reduce Reoffend-
ing?” Review of Economics and Statistics, 104(2): 232–245.
Yang, Crystal S. 2016. “Resource Constraints and the Criminal Justice System: Evidence from
Judicial Vacancies.” American Economic Journal: Economic Policy, 8(4): 289–332.
Yang, Crystal S. 2017. “Local labor markets and criminal recidivism.” Journal of Public Eco-
nomics, 147: 16–29.
Zapryanova, Mariyana. 2020. “The Effects of Time in Prison and Time on Parole on Recidivism.”
Journal of Law and Economics, 63(4): 699–727.
Zeng, Zhen. 2022. “Jail Inmates in 2021 – Statistical Tables.” Bureau of Justice Statistics. https:
//bjs.ojp.gov/library/publications/jail-inmates-2021-statistical-tables (accessed: December 22,
2022).

30
 Figure 1: Observed vs. Predicted Recidivism, Before and After IGNITE

.35

Observed three-month recidivism rate

.3

.25

.2

.15

.1

.05
.1 .2 .3 .4 .5
Predicted three-month recidivism rate (holdout sample)

Pre-IGNITE Post-IGNITE
Notes: This figure plots observed three-month recidivism rates in Genesee and Saginaw Coun-
ties by equal-sized bins of predicted three-month recidivism risk, before and after the start
of IGNITE. Predicted recidivism risk is estimated by a logit regression in the 2015 holdout
sample. Predictors are the design and auxiliary controls discussed in the main text.

31
Figure 2: Reduced-Form Effects of Court Delays by Booking Month

(a) Misconduct

.15
Start of IGNITE

.075
Weekly major misconduct

-.075

-.15

2016m1 2017m7 2019m1 2020m7 2022m1

Booking month

(b) Recidivism

.15

Start of IGNITE

.075
Three-month recidivism

-.075

-.15
2016m1 2017m7 2019m1 2020m7 2022m1
Booking month

Notes: This figure plots covariate-adjusted monthly average differences in weekly

major misconduct rates and three-month recidivism rates between individuals who
do and do not experience a court delay. Major misconduct rates are given by
the total number of major misconduct events observed for a unique incarceration
episode divided by the number of weeks spent in jail for that episode. Recidivism
is an indicator for whether the incarcerated individual was rebooked within three
months since the last episode’s release. Each point indicates the coefficient from
regressing the outcome on a court delay indicator and covariates among individu-
als booked in a given month. Covariates include the design and auxiliary controls
discussed in the main text. The fitted lines are obtained by local linear regression
with rule-of-thumb bandwidths, weighting by the number of incarcerated individ-
uals booked in a month. Shading indicates 95% confidence intervals derived from
individual-clustered standard errors. The vertical lines indicate the beginning of
the IGNITE program in September 2020.
32
 Figure 3: Misconduct and Recidivism Effects Over Time

(a) Misconduct

.1

Any major misconduct

-.1

-.2
1 2 3 4 5 6 7 8
Weeks since booking

(b) Recidivism

.1

0
Recidivism

-.1

-.2

3 6 9 12
Months since release

Notes: This figure plots estimated time-in-IGNITE effects from our main IV speci-
fications, with outcomes being the probability of involvement in major misconduct
in t weeks since booking (Panel A) and the probability of ever being rebooked
within t months of release (Panel B) for t = 1, . . . , 12. All specifications include
the design and auxiliary controls discussed in the main text. Blue dashed lines in-
dicate 95% confidence intervals derived from individual-clustered standard errors.

33
 Figure 4: Literature Comparisons

(a) Time in Programs

-46.4% US | Jail
Mueller-Smith and Schnepel 2021
Non-US | Jail
-23.3% US | Non-Jail
This paper Non-US | Non-Jail

-22.0%
Heller et al. 2017

-17.5%
Arbour, Marchand, and Lacroix 2023

-7.0%
Bhuller et al. 2020

-6.7%
Shem-Tov, Raphael, and Skog 2021

-3.1%
Augustine et al. 2022

-2.2%
Mastrobuoni and Terlizzese 2022

1.7%
Golestani, Owens, and Raissian 2022

4.6%
Lee 2023

-90 -60 -30 0 30 60 90

Effect Size (%)

(b) Time Incarcerated

-14.0%
Bushway and Owens 2013 US | Jail
-13.9% Non-US | Jail
Roach and Schanzenbach 2015
-11.5% US | Non-Jail
Estelle and Phillips 2018 Non-US | Non-Jail
-5.1%
Humphries et al. 2023
-4.5%
Zapryanova 2020
-3.9%
Lotti 2022
-3.8%
Kuziemko 2013
-2.7%
Hjalmarsson and Lindquist 2022
-2.4%
Tobón 2022
-2.3%
Rose and Shem-Tov 2021
-1.4%
Drago, Galbiati, and Vertova 2009
1.0%
Mueller-Smith 2015
1.6%
This paper
30.1%
Leslie and Pope 2017
35.3%
Dobbie, Goldin, and Yang 2018
51.1%
Aizer and Doyle 2015

-90 -60 -30 0 30 60 90

Effect Size (%)

Notes: This figure compares estimated treatment effect sizes from this paper to others from the literature.
See Appendix F for details on the papers and effect size construction. We distinguish between U.S. vs.
non-U.S. studies and studies in jail vs. non-jail contexts. Each point indicates the estimated effect of
treatment on recidivism as a percent of the control mean. When possible, we use one-year recidivism
outcomes and scale effects by time in treatment. 95% confidence intervals are shown around each point.

34
 Figure 5: Pre- and Post-Test Performance

(a) Math

.2

.15
Density

.1

.05

0
0 1 2 3 4 5 6 7 8 9 10 11 12
Math Grade Equivalence

Pre-Test Post-Test

(b) Reading

.2

.15
Density

.1

.05

0
0 1 2 3 4 5 6 7 8 9 10 11 12
Reading Grade Equivalence

Pre-Test Post-Test

Notes: This figure plots kernel density estimates of the distribution of math (N: 439) and
reading (N: 309) test scores before and after IGNITE program participation, of individuals who
completed both tests. Scores come from Comprehensive Adult Student Assessment Systems
(CASAS) exams administered by Mt. Morris Consolidated Schools and are normalized to
standard grade equivalents. Each distribution is estimated with a Gaussian kernel and 1.5
times the rule-of-thumb bandwidth.

35
 Table 1: Court Delay Balance Tests and First Stage
Overall Difference Standard
Mean in Means Error
(1) (2) (3)
Panel A: Individual Characteristics
Female 0.240 0.005 (0.007)
Age 25-34 0.378 0.007 (0.008)
Age 35-44 0.225 -0.008 (0.007)
Age 45-54 0.122 -0.009* (0.005)
Age 55-64 0.058 -0.001 (0.004)
Age 65+ 0.009 0.001 (0.001)
Black 0.534 -0.013 (0.008)
Booked in Past Year 0.433 -0.000 (0.007)
Has a Public Defender 0.116 0.005 (0.005)
Panel B: Census Tract Characteristics
Share with Elevated Blood Lead Level 0.031 -0.004 (0.004)
Share Black 0.429 -0.011 (0.008)
Share High School Graduate or Higher 0.848 -0.002 (0.007)
Log Median Household Income 10.322 -0.011 (0.044)
Missing Census Tract Information 0.055 0.003 (0.004)
F -Statistic for Joint Test [p-value] 1.353 [0.204]
Panel C: First Stage
Months in Jail 1.558 0.396*** (0.061)
Observations 23,610

Notes: Panels A and B summarize balance tests for the court delay instrument.
Column 1 reports the sample mean of different individual characteristics. Columns
2 and 3 report estimated coefficients and standard errors from regressing the char-
acteristics on the instrument. All regressions include the design controls discussed
in the main text. The census tract characteristics in Panel B are linked to an
individual’s residential address as recorded in the JMS data. A tract’s share with
elevated blood lead level refers to the proportion of individuals with above 4.5
micrograms of lead per deciliter of blood among those tested in the census tract in
2017. A tract’s share Black, share high school graduate or higher, and log median
household income are obtained from the 2016 American Community Survey. The
missing census tract information indicator equals one if an individual cannot be
matched to a census tract. The F -statistic is for the joint test of balance across all
individual and census tract characteristics. Panel C reports the coefficient from
a first-stage regression of months in jail on the instrument including the design
controls. Standard errors are clustered by individual. *, **, and *** refer to sta-
tistical significance at the 10, 5, and 1 percent level, respectively.

36
 Table 2: Complier Characteristics, Before and After IGNITE
Pre- Post-
IGNITE IGNITE Pre − Post Full Sample
(1) (2) (3) (4)
Panel A: Individual Characteristics
Female 0.151 0.086 0.064 0.240
(0.066) (0.054) (0.083)
Age 25-34 0.223 0.296 -0.073 0.378
(0.116) (0.100) (0.149)
Age 35-44 0.237 0.103 0.134 0.226
(0.091) (0.088) (0.125)
Age 45-54 0.163 0.213 -0.050 0.122
(0.084) (0.075) (0.111)
Age 55-64 0.116 0.156 -0.040 0.058
(0.047) (0.055) (0.070)
Age 65+ 0.061 0.044 0.017 0.009
(0.035) (0.027) (0.042)
Black 0.593 0.461 0.132 0.534
(0.115) (0.107) (0.154)
Booked in Past Year 0.407 0.320 0.086 0.433
(0.118) (0.102) (0.152)

Panel B: Crime Characteristics

Crimes against Persons 0.942 0.774 0.168 0.387
(0.100) (0.091) (0.132)
Crimes against Property 0.303 0.079 0.224 0.224
(0.103) (0.096) (0.137)
Crimes against Public Order 0.050 0.019 0.031 0.099
(0.059) (0.045) (0.073)
Drug Crimes 0.134 0.124 0.009 0.170
(0.082) (0.075) (0.108)
Weapons Crimes 0.328 0.360 -0.032 0.148
(0.117) (0.102) (0.152)
Traffic Crimes 0.098 0.179 -0.082 0.278
(0.070) (0.069) (0.095)
Other Crimes 0.252 0.133 0.119 0.166
(0.105) (0.097) (0.140)

Panel C: Census Tract Characteristics

Share with Elevated Blood Lead Level 0.038 0.033 0.005 0.031
(0.009) (0.006) (0.011)
Share Black 0.488 0.356 0.131 0.429
(0.093) (0.096) (0.132)
Share High School Graduate or Higher 0.862 0.836 0.026 0.848
(0.018) (0.016) (0.023)
Log Median Household Income 10.323 10.298 0.025 10.322
(0.106) (0.097) (0.143)
Missing Census Tract Information 0.167 0.215 -0.049 0.055
(0.062) (0.067) (0.089)
Notes: Columns 1 and 2 report estimated coefficients from IV regressions of the interaction
between a given characteristic and months in jail on months in jail, instrumenting with a
court delay indicator and controlling for the design and auxiliary controls, for individuals
booked pre- and post-IGNITE. As discussed in Appendix E, the estimated coefficients can
be interpreted as a weighted average of instrument complier characteristics along different
margins of time-in-jail response. Column 3 reports the difference between the pre- and post
coefficients while Column 4 reports the sample mean of the characteristic. Individual-clustered
standard errors are reported in parentheses. *, **, and *** refer to statistical significance at
the 10, 5, and 1 percent level, respectively.

37
 Table 3: Effects of IGNITE on Misconduct and Recidivism

Misconduct Recidivism
(1) (2) (3) (4)
Panel A: Reduced Form
Court Delay × Post-IGNITE -0.074*** -0.074*** -0.039*** -0.040***
(0.016) (0.016) (0.013) (0.012)
Court Delay 0.010 0.010 -0.002 -0.003
(0.007) (0.007) (0.007) (0.006)

Panel B: IV
Months in IGNITE -0.161*** -0.160*** -0.081*** -0.081***
(0.041) (0.041) (0.032) (0.032)
Months in Jail 0.079*** 0.077*** 0.014 0.009
(0.030) (0.029) (0.031) (0.029)

Months in IGNITE+Months in Jail -0.082*** -0.082*** -0.067*** -0.071***

(0.025) (0.025) (0.021) (0.021)

Control Complier Mean 0.329 0.329 0.457 0.457

Design Controls Yes Yes Yes Yes
Auxiliary Controls No Yes No Yes
Observations 23,610 23,610 22,191 22,191

Notes: This table reports reduced-form and IV estimates of effects on weekly major miscon-
duct incidents in jail (Columns 1 and 2) and three-month recidivism after release (Columns
3 and 4). Reduced-form estimates come from regressions of the outcome on the court delay
instrument interacted with an indicator for whether an individual was booked post-IGNITE.
IV estimates come from the specification discussed in the main text. All specifications include
the design controls discussed in the main text; Columns 2 and 4 also include the auxiliary con-
trols. Control complier means are computed as discussed in Appendix E. Individual-clustered
standard errors are reported in parentheses. *, **, and *** refer to statistical significance at
the 10, 5, and 1 percent level, respectively.

38
 Table 4: Robustness Checks

Misconduct Recidivism
(1) (2)
Baseline Specification -0.160*** -0.081***
(N = 23, 610) (0.041) (0.032)
High Predicted Risk Sample -0.123* -0.247**
(N = 5, 810) (0.064) (0.101)
Misconduct not Involving Others -0.048***
(N = 23, 610) (0.012)
Time Trend × Delay Control -0.069*** -0.101***
(N = 23, 610) (0.018) (0.025)
Excluding COVID Period -0.108*** -0.103***
(N = 20, 658) (0.033) (0.041)
Including Circuit Court Delay -0.223* -0.151**
(N = 23, 610) (0.120) (0.070)
COVID/Fiscal Crisis Delays Only -0.078* -0.091*
(N = 23, 610) (0.047) (0.050)
Delays Close to Holidays -0.124*** -0.127***
(N = 23, 610) (0.042) (0.038)
Multiple Delays per Day -0.163*** -0.094***
(N = 23, 610) (0.043) (0.035)
Multiple Delay Events Control -0.098*** -0.088**
(N = 23, 610) (0.028) (0.036)
Non-IGNITE Hours Misconduct -0.127***
(N = 23, 610) (0.039)
Non-Tablet Hours Misconduct -0.105***
(N = 23, 610) (0.041)
Notes: This table summarizes robustness checks for the primary IV
estimates of months-in-IGNITE effects on weekly major misconduct
and three-month recidivism. The first row reports estimates from our
baseline specification (Columns 2 and 4 of Table 3). The second row
restricts the sample to individuals with the top quartile of predicted
recidivism risk, as in Appendix Figure A16. The third row estimates
effects on rates of weekly major misconduct not involving other in-
dividuals. The fourth row adds a linear time trend interacted with
the court delay instrument as controls. The fifth row excludes obser-
vations from March 2020 to June 2021. The sixth row adds Circuit
Court delays to define the instrument. The seventh row uses only de-
lays associated with the COVID-19 pandemic or fiscal crises to define
the instrument. The eighth row uses only delays within two weeks be-
fore or after a federal holiday to define the instrument. The ninth row
only uses delays on days with there were multiple rescheduling events
to define the instrument. The tenth row controls for an individual
experiencing multiple delay events. The eleventh row estimates effects
on rates of weekly major misconduct restricting to times of day with
no IGNITE Programming. The twelfth row estimates effects on rates
of weekly major misconduct restricted to times of day with no tablet
access. Individual-clustered standard errors are reported in parenthe-
ses. *, **, and *** refer to statistical significance at the 10, 5, and 1
percent level, respectively.

39
 Table 5: Alternative Difference-in-IVs Estimates of Recidivism Effects

Difference-in-IVs
Post vs. Pre, Post, Genesee vs. Post vs. Pre, Double
Genesee (Baseline) Saginaw Saginaw Diff-in-IVs
(1) (2) (3) (4)
Months in IGNITE -0.081*** -0.064*** -0.075**
(0.032) (0.024) (0.030)
Months in Jail 0.009 0.021 0.009
(0.029) (0.017) (0.035)
Months in Jail × Post -0.006
(0.007)

Observations 23,610 6,380 14,227 37,837

Notes: This table reports IV estimates of effects on three-month recidivism with alternative IV
specifications. Column 1 reports the baseline estimates from Column 4 of Table 3. Column 2
reports estimates from a specification estimated in the post-IGNITE period comparing Genesee
County and Saginaw County, as described in the main text. Column 3 estimates our baseline
specification in Saginaw County, with the Months in Jail × Post treatment replacing the Months in
IGNITE treatment in Genesee County. Column 4 reports the difference in the estimated coefficient
on Months in IGNITE in column 1 and the estimated coefficient on Months in Jail × Post in Column
3. All estimates include the design and auxiliary controls discussed in the main text. Individual-
clustered standard errors are reported in parentheses. *, **, and *** refer to statistical significance
at the 10, 5, and 1 percent level, respectively.

40
 Table 6: Community Survey Results

Positive View Engaged in Hopeful about

of Law Enforcement Positive Activities the Future
(1) (2) (3) (4) (5) (6)

IGNITE Exposure 0.233** -0.126 0.087 0.093 -0.051 -0.103

(0.112) (0.214) (0.117) (0.242) (0.126) (0.256)
IGNITE Exposure × Months in Jail 0.187* -0.031 0.039
(0.094) (0.103) (0.102)

Control Mean 0.333 0.333 0.656 0.656 0.656 0.656

Observations 87 87 62 62 62 62
Notes: This table reports estimated effects of IGNITE exposure on the binary outcomes of a community
individual having a positive views of law enforcement (Columns 1 and 2), engagement in positive activities
post-incarceration (Columns 3 and 4), and hopefulness about the future post-incarceration (Columns 5
and 6). A survey wave indicator is included as a control in all specifications and Columns 2, 4, and 6
additionally control for months in jail. Questions on post-incarceration outcomes were only asked to those
who had themselves or had a close relation incarcerated recently. See Data Appendix D for details on
the outcomes and exposure variable. Robust standard errors are reported in parentheses. *, **, and ***
refer to statistical significance at the 10, 5, and 1 percent level, respectively.

41
 Table 7: Custody Staff Survey Results

High Mean of
IGNITE Exposure Control Group Observations
Outcome Variables (1) (2) (3) (4)
Rehab. Programs Worth Time & Money 0.018 (0.154) 0.611 45
Educ. Programs Worth Time & Money 0.344∗∗ (0.144) 0.333 45
Rehab. as Important as Punishment 0.045 (0.153) 0.611 45
F -Statistic for Joint Test [p-value] 2.278∗ [0.094] 45

Notes: This table reports estimated effects of high vs. low IGNITE exposure on whether custodial
staff in Genesee County Jail agree or strongly agree with a listed statement. The high IGNITE exposure
treatment is an indicator for a respondent answering the question “How often do you interact with inmates
in IGNITE?” with “Usually” or “Always.” Robust standard errors are reported in parentheses. *, **,
and *** refer to statistical significance at the 10, 5, and 1 percent level, respectively.

42
Appendix
Table of Contents
A Appendix Figures and Tables A.1

B Institutional Setting Appendix A.37

B.1 Origins of IGNITE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.37
B.2 IGNITE Coursework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.37
B.3 IGNITE Schedule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.38
B.4 IGNITE Expenses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.38
B.5 IGNITE Graduation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.38
B.6 IGNITE Culture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.38

C Data Appendix A.39

C.1 Construction of Sample and Key Variables . . . . . . . . . . . . . . . . . . . . . . . A.39
C.2 Other Administrative Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.40
C.3 Saginaw County Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.41
C.4 Sentiment Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.41

D Survey Appendix A.42

E Econometric Appendix A.43

E.1 Derivation of Equations (2) and (3) . . . . . . . . . . . . . . . . . . . . . . . . . . . A.43
E.2 Difference-in-IVs With Nonlinear Causal Effects . . . . . . . . . . . . . . . . . . . . A.43
E.3 Identification of ACR Weights and Complier Characteristics . . . . . . . . . . . . . A.45

F Literature Comparison A.46

F.1 Time in Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.46
F.2 Time Incarcerated . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.47
F.3 Alternatives to Incarceration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.49
A Appendix Figures and Tables
Appendix Figure A1: Residential Address of Individuals Booked into Genesee County Jail

(a) Share of Arrests, Pre-IGNITE

0%
0.0−0.3%
0.3−0.6%
0.6−0.9%
0.9−1.2%
1.2−1.5%
1.5−1.8%
1.8−2.1%
2.1−2.4%
>2.4%

(b) Share of Arrests, Post-IGNITE

0%
0.0−0.3%
0.3−0.6%
0.6−0.9%
0.9−1.2%
1.2−1.5%
1.5−1.8%
1.8−2.1%
2.1−2.4%
>2.4%

Notes: This figure shows the census tract of residence for individuals booked into Genesee
County Jail before and after the start of IGNITE. The map includes Genesee County and
adjacent counties (Lapeer, Tuscola, Saginaw, Shiawassee, Livingston, and Oakland). Borders
of Genesee County and Flint are outlined.

A.1
 Appendix Figure A2: Court Processes in Genesee County

(a) Misdemeanors

(b) Felonies

Notes: This figure illustrates the court process of typical cases in Genesee County by case
type. Red arrows indicate the stages at which District Court delays are possible. Purple
arrows indicate the stages at which Circuit Court Delays are possible.

A.2
 Appendix Figure A3: Photos of IGNITE

(a) Classroom

(b) Graduation

Notes: Panel A shows individuals participating in the IGNITE pro-

gram within Genesee County Jail, from NPI and NSA (2023). Panel
B shows an IGNITE graduation ceremony, from May (2021).

A.3
 Appendix Figure A4: Budget of Genesee County Corrections Division

20

Total appropriations (mil. USD)

18

16

14

12
2015-2016 2016-2017 2017-2018 2018-2019 2019-2020 2020-2021 2021-2022
Fiscal year
Notes: This figure plots the total appropriations of the Genesee County Sheriff Corrections
Division from fiscal year 2015-2016 to fiscal year 2021-2022. Data for fiscal years 2015-2016 to
2018-2019 are available at Genesee County Controller’s Office (2019) and data for fiscal years
2019-2020 onward are available at Genesee County Controller’s Office (2022). We subtract
capital outlay from total appropriations in fiscal years 2019-2020 onward for consistency with
fiscal years 2016-2016 to 2018-2019.

A.4
 Appendix Figure A5: Examples of IGNITE Programming Tracks

Notes: This figure shows example tailoring of IGNITE education to incarcerated individuals with different
educational backgrounds and goals. Source: NPI and NSA (2023).

A.5
 Appendix Figure A6: Examples of IGNITE Daily Schedules

Notes: This figure shows an example schedule of IGNITE program times and other Genesee County Jail
activities, by the cell floor of incarcerated individuals. Source: NPI and NSA (2023)

A.6
 Appendix Figure A7: Example ROA

(a) Case Information

(b) Court Delay

Notes: Panel A shows an example Genesee County ROA. Panel B shows an example court
delay as identified in the ROA description.

A.7
 Appendix Figure A8: Example Kites Messages

Example 1. “when will i see the judge for my court date really??? i
had an original court date to see judge kelly on 3/1/22 it got
cancelled along with yesterdays i guess??? what's up??”

Example 2. “when is my next court date or when do i get released”

Example 3. “do i have any new court dates? thank u.....”

Notes: This figure shows example Kites text messagess sent by incarcerated
individuals to custody staff in Genesee County Jail.

A.8
 Appendix Figure A9: Survey Recruitment Flyers

(a) Community Survey (b) Custody Staff Survey

Flint Community Correctional Facility

Survey Staff Survey
We Want to Hear From You! We Want to Hear From You!
The survey will ask questions about your views on the criminal The survey will ask questions about your experience working
justice system and the hopes and everyday activities of you or in the correctional facility as well as views on rehabilitation.
your loved ones.
Follow These Steps: Take the Survey
Follow These Steps: Take the Survey
1. Scan the QR code on your phone.
A.9

1.Scan the QR code on your phone 2. Complete survey immediately and

2.Complete survey immediately and in in one sitting (takes about 15
one sitting (takes about 15 minutes) minutes).
3.At the end of the survey, click 3. At the end of the survey, click
“Redeem eGift Card” “Redeem eGift Card.”
4.You will be directed to a $25 Walmart 4. You will be directed to a $25
eGift Card page Walmart eGift Card page.
5.Redeem by entering the eGift Card 5. Redeem by entering the eGift Card
Number and pin on Walmart.com or Number and pin on Walmart.com
scanning the bar code in store or scanning the bar code in store.

The survey is only for you and no one else. The survey is only for you and no one else.

Notes: This figure shows the recruitment flyers used for the Flint Community Survey and Genesee County Jail Custody Staff Survey. The QR
codes on the actual flyers linked to a personal survey page; the codes in this figure link to a survey demo.
 Appendix Figure A10: Main Analysis Sample Construction

Excluded:
All arrests in a) Booked before Jan 2016 (14%)
Jan 2015-Dec 2022 b) Booked after May 2022 (5%)

Booked in Jan
2016-May 2022 Excluded: Not charged (52%)
(81%)

Excluded:
a) Not Residing in Michigan (4%)
b) Residing in Michigan but race
Charged (48%)
unknown (1%)
c) Not yet sent to prison or released
for at least three months (1%)

Estimate Sample
(94%, booked in Jan 2016-May 2022
unique inmate-episodes = 23,610,
unique inmates = 14,794)

Notes: This figure summarizes the construction of our main analysis sample. See Appendix
C for more details.

A.10
 Appendix Figure A11: Share of Major Misconduct and Medical Incidents

(a) Major Misconduct (b) Medical Incidents

1 1
Share major misconduct among all incidents

Share medical incidents among all incidents

.8 .8

.6 .6

.4 .4

.2 .2

0 0
2016m1 2017m1 2018m1 2019m1 2020m1 2021m1 2022m1 2016m1 2017m1 2018m1 2019m1 2020m1 2021m1 2022m1
Month Month

Notes: Panel A plots the monthly share of major misconduct among all misconduct incidents in Genesee County
Jail. Panel B plots the monthly share of medical incidents among all incidents. The vertical dashed lines indicates
the start of IGNITE (September 2020). The horizontal lines indicate best-fit trends.

A.11
 Appendix Figure A12: Chromebook Usage by Floor

3rd Floor 4th Floor 5th Floor

8000
Number of Login Activities

6000

4000

2000

0
6 AM 9 AM 12 PM 3 PM 6 PM 6 AM 9 AM 12 PM 3 PM 6 PM 6 AM 9 AM 12 PM 3 PM 6 PM

Login Time

On-IGNITE Hours Off-IGNITE Hours

One Hour after IGNITE

Notes: This figure plots the distribution of login activities on IGNITE Chromebooks by time
of day and jail cell floor. Chromebook usage data come from Mt. Morris Consolidated Schools.

A.12
Appendix Figure A13: Average Causal Response Weights

(a) Estimated Weights

.12

Pr(incarcerated for at least t months)

.1

Rescheduling Effect on
.08

.06

.04

.02

0 5 10 15 20 25 30 35
Months

Pre-IGNITE Post-IGNITE

(b) Pre- vs. Post-IGNITE Difference

.12

.09
Pr(incarcerated for at least t months)
Rescheduling Effect on

.06

.03

-.03

-.06
0 5 10 15 20 25 30 35
Months

Notes: Panel A plots the estimated Average Causal Response (ACR)

weights for our baseline IV specification, following Appendix E. Each
point is an estimated coefficient from regressing indicators for whether
time in jail exceeds t months on the court delay instrument, for t =
1, . . . , 35. All regressions include the design and auxiliary controls
discussed in the main text. Panel B plots the difference between the
pre- and post-IGNITE estimates. Dashed lines show 95% confidence
intervals from individual-clustered standard errors.

A.13
Appendix Figure A14: First-Stage Effects of Court Delays by Booking Month

Start of IGNITE

1
Months in jail

-1
2016m1 2017m7 2019m1 2020m7 2022m1
Booking month

Notes: This figure plots covariate-adjusted monthly average differences in time in

jail between individuals who do and do not experience a court delay. Each point
indicates the coefficient from regressing time in jail on a court delay indicator
and covariates among individuals booked in a given month. Covariates include
the design and auxiliary controls discussed in the main text. The fitted line is
obtained by a local linear regression with a rule-of-thumb bandwidth, weighting by
the number of incarcerated individuals booked in a given month. Shading indicates
95% confidence intervals derived from individual-clustered standard errors. The
vertical line indicates the beginning of the IGNITE program in September 2020.

A.14
 Appendix Figure A15: Heterogeneity by Incarcerated Individual Characteristics

(a) Effect of Months in IGNITE on Misconduct (b) Effect of Months in IGNITE on Recidivism

Black Black
White White

Male Male
Female Female

Age 17-24 Age 17-24

Age >24 Age >24

Prior Offense Prior Offense

No Prior Offense No Prior Offense

High Lead Exposure High Lead Exposure

Low Lead Exposure Low Lead Exposure

-.4 -.2 0 .2 .4 -.3 -.15 0 .15 .3

(c) Effect of Months in Jail on Misconduct (d) Effect of Months in Jail on Recidivism

Black Black
White White

Male Male
Female Female

Age 17-24 Age 17-24

Age >24 Age >24

Prior Offense Prior Offense

No Prior Offense No Prior Offense

High Lead Exposure High Lead Exposure

Low Lead Exposure Low Lead Exposure

-.4 -.2 0 .2 .4 -.3 -.15 0 .15 .3

Notes: This figure plots IV estimates and associated 95% confidence intervals for the effects of months in
IGNITE or jail on either weekly major misconduct or three-month recidivism by individual characteristics.
Estimates are obtained by interacting either treatment with each observable characteristic as described in
the main text. The characteristics are an indicators for race, sex, age, having a prior offense, and being in
the fourth quartile of census tracts by elevated blood lead levels. All specifications include the design and
auxiliary controls discussed in the main text. 95% confidence intervals are derived from individual-clustered
standard errors.

A.15
 Appendix Figure A16: Heterogeneity by Predicted Recidivism Risk

(a) Misconduct (b) Recidivism

(High) 4 (High) 4
Quartiles of predicted recidivism risk

Quartiles of predicted recidivism risk

3 3

2 2

(Low) 1 (Low) 1

-.6 -.3 0 .3 -.6 -.3 0 .3

Notes: This figure plots IV estimates and associated 95% confidence intervals for the effects of months
in IGNITE by quartiles of predicted three-month recidivism risk. Panel A reports results for weekly major
misconduct while Panel B reports results for for three-month recidivism. Predicted recidivism risk is obtained
by a logit regression on the 2015 holdout sample. All specifications include the design and auxiliary controls
discussed in the main text. 95% confidence intervals are derived from individual-clustered standard errors.

A.16
 Appendix Figure A17: Heterogeneity by Predicted Time-in-Jail Quartiles

(a) Effect of Months in IGNITE on Misconduct (b) Effect of Months in IGNITE on Recidivism

(High) 4 (High) 4
Quartiles of predicted time in jail

Quartiles of predicted time in jail

3 3

2 2

(Low) 1 (Low) 1

-1.5 -.75 0 .75 1.5 -.6 -.3 0 .3 .6

(c) Effect of Months in Jail on Misconduct (d) Effect of Months in Jail on Recidivism

(High) 4 (High) 4
Quartiles of predicted time in jail

Quartiles of predicted time in jail

3 3

2 2

(Low) 1 (Low) 1

-1.5 -.75 0 .75 1.5 -.6 -.3 0 .3 .6

Notes: This figure plots IV estimates and associated 95% confidence intervals for the effects of months in
IGNITE or jail by quartiles of predicted time in jail. Panel A reports results for weekly major misconduct
outcome while Panel B reports results for for three-month recidivism. Predicted time in jail is produced by a
regression on the 2015 holdout sample. All specifications include the design and auxiliary controls discussed
in the main text. 95% confidence intervals are derived from individual-clustered standard errors.

A.17
Appendix Figure A18: Kites Court Closure Event Study

(a) Court Closure

.25

.15

Court-related texts

.05

-.05

-.15
-4 -3 -2 -1 0 1 2 3 4
Week since court closure

(b) Placebo

.25

.15
Court-related texts

.05

-.05

-.15
-4 -3 -2 -1 0 1 2 3 4
Week since placebo date

Notes: This figure plots coefficients and 95% confidence intervals from
an event study regression of an indicator for a Kites message includ-
ing the words “talk,” “speak,” “need,” “can,” “please,” “court,” or
“judge.” Panel A uses one week prior to the court closure due to
COVID-19 (March 17, 2020) as the base period. Panel B uses a
placebo date (May 17, 2019). 95% confidence intervals are derived
from individual-clustered standard errors.

A.18
Appendix Figure A19: Effects of IGNITE on Recidivism: Event Study Estimates

.1

.05
3-Month Recidivism

-.05

-.1

-.15
-12 -8 -4 0 4 8
4-Month Bins

Notes: This figure plots coefficients and 95% confidence intervals from an event study regres-
sion of three-month recidivism, measured in either Genesee or Saginaw County. The treatment
is an indicator for the individual being booked in Genesee County and the base period is De-
cember 2019. Outcomes are binned in four-month intervals. The shaded area denotes the
period when individuals booked pre-IGNITE saw nontrivial exposure to IGNITE after its
launch in September 2020.

A.19
 Appendix Figure A20: Probability of Exposed to IGNITE

1 Start of IGNITE

.8
Share exposed to IGNITE

.6

.4

.2

2016m1 2017m1 2018m1 2019m1 2020m1 2021m1 2022m1 2023m1

Booking month
Notes: This figure plots the probability of being exposed to IGNITE (i.e., being in Genesee
County Jail on or after September 2020) by booking month.

A.20
 Appendix Figure A21: Literature Comparison: Alternatives to Incarceration

-20.3% US | Jail
Williams and Weatherburn 2022 Non-US | Jail
US | Non-Jail
Non-US | Non-Jail
-14.0%
Blattman, Jamison, and Sheridan 2017

-5.2%
Di Tella and Schargrodsky 2013

-3.6%
Bhatt et al. 2023

-1.6%
Henneguelle, Monnery, and Kensey 2016

-0.7%
Lavecchia, Oreopoulos, and Spencer 2024

-0.0%
Packham and Slusky 2023

-90 -60 -30 0 30 60 90

Effect Size (%)

Notes: This figure summarizes estimated treatment effect sizes from the literature on alternatives to
incarceration. See Appendix F for details on the papers and effect size construction. We distinguish
between U.S. vs. non-U.S. studies and studies in jail vs. non-jail contexts. Each point indicates the
estimated effect of treatment on recidivism as a percent of the control mean. When possible, we use
one-year recidivism outcomes and scale effects by time in treatment. 95% confidence intervals are shown
around each point.

A.21
Appendix Figure A22: Custody Staff Views on Education Programs in Jail by IGNITE Exposure

1
Share agree
p-value = 0.022
Educ. programs in jail are worth time and money
Share neutral
p-value = 0.008
.75

.5 Share disagree
p-value = 0.726

.25

0
Low High Low High Low High

IGNITE exposure

Notes: This figure plots means and 95% confidence intervals for responses in the Genesee
County Jail Custody Staff Survey to the question of whether “Education programs in jail
are worth the time and money.” High IGNITE Exposure is an indicator for the respondent
answering the question “How often do you interact with inmates in IGNITE?” with “Usually”
or “Always.”

A.22
 Appendix Figure A23: Kites Sentiment Analysis

(a) Positive/Negative/Neutral Sentiment

.43
Share negative words
p-value = 0.643
Share neutral words
p-value = 0.253
.38
Share positive words
p-value = 0.023
Share of words

.33

.28

.23
Pre-IGNITE Post-IGNITE Pre-IGNITE Post-IGNITE Pre-IGNITE Post-IGNITE

(b) Emotion Prevalence, Post- vs. Pre-IGNITE

.08

.04
Change in word prevalence

-.04

-.08

Anticipation Anger Fear Sadness Disgust Joy Surprise Trust

Notes: This figure summarizes the sentiment of Genesee County Jail Kites mes-
sages, pre- and post-IGNITE, according to the NRC Word-Emotion Association
Lexicon from Mohammad and Turney (2010). Panel A plots the share of words in
text messages categorized as positive, negative, and neutral for each time period.
The p-values are obtained from a word-level regression on a post-IGNITE indica-
tor weighted by word frequency with standard errors clustered at the individual
level. Panel B plots the coefficient on post-IGNITE from analogous regressions of
an indicator for the word being associated wih an emotion. Emotions are catego-
rized as negative (blue) and positive (red) by ChatGPT.

A.23
 Appendix Table A1: Effects on Secondary Outcomes

Months in IGNITE Months in Jail

(1) (2)
Panel A: Alternative Recidivism/Misconduct Measures
Recharged -0.060*** 0.013
(N = 22, 191) (0.027) (0.014)
Reconvicted -0.051** 0.028
(N = 22, 191) (0.021) (0.018)
Minor Misconduct -0.021** 0.011***
(N = 23, 610) (0.010) (0.004)
More Serious Misconduct -0.029** 0.011***
(N = 23, 610) (0.012) (0.005)

Panel B: Other Outcomes

Tether 0.002 -0.000
(N = 23, 610) (0.009) (0.007)
Bail Posted -0.078 0.154***
(N = 23, 610) (0.053) (0.048)
Sentenced to Prison 0.005 -0.024
(N = 23, 610) (0.020) (0.017)
Convicted -0.048 0.190***
(N = 23, 610) (0.060) (0.055)
Released to Rehab. Centers -0.001 -0.003
(N = 23, 610) (0.010) (0.008)
Suicide -0.008 -0.020
(N = 23, 610) (0.030) (0.022)
Other Medical -0.041 -0.006
(N = 23, 610) (0.040) (0.027)
Notes: This table reports reduced form and IV estimates for alterna-
tive measures of recidivism and misconduct, as well as other outcomes.
Panel A reports results for Recharged (charged within three months
since release), Reconvicted (convicted within three months since re-
lease), Minor Misconduct, and More Serious Misconduct (as defined in
the text). Panel B reports results for Tether (an indicator for released
with electronic monitoring), Bail Posted (an indicator for released on
bond), Sentenced to Prison (an indicator for sentenced to prison), Con-
victed (an indicator for either found or plead guilty), Released to Reha-
bilitation Centers (an indicator for released to rehabilitation centers),
Number of Suicide Attempts per Week, and Number of Other Medical
Incidents per Week. All rows include the design controls and auxil-
iary controls discussed in the main text. Individual-clustered standard
errors are reported in parentheses. *, **, and *** refer to statistical
significance at the 10, 5, and 1 percent level, respectively.

A.24
 Appendix Table A2: Representativeness of Test Takers

Overall Difference Standard

Mean in Means Error
(1) (2) (3)
Panel A: Individual Characteristics
Female 0.209 0.004 (0.019)
Age 25-34 0.388 -0.024 (0.023)
Age 35-44 0.247 0.004 (0.021)
Age 45-54 0.116 -0.030* (0.016)
Age 55-64 0.058 -0.020* (0.011)
Age 65+ 0.012 -0.004 (0.005)
Black 0.532 -0.067 (0.053)
Has a Public Defender 0.299 0.026 (0.020)

Panel B: Census Tract Characteristics

Share with Elevated Blood Lead Level 0.031 -0.020 (0.014)
Share Black 0.417 -0.034 (0.024)
Share High School Graduate or Higher 0.849 -0.039 (0.025)
Log Median Household Income 10.336 -0.231 (0.151)
F -Statistic for Joint Test [p-value] 1.099 [0.122]
Observations 4451
Notes: This table summarizes the sample of incarcerated individuals who can be linked to the
Mt. Morris data and compares those who do and do not have CASAS test scores. See Table 1
for details on the variables. standard errors are reported in parentheses. *, **, and *** refer
to statistical significance at the 10, 5, and 1 percent level, respectively.

A.25
Appendix Table A3: IGNITE Exposure Balance Test: Community Survey
Overall Difference Standard
Mean in Means Error
(1) (2) (3)
Panel A: Site
Barbershop 0.092 -0.039 (0.063)
Church 0.230 0.158 (0.102)
General Store 0.253 -0.132 (0.091)
Other 0.414 0.031 (0.112)
Social Security Office 0.011 -0.017 (0.018)

Panel B: Demographics
Male 0.287 0.121 (0.107)
Black 0.793 -0.041 (0.095)
White 0.207 0.041 (0.094)
Hispanic 0.023 -0.035 (0.025)
Completed college or more 0.391 -0.086 (0.109)
High school degree or GED 0.322 0.067 (0.106)
No high school degree 0.034 0.049 (0.049)
Some college 0.253 -0.030 (0.098)
Age 18-24 0.080 0.080 (0.070)
Age 25-34 0.299 0.052 (0.104)
Age 35-44 0.184 -0.026 (0.087)
Age 45-54 0.264 -0.047 (0.099)
Age 55-64 0.149 -0.075 (0.076)
Age 65+ 0.023 0.016 (0.038)
F -Statistic for Joint Test [p-value] 0.851 [0.626]
Observations 87
Notes: This table summarizes the Flint Community Survey sample and reports
balance tests for IGNITE exposure. Column 1 reports sample means and Columns
2 and 3 report regression coefficients and associated standard errors from regressing
respondent characteristics on an indicator for whether the individual was person-
ally released from Genesee County Jail after September 2020 or had a friend or
family member who was released from Genesee County Jail after September 2020.
Panel A summarizes the survey recruitment location. Panel B reports results for
respondent-level demographics, including sex, race, educational attainment, and
age. All estimations include a survey cohort wave control. Robust standard errors
are in parentheses. *, **, and *** refer to statistical significance at the 10, 5, and
1 percent level, respectively.

A.26
Appendix Table A4: IGNITE Exposure Balance Test: Custody Staff Survey

Overall Difference Standard

Mean in Means Error
(1) (2) (3)
Female 0.267 -0.006 (0.135)
Black 0.022 0.035 (0.035)
Age 18-24 0.089 -0.039 (0.090)
Age 25-34 0.289 -0.267∗ (0.143)
Age 35-44 0.356 0.158 (0.141)
Age 45-54 0.200 0.128 (0.110)
Started Work after IGNITE 0.400 0.009 (0.153)
Started Work after COVID-19 0.422 0.044 (0.154)
F -Statistic for Joint Test [p-value] 1.067 [0.408]
Observations 45
Notes: This table summarizes the Genesee County Jail Custody Staff Survey
and reports balance tests for high IGNITE exposure. Specifically, Column 1 re-
ports sample means and Columns 2 and 3 report regression coefficients and asso-
ciated standard errors from regressing respondent characteristics on an indicator
for whether the individual answered the question “How often do you interact with
inmates in IGNITE” with “Usually” or “Always.” Robust standard errors are in
parentheses. *, **, and *** refer to statistical significance at the 10, 5, and 1
percent level, respectively.

A.27
 Appendix Table A5: Summary Statistics, Main Analysis Sample

Mean SD N
(1) (2) (3)
Panel A: Instrument and Outcomes
Any District Court Delay 0.381 (0.486) 23,610
Months in IGNITE 0.434 (2.091) 23,610
Months in Jail 1.558 (4.212) 23,610
Ever Rebooked in 3 Months after Release 0.175 (0.380) 22,147
Convicted 0.495 (0.500) 23,610
Sentenced to Prison 0.060 (0.238) 23,573
Any Major Misconduct 0.092 (0.289) 23,610
Any Medical or Suicidal Incident 0.061 (0.240) 23,610
Any Incident 0.160 (0.366) 23,610
Panel B: Individual and Case Characteristics
Female 0.240 (0.427) 23,610
Age 25-34 0.378 (0.485) 23,610
Age 35-44 0.225 (0.418) 23,610
Age 45-54 0.122 (0.327) 23,610
Age 55-64 0.058 (0.234) 23,610
Age 65+ 0.009 (0.092) 23,610
Black 0.534 (0.499) 23,610
Booked in Past Year 0.433 (0.496) 23,610
Felony Charge 0.534 (0.499) 23,610
Number of Charges 1.385 (0.867) 23,610
Panel C: Census Tract Characteristics
Share with Elevated Blood Lead Level 0.031 (0.028) 22,318
Share Black 0.429 (0.354) 22,320
Share High School Graduate or Higher 0.848 (0.066) 22,320
Log Median Household Income 10.322 (0.425) 22,318
Missing Census Tract Information 0.055 (0.228) 23,610
Notes: This table reports the sample mean, standard deviation, and number of non-
missing observations of variables in the main analysis sample.

A.28
 Appendix Table A6: Differential Attrition
Observed Observed Observed Observed
for 3 Months for 6 Months for 9 Months for 12 Months
after Release after Release after Release after Release
(1) (2) (3) (4)
Court Delay 0.000 -0.001 -0.004*** -0.011***
(0.000) (0.001) (0.001) (0.002)

Control Mean 1.000 0.998 0.995 0.976

Observations 23,610 23,610 23,610 23,610
Notes: This table reports differential attrition by the court delay instrument over various
time horizons. Each column reports a regression of an indicator that equals one if an
individual is observed for t months after release on the Court Delay instrument for
t = 3, 6, 9, 12. All regressions include the design and auxiliary controls discussed in the
main text. Individual-clustered standard errors are reported in parentheses. *, **, and
*** refer to statistical significance at the 10, 5, and 1 percent level, respectively.

A.29
 Appendix Table A7: First-Stage Effects of Court Delays
Months Months
in IGNITE in Jail
(1) (2) (3) (4)
Court Delay × Post-IGNITE 0.521*** 0.519*** 0.120 0.113
(0.096) (0.096) (0.119) (0.119)
Court Delay 0.128*** 0.130*** 0.393*** 0.401***
(0.029) (0.029) (0.073) (0.073)

Control Mean 0.102 0.102 1.311 1.311

F -Stat.: Court Delay × Post-IGNITE 79.909 79.611 79.909 79.611
F -Stat.: Court Delay 55.159 56.549 55.159 56.549
Design Controls Yes Yes Yes Yes
Auxiliary Controls No Yes No Yes
Observations 23,610 23,610 23,610 23,610
Notes: This table reports first-stage estimates for months in IGNITE (Columns 1 and 2) and
months in jail (Columns 3 and 4). All columns include the design controls discussed in the
main text. Columns 2 and 4 also include the auxiliary controls. Individual-clustered standard
errors are reported in parentheses. F-statistics are from Sanderson and Windmeijer (2016).
*, **, and *** refer to statistical significance at the 10, 5, and 1 percent level, respectively.

A.30
 Appendix Table A8: Judge IV Estimates
Misconduct Recidivism
Pre-IGNITE Post-IGNITE Pre-IGNITE Post-IGNITE
(1) (2) (3) (4)
Months in Jail 0.010 -0.011 0.073 -0.054**
(0.026) (0.024) (0.062) (0.024)

Post-Pre -0.021 -0.127*

(0.035) (0.067)

Design Controls Yes Yes Yes Yes

Auxiliary Controls Yes Yes Yes Yes
Observations 19093 4751 17887 4484
Notes: This table reports judge IV estimates for the effects of time in jail on weekly major
misconduct and three-month recidivism in the pre- and post-IGNITE periods, estimated as
in the main text. All columns include the design controls and auxiliary controls discussed in
the main text. Individual-clustered standard errors are reported in parentheses. *, **, and
*** refer to statistical significance at the 10, 5, and 1 percent level, respectively.

A.31
 Appendix Table A9: Judge IV Specification Tests
Number of Spline Knots
One Two Three Four
(1) (2) (3) (4)
Panel A: Pre-IGNITE
Test Statistic 32.4 31.9 31.1 30.9
Deg. of Freedom 17 16 15 14
p-value 0.013 0.010 0.008 0.006
Panel B: Post-IGNITE
Test Statistic 27.8 24.6 23.8 25.3
Deg. of Freedom 14 13 12 11
p-value 0.019 0.020 0.019 0.008
Panel C: Overall
Test Statistic 50.2 48.7 50.6 50.5
Deg. of Freedom 17 16 15 14
p-value <0.001 <0.001 <0.001 <0.001
Notes: This table reports the results of the tests of judge IV monotonic-
ity and exclusion proposed by Frandsen, Lefgren and Leslie (2023),
computed both for the overall sample and separately by pre- versus
post-IGNITE. Test statistics are based on quadratic b-spline estimates
of the relationship between recidivism outcomes and District Court
judge leniency, with the number of knots specified in each column. All
specifications include the design controls.

A.32
 Appendix Table A10: County Difference-in-Differences Estimates

Recidivism Measured Recidivism Measured

in Each County in Both Counties
(1) (2) (3) (4)
Post × Genesee -0.039*** -0.026*** -0.035*** -0.023***
(0.007) (0.007) (0.007) (0.007)
Genesee 0.069*** 0.040*** 0.072*** 0.042***
(0.004) (0.003) (0.004) (0.004)
Post 0.006 -0.020 0.004 -0.023
(0.010) (0.016) (0.010) (0.016)

Control Mean 0.182 0.182 0.189 0.189

Ind. Chars. × Post No Yes No Yes
Observations 40,655 40,655 40,655 40,655
Notes: This table reports the results from regressing, in a sample of individuals
booked in either Genesee County or Saginaw County, three-month recidivism on
an indicator for being booked in Genesee County, an indicator for being booked
after January 2020, and their interaction. Columns 1 and 2 measure recidivism
separately for Genesee and Saginaw Counties while Columns 3 and 4 combine
recidivism in both counties. Month and year fixed effects are included in all
estimates. Columns 2 and 4 additionally control for individual characteristics
(indicators for Black, female, age groups, whether an individual was booked in the
past year) interacted with the post indicator. The control mean is for Genesee
County in the pre-period. Individual-clustered standard errors are reported in
parentheses. *, **, and *** refer to statistical significance at the 10, 5, and 1
percent level, respectively.

A.33
 Appendix Table A11: Social Cost of Crime Effects

3 Months 6 Months 9 Months 12 Months

(1) (2) (3) (4)
Months in IGNITE -2957.46** -3943.38** -5293.63*** -5614.70**
(1238.02) (1653.90) (1972.45) (2197.60)

Control Complier Mean 12212.62 26148.83 39364.44 45535.49

Observations 22,191 21,525 21,139 20,766
Notes: This table reports IV estimates for the social costs induced by future crimes. Future
crimes are obtained from ROAs associated with observed future bookings and divided into the
following categories: DUIs, drug offenses, motor vehicle offenses, persons offenses, property
offenses, public order offenses, weapons offenses, and other offenses. Within each of these
crime types, we take the lowest social cost estimate from Miller et al. (2021). We then use
the total social cost (sum of frequency of crime × cost crime) as an outcome in the main IV
specification discussed in the text. Individuals sentenced to prison or not observed for enough
time since their last release are excluded from the analysis. All columns include the design
controls and the auxiliary controls discussed in the main text. Individual-clustered standard
errors are reported in parentheses. *, **, and *** refer to statistical significance at the 10, 5,
and 1 percent level, respectively.

A.34
Appendix Table A12: Effects of High IGNITE Exposure: Views on Own Work Experience

High Mean of
IGNITE Exposure Control Group Observations
Outcome Variables (1) (2) (3) (4)
Satisified with Own Job 0.135 (0.151) 0.588 45
Benefit Package is Competitive 0.044 (0.155) 0.353 45
Pay is High Enough -0.013 (0.133) 0.235 45
Joint F -Test [p-Value] 0.306 [0.821] 45
Notes: This table reports estimated effects of high vs. low IGNITE exposure on whether custodial staff
in Genesee County Jail agree or strongly agree with the listed statement. The high IGNITE exposure
treatment is an indicator for a respondent answering the question “How often do you interact with inmates
in IGNITE?” with “Usually” or “Always.” Robust standard errors are reported in parentheses. *, **,
and *** refer to statistical significance at the 10, 5, and 1 percent level, respectively.

A.35
 Appendix Table A13: Staff Quotes on IGNITE

Selected Quotes
Positive “Ignite [sic] has changed the culture at the Genesee County jail.”
“It has given inmates something to focus on and invest instead of
worrying about drama and wrong doings. Their mind is being put to
use and the rewards helps encourage them to do better.”
“It has created a safer work place.”
“I’ve seen inmates that are constant problems change their behavior
dramatically once they have seen the benefits of Ignite. Ignite seems
to provide self-worth to people that maybe never had any.”
“I believe it has helped to bring the unity and humanity aspect back
to the jail.”

Neutral “The program is a great idea, however, we seem to be neglecting other

areas of jail operations to cater to IGNITE.”
“It makes things a little more difficult for staff and the security of the
facility. Helps some with the behavioral aspects.”
“Ignited [sic] sometimes made it harder for us to do our jobs and
decrease the Deputy security and safety. But at the same time it has
reduced some recidivism.”

Negative “I have experience added [sic] mandatory overtime, time away from
my family. We have had security breaches putting the safety and
security of the facility at risk.”
“There are additional duties placed on the housing unit deputies.
Outside of that, IGNITE does not directly effect [sic] my day to day
experience or operation.”
“Ignite has taken over most of the priorities at the jail. Ignite comes
before anything else.”

Notes: This table lists example custody staff quotes on their overall views on IGNITE, grouped by
positive, neutral, or negative views, from the Genesee County Jail Custody Staff Survey. Quotes are
selected from the answers to the question: “Can you describe any ways in which IGNITE has changed
your experience working at the Genesee County Jail?”

A.36
B Institutional Setting Appendix
This appendix details the IGNITE program and its launch. Most information comes from original
field interviews with Genesee County community members and custody staff.

B.1 Origins of IGNITE

Prior to IGNITE, there were minimal educational opportunities available to incarcerated individuals
in Genesee County Jail. Mt. Morris Consolidated Schools, the school partner for Genesee County
Jail since 2004, offered a small GED prep program that worked with 2 to 20 people at a time.
Participation in the GED prep program was limited to individuals chosen by custody staff. There
were no educational activities or training opportunities for those who needed more elementary
educational training or who had already obtained their GED or high school diploma.
IGNITE was a concept originated by Genesee County Sheriff Chris Swanson. A long-time
undersheriff for the county, Sheriff Swanson was elected in January 2020 into an interim-sheriff po-
sition to complete the 1-year remaining term of retiring Sheriff Robert Pickell. In July 2019, Sheriff
Swanson planned a new educational program within the jail which was informed by his master’s
thesis at the University of Michigan. Having seen three generations of families go through the jail
system over his career in corrections, Sheriff Swanson was determined to transform incarceration
through education.
Sheriff Swanson had originally hoped to start an education program in the jail later in his
tenure, but the pandemic and the murder of George Floyd in May 2020 accelerated these plans.35
Escalating racial tensions, difficult community-police relations, and a “broken” jail system led
Sheriff Swanson and jail administrators to launch IGNITE on September 8th, 2020. According to
the National Policing Institute and National Sheriff’s Association (NPI and NSA, 2023):

“The goal of the program is to create a structured and safe environment that is conducive
to education. Having this type of environment in place can facilitate the implementation
of more diverse educational programming, which can help re-energize inmates, motivate
them once again, and provide them with an education and new opportunities once they
return to their communities.”

B.2 IGNITE Coursework

IGNITE includes traditional educational coursework as well as skill-building and lifestyle promotion.
The traditional educational coursework occurs in the jail’s open activities space. Incarcerated
individuals take courses offered through Odysseyware software on Chromebooks. These courses
include enrichment classes, GED preparation, high school classes, as well as classes in basic reading,
writing, and arithmetic. Prior to beginning coursework, incarcerated individuals take a placement
test using the CASAS testing system. After individuals are placed into a track, Mt. Morris teachers
35
Viral YouTube videos even showed Sheriff Swanson marching through the streets of Flint alongside protesters.
See, e.g., https://www.youtube.com/watch?v=fBo9E9Be kw (7NEWS Australia, 2020).

A.37
monitor their progress and provide support. Mt. Morris also conducts post-assessment testing with
CASAS. Ideally, all incarcerated individuals would receive both pre and post-assessments. However,
individuals are often released before post-assessment testing can be arranged.
In addition to the traditional curriculum, IGNITE offers enrichment classes including nutritional
courses, financial literacy, and training for certain trades (e.g., welding) via virtual reality software.
Aramark, a food service provider, also offers a ServSafe food certification course and there are
opportunities to work on obtaining a commercial driver’s license (CDL) and even to complete
college level coursework.

B.3 IGNITE Schedule

IGNITE coursework runs five days a week in two hour-long program blocks. All other jail activities
are paused during IGNITE study time. Tablet computers, provided by ViaPath, are distributed
daily to IGNITE participants who completed their coursework. Tablets are typically available from
5PM to 9PM and used as rewards for IGNITE participation with approved content such as games
and puzzles.

B.4 IGNITE Expenses

ViaPath provides internet and communications technology to Genesee County Jail, including the
Chromebooks and tablets used for IGNITE. The additional costs of the program are partially offset
by service fees paid from commissary funds. Philanthropy also plays a role (e.g., the virtual reality
training system was donated). IGNITE also saves on administrative costs by relying on existing
custody staff to implement the program. Custody staff provide security for teachers, supervise
tablet time, and run the virtual reality system.

B.5 IGNITE Graduation

Genesee County Jail holds graduation ceremonies every 1-3 months to celebrate the achievements of
individuals who participated in IGNITE. As part of the ceremony, the Sheriff hands out certificates
to graduating individuals. Family and friends are invited to attend, with a Deputy reaching out to
them as part of the event organization. Incarcerated individuals wear gowns over their jumpsuits
and listen to a keynote speaker. The commencement ceremonies are frequently broadcast on social
media and covered by the local press.36 .

B.6 IGNITE Culture

IGNITE is described by administrators as a law-enforcement-led program that changes attitudes
and the culture of corrections both directly and indirectly through education and opportunities.
36
An example of an IGNITE graduation ceremony can be viewed at: https://www.abc12.com/video/
genesee-county-sheriffs-office-holds-ignite-graduation/video 7badfb62-21d7-59d5-8eea-a0f0b7faece0.html (ABC 12
News, 2024)

A.38
The jail views IGNITE as a vehicle to rebuild trust in a community that has been rocked by a public
health crisis, fiscal crises, and government corruption. IGNITE purports to institute this cultural
change by giving value and respect to incarcerated individuals and operating under a meritocratic
system where those who work hard can succeed no matter what their background.

C Data Appendix
C.1 Construction of Sample and Key Variables
To construct our analysis sample, we start with the universe of individuals in the JMS who were
booked and detained in Genesee County Jail from January 2016 to May 2022. The JMS data
include case numbers which we match to ROAs from the online Michigan court records database
(Michigan Judiciary, 2024). We scrape records for cases seen in the 67th Judicial District Courts,
where both criminal misdemeanor and felony cases are seen in Genesee County along with traffic
and civil infractions, and in the Seventh Judicial Circuit Court where felony cases are bound over
and tried. These records create a timeline of court appointments for jailed individuals from the
time their case was filed to when it closed, including all hearings, trials, and sentencing motions
and proceedings.
Among the set of individuals booked and charged between January 2016 and May 2022, we
exclude 4% of individuals not residing in Michigan, 1% of individuals with incomplete demographic
information (specifically, race), and 1% of individuals who have not yet been sent to prison or
released for at least three months. This leaves a sample of 23,610 incarceration spells involving
14,794 unique individuals.

Instrument: Court Delay. The ROAs are used to define court delay incidents. In Genesee
County, when a court date is assigned, a scheduling notice appears in the Description section of an
event entry in the ROA, along with the date and time of court appearance in the Comment section
(see Appendix Figure A7). We define a specific court event, such as a pretrial arraignment as
being delayed if, for the same date in the Comments section, a new event appears in the ROA with
“Removed from Calendar” in the Description section.37 We define our instrument by the presence
of any such delays in an individual’s District Court history. We also use the number of delays
across an individual’s court history, as well as the number of delays occurring across individuals on
a given day, to construct controls for certain robustness checks. In another check, we construct an
alternative instrument by the presence of delays in either District or Circuit Court ROAs.

Treatments: Months in Jail/IGNITE. The JMS dataset is used to determine an individual’s

time spent in jail, defined either as the difference between their release date and booking date or the
37
This measure of court delay is not without limitations. Most importantly, we do not capture all forms of possible
court delays which could include adjournments. One reason for excluding adjournments is that they are less likely
to be as-good-as-randomly assigned, since they are more likely to be requested by the prosecutor or defense counsel
than are removed events.

A.39
difference between their transfer date and booking date if the individual experienced an external
transfer to another facility (such as prison or rehabilitation center). In some cases the JMS data
records multiple release dates present for the same booking date; we use the most recent release
date in these cases. For the very small minority of individuals still in jail in May 2023 (the last
month in our sample), we define time spent in jail as the difference between May 2023 and their
booking date. We determine the reason for release through the Release Checklist file in JMS.
To construct our measure of IGNITE exposure, we use the program’s start date of September
8, 2020. We define an individual’s months in IGNITE as the total amount of time in jail occurring
since this date.

Outcomes: Misconduct and Recidivism. We obtain information on within-jail misconduct

incidents from JMS Jail Incident Reports. These reports contains an incident log number, the date
and time of the incident, the name of involved individual(s), the type of incident, the incident code
associated with the action, and a description of the event. These reports also include the location
of the incident, the status or resolution of the incident, and whether force was used.
We define an incident occurring for an individual in the jail if the associated incident log number
is present and non-missing. Incidents are classified as major, minor, medical, or suicide attempt.
We order all incidents by date. The number of incidents and the number of incidents of a single
incident type is defined as the sum of the incidents an individual has during their jail stay. We
divide this number by an individual’s number of weeks in jail to obtain our measures of weekly
misconduct. We further categorize incidents into disobedience and violent incidents, based on the
incident classification numbers used in the incident description.
Our main three-month recidivism outcome is defined as an individual for whether an individual
is rebooked in either Genesee County Jail or Saginaw County Jail within three months after release.
Recidivism outcomes over longer horizons are constructed similarly. In most cases, individuals are
brought to the jail for booking from local municipal arresting agencies throughout the county that
operate independently of the county Sheriff’s office.

C.2 Other Administrative Data

Kites Messaging. We merge in Kites Messaging data, which consists of all time-stamped mes-
sages sent between incarcerated individuals and the monitoring correctional staff, by individual
identification number and jail stay.

Mt. Morris Consolidated Schools. The Mt. Morris Schools education data consists of two
components. The first component consists of the amount of Chromebook usage, organized by
individual-login time. Mt. Morris Schools use Chromebooks to administer classes in Genesee
County Jail and each login reflects a student accessing their schooling program. We link these data
at the individual-case level by first and last name and time of jail stay.

A.40
 The second component consists of individual-level data on incarcerated individuals who partic-
ipated in adult education programming in Genesee County Jail. Variables include High school and
GED completion status, pre- and post-reading and math assessments, and GED subject test scores
and dates. We convert the pre-and-post reading and math scores to their grade-level equivalence
using the derived Grade Level Equivalencies for CASAS standardized exams. We link this dataset
by first name, last name, and date of birth to the JMS data.

ViaPath Data. ViaPath, formally known as Global Tel Link (GTL), provides phone, tablets,
and internet service to the jail. ViaPath data provides individual-level measures of total phone, app,
and video visit usage. We observe the number of logins or calls an individual has ever made and
the total number of logins or minutes used for each ViaPath service. Apps, which are accessible
to incarcerated individuals via tablet computers, are further disaggregated into education and
entertainment. We merge the ViaPath data by individual identifier to the JMS data.

Lead Exposure and Census Data. The residential address of incarcerated individuals recorded
in JMS data are used to link census-tract variables, by zip code. Elevated Blood Lead Level data
comes from the Michigan Department of Health and Human Services. Data are available at https://
catalog.data.gov/no/dataset/leadbloodlevels-2017-bytract-20181129-dec8f (Data Driven Detroit,
2022). We define a missing indicator variable that equals one if an observation is not matched
to a census tract. We obtain the population shares of Black, High School Graduate or Higher, and
the Log Median Household Income from 2016 ACS 5-year Estimates (U.S. Census Bureau, 2016),
linking individuals to the lead exposure data by zipcode of residence.

C.3 Saginaw County Data

For certain robustness checks we construct an analogous analysis sample using JMS data for Saginaw
County Jail from 2016 to 2023. The booking records, organized at the individual-booking level
include the defendant’s full name, date of birth, race and sex identifiers, booking and release dates,
arresting officer, jail classification, and case number.
We merge these data to an analogous court date delay measure for Saginaw County, using case
numbers from the Saginaw County booking records to scrape ROAs from the 10th Judicial Circuit
Court and 70th Judicial District Court. We parse the ROAs in a similar as was done for Genesee
County. Since Genesee County and Saginaw County are served by different courts with their own
practices for describing events, we alter our definition of a court date delay in Saginaw County to
include the use of “Adjourn” or “Resched” in the event description.

C.4 Sentiment Analysis

Variable definition for Table 6 include the following:
• Positive View on Law Enforcement is an indicator for whether the respondent answered
“Somewhat agree” or “Strongly agree” to the following question: “To what extent do you

A.41
 agree or disagree with the following statement: Law enforcement looks out for me and my
community.”

• Engagement in Positive Activities is an indicator for whether the respondent said they were
“Looking for work,” “Taking classes,” “Working for pay that does not involve crime,” or
“Taking care of children or elderly family members,” in response to the question: “Are you
currently doing any of the following [Select all the apply].”

• Hopeful about the Future is an indicator for whether the respondent answered “More Hopeful”
to the question, “Are you more or less hopeful about your future compared to before you were
incarcerated?”

D Survey Appendix
Flint Community Survey. The Flint Community Survey was distributed in Flint, MI by
four local community members between October 2023 and December 2023. Participants were
recruited from locations throughout the the city, including grocery stores, libraries, barbershops,
and churches. An example of the recruitment flyer is shown in Panel A of Appendix Figure A9.
Participants scanned the QR code with their phone to access the survey. After providing informed
consent, the participants were asked about their personal incarceration experience, and if not ap-
plicable, the incarceration experience of close friends and family members. Our primary outcome
of interest was views on local law enforcement. Community members who had personally spent
time in jail were also asked about their current activities. The survey was anonymous. A copy of
the survey can be found at:

https://harvard.az1.qualtrics.com/jfe/form/SV 2fOdxQEsvmAO5U2

Custody Staff Survey. The Custody Staff Survey was distributed as a QR code on a flyer
that was circulated in Genesee County Jail in January 2024 (see Panel B of Appendix Figure A9).
Custody staff were asked about their tenure at the facility, their exposure to IGNITE participants,
and their views on rehabilitation and reform in jails. Given ongoing negotiations with the jail
administration, we included questions on job satisfaction—including whether staff thought they
received a competitive benefits package and whether they thought pay was high enough. The
survey was anonymous. A copy of the survey can be found at:

https://harvard.az1.qualtrics.com/jfe/form/SV 3xQqkubYEsu6zeS

A.42
E Econometric Appendix
E.1 Derivation of Equations (2) and (3)
Equations (2) and (3) follow by substituting the causal model (1) into the two IV estimands:

Cov(Zi , Yi (0) + γi MiJ + βi MiI | Pi = 0) Cov(Zi , γi MiJ | Pi = 0)

β P re = = ,
Cov(Zi , MiJ | Pi = 0) Cov(Zi , MiJ | Pi = 0)

and

Cov(Zi , Yi (0) + γi MiJ + βi MiI | Pi = 1) Cov(Zi , (γi + βi )MiI | Pi = 1)

β P ost = = .
Cov(Zi , MiI | Pi = 1) Cov(Zi , MiI | Pi = 1)

⊥ (Pi , Yi (0), γi , βi ) and MiI = MiJ × Pi . Let

The second equalities follow from the facts that Zi ⊥
MiJ (z) be individual i’s potential time in jail when Zi = z. Then, by independence of Zi :

Cov(Zi , γi MiJ | Pi = 0) = E[γi MiJ (1) | Zi = 1, Pi = 0] − E[γi MiJ (0) | Zi = 1, Pi = 0]

= E[(MiJ (1) − MiJ (0))γi | Pi = 0].

Similarly, with MiI (z) denoting individual i’s potential time in IGNITE when Zi = z:

Cov(Zi , (γi + βi )MiI | Pi = 1) = E[(γi + βi )MiI | Zi = 1, Pi = 1] − E[(γi + βi )MiI | Zi = 1, Pi = 1]

= E[(MiI (1) − MiI (0))(γi + βi ) | Pi = 1].

Moreover:

Cov(Zi , MiJ | Pi = 0) = E[MiJ (1) | Zi = 1, Pi = 0] − E[MiJ (0) | Zi = 1, Pi = 0]

= E[MiJ (1) − MiJ (0) | Pi = 0]

and

Cov(Zi , MiI | Pi = 1) = E[MiI (1) | Zi = 1, Pi = 1] − E[MiI (0) | Zi = 1, Pi = 1]

= E[MiI (1) − MiI (0) | Pi = 1].

MiJ (1)−MiJ (0)

Hence β P re = E[ωiP re γi | Pi = 0] and β P ost = E[ωiP ost (γi +βi ) | Pi = 1] for ωiP re = E[MiJ (1)−MiJ (0)|Pi =0]
MiI (1)−MiI (0)
and ωiP ost = E[MiI (1)−MiI (0)|Pi =1]
.

E.2 Difference-in-IVs With Nonlinear Causal Effects

Consider a general (partially linear) causal model, in place of Equation (1):

Yi = Gi (MiJ ) + Bi (MiI )

A.43
 ∂
where Gi (·) and Bi (·) are unconstrained potential outcome functions. Let γi (m) = ∂m Gi (m) and
∂
βi (m) = ∂m Bi (m) denote marginal effects of MiJ ≥ 0 and MiI ≥ 0, respectively, for individual i at
margin m. Assume Zi is independent of (Pi , Gi (·), Mi (·)) and that MiI = MiJ × Pi . Consider:

Cov(Zi , Yi | Pi = 0) = Cov(Zi , Gi (MiJ ) | Pi = 0)

!
Z MiJ
∂
= Cov(Zi , Gi (0) | Pi = 0) + Cov Zi , Gi (m)dm | Pi = 0
0 ∂m
Z ∞ 
J J


=E 1[Mi (1) ≥ m] − 1[Mi (0) ≥ m] γi (m)dm | Pi = 0 ,
0

where again MiJ (z) denotes individual i’s potential time in jail when Zi = z. The first equality
uses MiI = MiJ × Pi , the second equality applies the causal model, and the third equality uses
independence of Zi . By the same steps:

Cov(Zi , Yi | Pi = 1) = Cov(Zi , Gi (MiI ) + Bi (MiI ) | Pi = 1)

Z ∞ 
I I


=E 1[Mi (1) ≥ m] − 1[Mi (0) ≥ m] (γi (m) + βi (m))dm | Pi = 1 ,
0

where again MiI (z) denotes individual i’s potential time in IGNITE when Zi = z. Moreover:
Z ∞ 
Cov(Zi , MiJ | Pi = 0) = E 1[MiJ (1) ≥ m] − 1[MiJ (0) ≥ m] dm | Pi = 0


0

and
Z ∞ 
Cov(Zi , MiI 1[MiI (1) 1[MiI (0)


| Pi = 1) = E ≥ m] − ≥ m] dm | Pi = 1 .
0

Now consider the following generalization of Equation (4):

Z ∞  Z ∞ 
E ωiP re (m)γi (m) | Pi = 0 = E ωiP ost (m)γi (m) | Pi = 1
0 0

for

1[MiJ (1) ≥ m] − 1[MiJ (0) ≥ m]

ωiP re (m) = R∞
1[MiJ (1) ≥ m] − 1[MiJ (0) ≥ m] dm | Pi = 0]


E[ 0

and

1[MiI (1) ≥ m] − 1[MiI (0) ≥ m]

ωiP ost (m) = R∞ .
1[MiI (1) ≥ m] − 1[MiI (0) ≥ m] dm | Pi = 1]


E[ 0

A.44
Under this condition, the difference-in-IVs identifies:

Cov(Zi , Yi | Pi = 1) Cov(Zi , Yi | Pi = 0)
β∆ = −
Cov(Zi , Mi | Pi = 1) Cov(Zi , MiJ | Pi = 0)
I
Z ∞  Z ∞ 
P ost P re
=E ωi (m)(γi (m) + βi (m)) | Pi = 1 − E ωi (m)γi (m) | Pi = 0
0 0
Z ∞ 
P ost
=E ωi (m)βi (m) | Pi = 1 ,
0

generalizing Equation (5). Here β ∆ captures a weighted average of incremental IGNITE effects
βi (m) at different margins of exposure time m. The ωiP ost (m) weights are convex when court
delays weakly increase time in IGNITE, i.e., 1[MiI (1) ≥ m] − 1[MiI (0) ≥ m] ≥ 0.

E.3 Identification of ACR Weights and Complier Characteristics

As in Angrist and Imbens (1995), the difference-in-IVs estimand in Appendix E.2 can be written:
Z ∞
β∆ = ϕ(m)E[βi (m) | MiI (1) ≥ m > MiI (0), Pi = 1]dm,
0

where

P r(MiI (1) ≥ m > MiI (0) | Pi = 1)

ϕ(m) = R ∞ .
0 P r(MiI (1) ≥ m′ > MiI (0) | Pi = 1)dm′

The weight that β ∆ puts on E[βi (m) | MiI (1) ≥ m > MiI (0), Pi = 1], i.e. the complier-average
effect for margin m, is identified by a conditional IV regression of 1[MiI ≥ m] on MiI :

Cov(Zi , 1[MiI ≥ m] | Pi = 1) P r(MiI (1) ≥ m > MiI (0) | Pi = 1)

= ∞ = ϕ(m).
Cov(Zi , MiI | Pi = 1) I I
R
′ ′
0 P r(Mi (1) ≥ m > Mi (0) | Pi = 1)dm

Moreover, the average characteristics of compliers with the same weighting scheme are identified.
Letting Xi be an observed characteristic, with Xi ⊥
⊥ Zi :
∞
Cov(Zi , Xi MiI | Pi = 1)
Z
= ϕ(m)E[Xi | MiI (1) ≥ m > MiI (0), Pi = 1]dm,
Cov(Zi , MiI | Pi = 1) 0

following the same steps as above. The left-hand side of this expression comes from a conditional IV
regression of Xi MiI on MiI that instruments with Zi . Analogous versions of these two IV regressions
identify pre-IGNITE ACR weights and complier characteristics. In practice, we estimate ACR
weights and complier characteristics by versions of these IV regressions that include the design and
auxiliary controls from our baseline estimation procedure.

A.45
F Literature Comparison
We compare our baseline estimated effect of IGNITE on recidivism with the effects of other pro-
grams evaluated in the literature. To benchmark our findings, we calculate the effect sizes for
related work against the control complier mean, the complier mean, the control mean, or the mean
value of the recidivism measure, in that order of priority based on availability. When possible, we
also compute one-month treatment effects of these programs by assuming linear effects. We apply
the same transformations to the confidence intervals. Below we detail this calculation for each
paper included in the literature comparison plots in Figure 4.

F.1 Time in Programming

1. Mueller-Smith and Schnepel (2021) use a fuzzy regression discontinuity (RD) design to study
two two natural experiments in Harris County, TX, that shifted the probability of diversion
referral for first-time felony defendants. They find that diversion reduced having any convic-
tions 10 years post-disposition by 46.4% ([Table 3, Column 2] -0.26/0.56). The treatment is
not comparable so we use this effect size.

2. Heller et al. (2017) use an experiment to study the effects of a cognitive behavioral therapy
(CBT) Program for youth in a juvenile detention center in Cook County, Illinois. They find
that CBT participation reduced the probability of readmission in the 12 months after release
by 22.0% ([T8C6] -0.1689/0.768). Given the program lasted for approximately one month,
we use this effect size.

3. Arbour, Marchand and Lacroix (2023) study prison rehabilitation programs in Canada using
variation in program availability as an instrument for participation. They find that partici-
pation in one additional program reduces the probability of recidivism within one year after
release by 17.5% ([T5C1] -0.038/0.217). The treatment is not comparable so we use this effect
size.

4. Bhuller et al. (2020) study prisons in Norway. Leveraging random assignment of cases to
judges and the variation in judge stringency in an IV design, they find that incarceration
reduced the probability of being charged with at least one crime within two years of the case
decision by 42.7% ([T4C1] -0.239/0.56). Being incarcerated increases the number of prison
days served to 183.83 days or 6.1 months. We convert the estimate to a one month treatment
effect of -7.0%.

5. Shem-Tov, Raphael and Skog (2022) study randomized assignment to a restorative justice
intervention for youth facing felony charges in San Francisco, CA. The program replaced
traditional felony prosecution. They find that program participation reduces the likelihood
of being arrested in the 12 months after randomization by 40.3% ([T3C2] -0.228/0.566). The
intervention lasts approximately 6 months from enrollment to completion. We convert the
estimate to a one month treatment effect of -6.7%.

A.46
 6. Augustine et al. (2022) leverage the random assignment of felony cases to arraignment judges
and the variation in diversion referral rates as an instrument for diversion referral in San
Francisco County, CA. They find that being referred to a diversion program reduced new ar-
rests within one year post-arraignment by 29.8% ([T7C4 and T7C1] -0.150/0.503). Diversion
programming increases time to disposition by 288.5 days or 9.6 months [T6C4]. We convert
the estimate to a one month treatment effect of -3.1%.

7. Mastrobuoni and Terlizzese (2022) leverage prison overcrowding in an IV design to study

the use of open prisons in Italy. They find that one extra year in open prisons reduced
reincarceration within 3 years from the end of an individual’s custodial term by 26.6% ([T4C4
and T4-Notes] -0.105/0.395). Considering open prisons as a program, we convert the estimate
to a one month treatment effect of -2.2%.

8. Golestani, Owens and Raissian (2024) exploit random court room assignment of low-income
defendants in Nashville and Davidson County, TN in an IV design. They find that having a
domestic violence case heard in a specialized domestic violence division court increases the
probability of subsequently appearing in court for a new crime within three years of initial
case disposition by 1.7% ([T8C3] 0.015/0.865). The treatment is not comparable so we use
this effect size.

9. Lee (2023) estimates the effect of residential housing on reincarceration in Iowa using the
housing recommendation rate of randomly assigned case managers as an IV. He finds that
former prisoners assigned to housing after release had a 18.4% ([T4C6] 0.082/0.446) higher
probability of returning to prison within 3 years of release. Individuals typically stay in their
assigned housing facility for 4 months. We convert the estimate to a one month treatment
effect of 4.6%.

F.2 Time Incarcerated

1. Bushway and Owens (2013) use a law change in Maryland that altered recommended (but not
actual) sentences for a subset of offenders in jail or prison to study the impact of expectations
on future criminal behavior in an IV design. They find that holding actual punishment con-
stant, a 100 percent reduction in recommended sentences reduced the probability of rearrest
for men within 3 years after release by 14.0% ([T4C3] -0.090/0.645). The treatment is not
comparable so we use this effect size.

2. Roach and Schanzenbach (2015) use random judge assignment for defendants in Seattle, WA,
who plead guilty as an IV for time in prison. They find that extending the prison sentence by
an additional month reduces another sentencing for one year post-release by 13.9% ([T5C1
and T1C1] -0.0167/0.12).

3. Estelle and Phillips (2018) study the effect of additional time in a Michigan jail or prison
using both a judge IV and sentence guideline discontinuities. They find that an additional

A.47
 day spent incarcerated reduces the number of future felony convictions 5 years after the start
of initial sentence by 0.4% ([T10C1] -0.0055/1.43). We convert this estimate to a one month
treatment effect of -11.5%.

4. Humphries et al. (2023) study the effects of Virginia felony conviction and incarceration
using both a judge IV and sentencing guideline discontinuities. They find that one year after
sentencing, those right above the sentence guideline cutoff spend 8 more months incarcerated
and experience a reduction in recidivism by 40.8% ([T7C3] -0.049/0.12). We convert the
estimate to a one month treatment effect of -5.1%.

5. Zapryanova (2020) uses both a judge IV and a fuzzy RD design to study the effects of time in
prison and time on parole in Georgia. She finds that an additional month in prison decreases
the probability of returning to prison within three years of release while on parole by 4.5%
([T5C4 and T1C1] -0.0104/0.23). We use this effect size.

6. Lotti (2022) uses a fuzzy RD design around an offender’s 21st birthday to compare harsh
and rehabilitative criminal incarceration practices among youthful offenders in England and
Wales. She finds that young offenders at the margin of the age cutoff and who experience
custody in prison were 36.9% less likely to reoffend than those exposed to youth custody
centers over an 8-year time span ([T3C1 and T1C1] -0.265/0.719). The mean sentence length
for the 1963 cohort was 9.5 months [T1C1]. We convert the estimate to a one month treatment
effect of -3.9%.

7. Kuziemko (2013) studies discontinuities in Georgia’s parole guidelines using an RD design.

She finds that each month in prison reduces recidivism risk within 3 years after release by
3.8% ([T2C3] -0.0130/0.344).

8. Hjalmarsson and Lindquist (2022) study Swedish prison reforms which changed the share of
time spent in prison without shifting sentence length. They find that an increase in the share
of time individuals were required to serve reduced the probability of having any convictions
in the one year post-release by 2.7% ([T6C1] -0.015/0.563). The average prison sentence was
11.7 months [T1C1], and since the reform would hypothetically increase the share of time
served from 52% to 62%, we interpret the results as a one-month treatment effect.

9. Tobón (2022) studies the quasi-random assignment of individuals in the same judicial dis-
trict to newer and higher quality prisons in Columbia. He finds that being released from a
new prison reduced prison reentry in the twelve months after release by 35.7% ([T4C1-2] -
0.035/0.098). Given an average prison length of 15.11 months [T2C1, 453.32 days], we convert
the estimate to a one month treatment effect of -2.4%.

10. Rose and Shem-Tov (2021) use discontinuities in North Carolina’s sentencing guidelines to
study the effect of increased time in prison. They find that one month of prison exposure

A.48
 reduces the likelihood of reincarceration within 5 years by 2.3% ([T2C3] -0.0115/0.5). We use
this effect size.

11. Drago, Galbiati and Vertova (2009) study a collective clemency bill in Italy that exogenously
shifts remaining sentences at the time of pardon using a regression analysis conditioning on
initial sentence length. They find that an additional month in the residual sentence decreases
the probability of returning to prison 7 months after release by 1.4% ([T2C1 and T1C1]
-0.0016/0.115). We use this effect size.

12. Mueller-Smith (2015) uses random assignment of defendants to courtrooms in Harris County,
Texas in an IV design. He finds that for an additional year a felony defendant is incarcerated,
the probability of being rebooked in county jail for a new arrest increases by 12.2% ([T4C2
and T1C2] 0.067/(1 - 0.45)). The outcome measures recidivism per quarter and the treatment
margin is an additional year incarcerated. We convert the estimate to a one month treatment
effect of 1.0%.

13. Leslie and Pope (2017) instrument for pretrial detention status using variation in judge de-
tention rates across judges in New York City criminal courts. They find that being detained
leads to a 30.1% increase in being rearrested within two years for misdemeanor defendants
([T5C6 and T1C3-4] 0.118/[1-(0.66×639,141 + 0.24×89,614)/(639,141+89,614)]). The me-
dian length of detention was approximately 1 month for misdemeanor defendants, so we use
this effect size.

14. Dobbie, Goldin and Yang (2018) leverage the detention tendencies of quasi-randomly assigned
bail judges in Philadelphia County, PA and Miami-Dade County, FL in an IV design. They
find that pretrial release decreases the probability of rearrest two years following case dispo-
sition by 35.3% ([T4C6 and T4C1] -0.121/0.343). The treatment is not comparable so we use
this effect size.

15. Aizer and Doyle Jr (2015) study the effect of juvenile detention on adult incarceration, using
random assignment of judges in Chicago, IL. Using an IV design, they find that juvenile
incarceration increases the probability of adult imprisonment by age 25 by 71.6% ([T5C7
and T5C4] 0.234/0.327). The average incarceration length was 42 days, so we convert the
estimate to a one-month treatment effect of 51.1%.

F.3 Alternatives to Incarceration

1. Williams and Weatherburn (2022) study electronic monitoring as an alternative to prison for
nonviolent offenses in Sydney, Australia. Leveraging quasi-random assignment of cases to
judges in an IV design, they find that being sentenced to electronic monitoring reduces the
probability of committing any reoffense within 84 months of their case being finalized by 20.2
percent ([T5C3] -0.15/0.74). The treatment is not comparable so we use this effect size.

A.49
2. Blattman, Jamison and Sheridan (2017) explore the use of two randomly assigned interven-
tions in Monrovia, Liberia, which included an eight-week program of group CBT and a $200
grant. They find that participants who received the treatment saw a 28.0% reduction in
participants saying they have been arrested within the past two weeks in a survey conducted
12-13 months after receipt of grants ([T2C10 and T2C1] -0.033/0.118). The treatment period
was two months long. We convert the estimate to a one month treatment effect of -14.0%.

3. Di Tella and Schargrodsky (2013) use randomly assigned judges in an IV design to study the
effects of electronic monitoring in Buenos Aires, Argentina. They find that assignment to
electronic monitoring decreases the probability of returning to prison for a new crime after
supervision by 71.5% ([T5C3 and Section IV Paragraph 1] -0.16/0.2237). The mean post-
release period was 2.85 years and individuals spent on average 420 days or 14.0 months on
electronic monitoring. We convert the estimate to a one month treatment effect of -5.2%.

4. Bhatt et al. (2024) use an experiment of an 18-month long program in Chicago, IL that couples
short-term employment with cognitive behavioral therapy and other social support. They find
that participation in the program leads to a 64.7% reduction in shooting and homicide arrests
20 months post-randomization ([T4C4 and T4C3] -0.0220/0.0340). We convert the estimate
to a one month treatment effect of -3.6%.

5. Henneguelle, Monnery and Kensey (2016) use two IVs that exploit the staggered rollout
of electronic monitoring in French courts and the tendency of courts to utilize electronic
monitoring. They study the effects of electronic monitoring in place of serving prison time
and find that electronic monitoring decreases any re-conviction within 5 years of release
for prisoners by 8.6% ([T2C6 and TA1C3] -0.0571/0.662). The average initial sentence for
electronic monitoring is 5.4 months, so we convert the estimate to a one month treatment
effect of -1.6%.

6. Lavecchia, Oreopoulos and Spencer (2024) use a difference-in-differences framework to study

a high school support program for disadvantaged youth living in public housing projects in
Toronto, Canada. They find that being eligible for the program in one’s neighborhood reduces
ever being charged with a crime by 31.6% ([T1C1-2] -0.06/0.19). The program lasted for the
entirety of a child’s high school career (4 years). We convert the estimate to a one month
treatment effect of -0.66%.

7. Packham and Slusky (2023) explore the effects of Medicaid access in South Carolina using
an RD design and changes in access policies. They find that Medicaid enrollment within
six months of release reduces offenses committed within 1 year of release by 0.02% ([T3C1]
-0.00002/ 0.097). The treatment is not comparable so we use this effect size.

A.50