Skip navigation
The Habeas Citebook Ineffective Counsel - Header

Distortion of Justice How the Inability to Pay Bail Affects Case Outcomes, Stevenson, 2016

Download original document:
Brief thumbnail
This text is machine-read, and may contain errors. Check the original document to verify accuracy.
Distortion of Justice: How the Inability to Pay Bail
Affects Case Outcomes
Megan Stevenson∗†
May 2, 2016

Abstract
Instrumenting for detention status with the bail-setting propensities of rotating
magistrates I find that pretrial detention leads to a 13% increase in the likelihood of
being convicted, an effect explained by an increase in guilty pleas among defendants
who otherwise would have been acquitted or had their charges dropped. On average,
those detained will be liable for $128 more in court fees and will receive incarceration
sentences that are almost five months longer. Effects can be seen in both misdemeanor
and felony cases, across age and race, and appear particularly large for first or second
time arrestees. Case types where evidence tends to be weaker also show pronounced
effects: a 30% increase in pleading guilty and an additional 18 months in the incarceration sentence. While previous research has shown correlations between pretrial detention and unfavorable case outcomes, this paper is the first to use a quasi-experimental
research design to show that the relationship is causal.

∗

Fellow, University of Pennsylvania Law School, mstevens@law.upenn.edu
Many people helped considerably on this paper and for that I am grateful: Mira Baylson, Keir BradfordGrey, Paul Heaton, John Hollway, Mark Houldin, Charles Loeffler, Sandy Mayson, Chris Welsh, participants
at the Quattrone Lunch and the Penn Economics of Crime Brown Bag.
†

1

I have had the ‘you can wait it out or take the deal and get out’ conversation with way too many clients. -a public defender, Philadelphia

1

Introduction
The use of money bail has long been contentious, as it implies that freedom is re-

served for those who can afford to pay for it. According to Bureau of Justice Statistics,
five out of six people detained before trial are there because they could not afford their
bail (Cohen and Reaves, 2007). For some, bail was intentionally set at an unaffordable
level to keep them behind bars. But many have bail set at amounts that would be
affordable for the middle or upper-middle class but are simply beyond the reach of the
poor. In Philadelphia, the site of this study, only 51% of those who were assigned a bail
amount less than or equal to $500 were able to pay the $50 deposit required for release
within the first three days, and 25% remained in jail at the time of disposition. While
the loss of freedom during the pretrial period is a serious concern, there are indirect
consequences of pretrial detention which are potentially just as serious. If detention
increases the likelihood that the defendant will plead guilty although innocent, accept
an excessively-punitive plea deal, or lose at trial because detention impaired her ability
to mount a successful defense, then the socio-economic disparities induced by a money
bail system extend beyond the pretrial period.
Right now there is a wave of momentum in bail reform that dwarfs any seen in
decades. Media attention to the recent deaths of Kalief Browder and Sandra Bland
and budget pressures imposed by the 11.4 million yearly admissions to local jails have
spurred jurisdictions all over the country to reconsider long-standing practices.12 Yet
despite all this activity, research on the pretrial period is limited. The work that
does exist is almost entirely qualitative or correlational; the few experimental or quasiexperimental studies that have been conducted either did not examine case outcomes
or did not report the results in a way that allowed for causal inference (Ares et al.,
1963; Goldkamp, 1980; Abrams and Rohlfs, 2011).
In this paper I present the first quasi-experimental evidence that pretrial detention
increases the likelihood of being convicted, pleading guilty, being sentenced to incarceration, and being required to pay hundreds or thousands of dollars in court fees. In
Philadelphia, the bail decision is made by one of six Arraignment Court Magistrates
1

In the last several years, pretrial reform has been committed to or implemented in New Jersey, Kentucky,
Colorado, Maryland, New Mexico, Connecticut, Chicago, New York City. 26 cities are implementing new
pretrial risk assessment regimes in partnership with the Laura and John Arnold Foundation and 20 cities
are developing pretrial reform proposals with a $75 million fund from the MacArthur Foundation.
2
Many of the 11.4 million admissions will spend only short periods in jail. The point-in-time count of
jail inmates is 750,000, two thirds of whom are awaiting trial. BJS estimates that 95% of jail growth since
the year 2000 has come from pretrial detainees (Minton and Zeng, 2015).

2

who have broad discretion in setting bail. These magistrates work on a rotating schedule which ensures that over time, the sample of defendants seen by each magistrate
will be statistically near identical. By determining the bail amount, the magistrate
has influence over whether a defendant is detained pretrial, but is unlikely to affect
case outcomes through any other channel. All of the other activities that take place
in the bail hearing are pro-forma and the schedules of the magistrates do not correlate
with the schedules of the judges or attorneys that are influential later in the criminal
proceedings. Thus the bail preferences of the magistrate who presides over the bail
hearing provide an exogenous shock to the likelihood of being detained pretrial: a suitable instrument with which to identify the causal impact of pretrial detention on case
outcomes.
The data used in this analysis covers all criminal cases originated in Philadelphia
between September 2006 and February 2013. The sample size is 331,615 criminal cases
with eight total bail magistrates. I use a jackknife instrumental variables technique
which, in my preferred specification, allows the preferences of the magistrates to vary
over time and according to offense, criminal history, race, and gender of the defendant (Mueller-Smith, 2015). This allows me to exploit considerable within-magistrate
variance in detention rates. For example, a defendant with a shoplifting charge is 30
percentage points more likely to be detained if she is seen by the magistrate who is
most strict on shoplifting charges instead of the one who’s most lenient. However the
magistrate that is most lenient on shoplifting charges is most strict on robbery charges,
the magistrate that is most lenient on robbery is strict on drug possession, and so forth.
Using variation in the propensity to detain both within and across the magistrates, I
estimate that pretrial detention leads to a 6.6 percentage point increase in the likelihood
of being convicted on at least one charge, over a mean 50% conviction rate. The effect
on conviction (being found guilty either through plea or at trial) is almost entirely
explained by a 5.3 percentage point increase in the likelihood of pleading guilty among
those who would otherwise have been acquitted, diverted, or had their charges dropped.
Those detained will be liable for $128 more in non-bail court fees and will be sentenced
to an additional 140 days of incarceration.3
Effects appear to be present in both misdemeanors and felonies, although the effects
are much more precisely estimated in the former. Those detained on a misdemeanor
charge will be 8 percentage points more likely to be convicted and 8 percentage points
more likely to receive a sentence of incarceration. Effects are similar in magnitude for
white and black defendants, don’t vary substantially according to defendant’s age, but
appear largest for first or second time arrestees. For those with very limited experience
3

Average court fees for the convicted are $775 and those who receive a prison sentence will serve almost
two years on average before being eligible for parole.

3

in the criminal justice system, pretrial detention leads to a 17 percentage point increase
in the likelihood of being convicted.
I divide the sample according to the strength of the evidence that tends to be
available in different crime types. My hypothesis is that extra-legal factors such as
detention will be more influential among cases where the facts are in dispute than
among cases where the evidence is difficult to contest. I find that effects among ‘strongevidence’ crimes (DUI, drugs, illegal firearms) are generally small and not statistically
significant in the IV regressions, but effects among ‘weak-evidence’ crimes (assault,
vandalism, burglary) are more pronounced. Those detained on a weak-evidence charge
are 7 percentage points more likely to plead guilty and will receive an additional 18
months of incarceration. Since weak-evidence crimes tend to have higher rates of
wrongful convictions, these findings are consistent with the claim that pretrial detention
increases the likelihood that innocent people will plead guilty.4
Previous studies using multivariate regression techniques usually show that pretrial
detention is correlated with unfavorable case outcomes even after controlling for a number of defendant and case characteristics (Ares et al., 1963; Rankin, 1964; Goldkamp,
1980; Williams, 2003; Free, 2004; Phillips, 2007, 2008; Tartaro and Sedelmaier, 2009;
Sacks and Ackerman, 2012; Oleson et al., 2014; Heaton et al., 2016). Many have found
very large correlations. An often-cited study by the Arnold Foundation concludes that
low risk defendants who are detained throughout the pretrial period are 5.41 times more
likely to be convicted and sentenced to jail than those who are released (Lowenkamp
et al., 2013). They control for offense within eight broad categories as well as basic
demographics and criminal history measures, but variations in charge severity, strength
of the evidence, quality of the representation, wealth of the defendant, and a variety
of other unobservable differences may have biased the results.
The IV estimates I present here are considerably lower than most previous estimates. This may be due to omitted variable bias in previous research, differences in
effect sizes across jurisdictions, or the fact that the IV estimates capture the effect of
pretrial detention only for those on the margins of being detained. I also conduct OLS
regressions of case outcomes on pretrial detention, controlling for narrowly defined offense categories as well as many other defendant and case characteristics. The OLS
results do not differ substantially from the IV results, suggesting that researchers who
are able to control for narrowly defined offense categories as well as a variety of criminal history and demographic variables may be able to produce reasonable estimates
of the effect of pretrial detention on case outcomes even in the absence of a natural

4

A forthcoming study by Charles Loeffler and Jordan Hyatt examines wrongful convictions through the
use of anonymous surveys of prison inmates. Their preliminary findings suggest that the rate of wrongful
convictions is higher among assaultive crimes and lower among DUIs and some drug crimes.

4

experiment.5
The impact that pretrial detention has on case outcomes could come through a
variety of channels. A common claim is that people plead guilty just to get out of jail
(Bibas, 2004). While this is a reasonable mechanism for misdemeanor cases, effects are
also seen in much more serious charges, where it is unlikely that a defendant would
be able to avoid a prison sentence. It may be that since some of the disruptions of
incarceration have already occurred – loss of job/housing, the initial adjustment to life
behind bars – the incentives to fight the charges are lower. Jail may affect optimism
about the likelihood of winning the case, or may affect risk preferences in such a way
that the certainty of a plea deal seems preferable to the gamble of a trial.6 Detention
also impairs the ability to gather exculpatory evidence, makes confidential communication with attorneys more difficult, and limits opportunities to impress the judge with
gestures of remorse or improvement (taking an anger management course, entering rehab, etc.) (Goldkamp, 1980). Detained defendants are likely to attend pretrial court
appearances in handcuffs and/or prison garb, creating superficial impressions of criminality. Furthermore, if a defendant must await trial behind bars he may be reluctant
to employ legal strategies that involve delay. Whereas a released defendant may file
continuances in the hopes that the prosecution’s witnesses will fail to appear, memories
will blur, or charges eventually get dropped, a detained defendant pays a much steeper
price for such a strategy. More nefariously, those detained have less opportunity to
coerce witnesses, destroy evidence or otherwise impede the investigation (Allen and
Laudan, 2008). While pretrial detention may increase wrongful conviction, pretrial
release may decrease the likelihood of successfully prosecuting the truly guilty.
While this paper does not argue for or against the use of pretrial detention, it
shows that detention has significant downstream consequences. Being found guilty in
the court of law comes with hefty court fees, periods of incarceration or probation,
potential disenfranchisement, a criminal record, and challenges securing future employment, school loans, or housing (Uggen et al., 2012; SHRM, 2012; Thacher, 2008;
ONDC, n.d.). If wealth or race influence the likelihood of being detained pretrial – and
both previous research as well as evidence presented in this paper suggest that they do
– then pretrial detention exacerbates socio-economic inequalities in the criminal justice
system (Schlesinger, 2005; Wooldredge, 2012).
In Section 2 I give a brief overview of the pretrial process, in Section 3 I describe
5

In forthcoming work with Paul Heaton and Sandra Mayson (Heaton et al., 2016) we use OLS methods
with extensive controls to estimate the impact pretrial detention has on case outcomes in Harris County,
Texas. In general, results are similar to those found in Philadelphia; effect sizes are slightly larger but this
may be explained by institutional differences across the different jurisdictions.
6
Detention may change the reference point so that freedom is seen as a gain instead of incarceration being
seen as a loss. Prospect theory would thus predict a shift towards more risk averse decisions (Daniel Kahneman, 1979).

5

the natural experiment, and in Section 4 I discuss the data and provide descriptive
statistics and graphs. Section 5 tests the identifying assumption and Section 6 discusses
the empirical strategy and shows a visual representation of the first and second stage.
Section 7 presents the results for the full sample and Section 8 shows results for various
subgroups. Section 9 provides robustness tests and shows how results vary with the
number of days detained. Section 10 concludes.

2

Bail, pretrial detention, and an overview of

criminal proceedings
Pretrial detention is the act of keeping a defendant confined during the period between arrest and disposition for the purposes of ensuring their appearance in court
and/or preventing them from committing another crime. The vast majority of jurisdictions use a money bail system to govern whether or not a defendant is detained. In
such a system a judge or a magistrate determines the amount of the bail required for
release and the defendant is only released if she pays that amount and agrees to certain
behavioral conditions. In some cases the defendant will be released without having to
pay anything, in others (usually only the most serious cases) she will be denied bail
and must remain detained. While the defendant is liable for the full amount of the
bail bond if she fails to appear at court or commits another crime during the pretrial
period, she usually does not need to pay the full amount in order to secure release. In
many jurisdictions she will borrow this sum from a bail bondsman, who charges a fee
and holds cash or valuables as collateral. In some jurisdictions, Philadelphia included,
the courts act as a bail bondsman and will release the defendant after the payment of
a deposit.
Bail hearings are generally quite brief – in Philadelphia most last only a minute
or two – and often do not have any lawyers present. After the bail hearing there are
a series of pretrial court appearances that defendants must attend. Although the exact procedure varies across jurisdictions these usually include at least an arraignment
(where formal charges are filed) and some sort of preliminary hearing or pretrial conference (where the case is discussed and plea deals can be negotiated). Plea bargaining
usually begins around the time of arraignment and can continue throughout the criminal proceedings. In some jurisdictions, like New York City, the arraignment happens
simultaneous to the bail hearing and it is not uncommon to strike a plea deal at this
first appearance. In other jurisdictions, such as New Orleans, arraignments often do
not happen until six months after the bail hearing and a defendant who is unable to
make bail must wait until then to file a plea. In Philadelphia, arraignments usually
happen within a month of the bail hearing.

6

Plea negotiation is a process in which the defendant receives reduced charges or
shorter sentences in return for pleading guilty and waiving her right to a trial. Since
defendants often face severe sentences if found guilty at trial, the incentives to plead
are strong. It’s estimated that 90-95% of felony convictions are reached through a plea
deal (Devers, 2011). Philadelphia differs from many other jurisdictions in its wide use
of bench trials on felony cases. Since sentencing tends to be more lenient in bench
trials than jury trials, this reduces the incentive to plead guilty.7 As such, only about
75% of felony convictions are reached through plea in Philadelphia. Trial by jury is not
constitutionally required if the maximum incarceration sentence is less than six months,
and the use of bench trials for misdemeanors, as is the custom in Philadelphia, is more
common across jurisdictions.
This paper shows that pretrial detention increases the likelihood that a defendant
will plead guilty and will plead to less favorable terms. While there is little reason to
believe this result is unique to Philadelphia, the magnitude of the effect is likely to
differ across jurisdictions due to variations in the criminal justice proceedings.

3

The natural experiment
Immediately after arrest, arrestees are brought to one of seven police stations around

the city. There, the arrestee will be interviewed via videoconference by Pretrial Services. Pretrial Services collects information about various risk factors as well as financial information to determine eligibility for public defense. Using risk factors and the
current charge, Pretrial Services will determine the arrestee’s place in a 4 by 10 grid of
bail recommendations. These bail guidelines suggest a range of appropriate bail, but
are only followed about 50% of the time (Shubik-Richards and Stemen, 2010). Once
Pretrial Services has entered the bail recommendation and the financial information
into the arrest report the arrestee is ready for her bail hearing.
Once every four hours the magistrate will hold bail hearings for all arrestees on
the ‘list’ (those who are ready to be seen). The bail hearing will be conducted over
videoconference by the magistrate, with a representative from the district attorney’s
office, a representative from the Philadelphia Defender Association (the local public
defender), and a clerk also present. In general, none are attorneys. The magistrate
makes the bail determination on the basis of information in the arrest report, the
pretrial interview, criminal history, bail guidelines, and advocacy from the district
attorney and public defender representatives.
7

In Philadelphia, a bench trial is the default for all but the most serious felonies. The right to a jury
trial can be asserted upon request, but this is uncommon. While there is no formal mechanism that ensures
that a bench trial will lead to better outcomes for the defendant than a jury trial, all defense attorneys
interviewed assured me that this was the case.

7

There are four things that happen during the bail hearing: the magistrate will
read the charges to the arrestee, inform her of her next court appearance, determine
whether the arrestee will be granted a court-appointed defense attorney, and set the
bail amount. The first two activities are formalities that ensure the defendant is aware
of what she is being charged with and where her next court date is. Eligibility for
public defense is determined by income. If the defendant is deemed eligible, she will
be assigned either to the Defender Association, or to a private attorney who has been
approved to accept court appointments by the City of Philadelphia. The default is to
appoint the Defender Association; if procedural rules require the court to appoint an
attorney outside of the Defender Association the magistrate’s clerk will appoint the
attorney at the top of a rotating list of eligible attorneys known as a ‘wheel’.8
A typical bail hearing lasts only a minute or two and the magistrate has broad
authority to set bail as she sees fit.9 Bail decisions fall into three categories: release
with no payment required, cash bail or no bail.10 Those with cash bail will be required
to pay a 10% deposit on the total bail amount in order to be released. After disposition,
and assuming that the behavioral conditions of the pretrial period were met, 70% of
this deposit will be returned. The City of Philadelphia retains 30% of the deposit,
even if charges get dropped or the defendant is acquitted on all charges. Those who
do not have the 10% deposit in cash can borrow this amount from a commercial bail
bondsman, who will accept cars, houses, jewelry and other forms of collateral for their
loan. If the defendant’s arrest occurred while she is already on parole, her parole officer
may choose to file a detainer. If a detainer is filed she may not bail out until a judge
removes the detainer.11
The research design uses variation in the propensity of the magistrates to assign
affordable bail as an instrument for detention status. The validity of the instrument
rests on several factors, including that the magistrate received is essentially random
and that the instrument not affect outcomes through a channel other than pretrial
detention. The following details help ease concerns along these lines.
8

If there are multiple codefendants, such that representing all of them would pose a conflict of interest,
one defendant will be randomly selected to be served by the Defender Association and the others will receive
a court-appointed attorney. For opaque historical reasons, four out of five defendants charged with murder
will be represented by court-appointed attorneys and the fifth will be represented by the homicide division
of the Defender Association (Anderson and Heaton, 2012). This decision is made by the order in which
defendants are entered into the data system and the court-appointed attorney is chosen by a Municipal
Court judge, not a magistrate.
9
If either the defense or the prosecution is unhappy with the decision they can make an appeal to a judge
immediately after the bail hearing. However the bar is high for overturning the original bail decision so this
is not very common.
10
Holding a defendant without bail is uncommon, although bail is sometimes set at prohibitively high
rates.
11
The detainer hearing usually happens within a week of arrest. Detainer cases are evenly distributed
across magistrates and should not bias the results.

8

Philadelphia employs six Arraignment Court Magistrates at a time, and one of the
six will be on duty 24 hours a day, 7 days a week, including holidays. Each day is
composed of three work shifts: graveyard (11:30 pm-7:30 am), morning (7:30 am-3:30
pm) and evening (3:30 pm-11:30 pm). Each magistrate will work for five days on a
particular shift, take five days off, then do five days on the next shift, five days off, and
so forth. For example, a magistrate may work the graveyard shift from January 1st
to January 5th, have January 6th-10th off, then work the morning shift from January
11th-15th, have the 16th-20th off, do the evening shift from January 21st-25th, take
the next five days off, and then start the cycle all over again.
This rotation relieves concerns that certain magistrates set higher bail because
they work during shifts which see higher-risk defendants. Over time, each magistrate
will be scheduled to work a balanced number of weekends, graveyard shifts, and so
forth. However the magistrates do not always work their appointed shifts; in fact,
about 20% of the time there is a substitute (usually one of the other magistrates).
To avoid potential confounds I instrument with the magistrate who was scheduled to
work instead of the magistrate that actually worked. Furthermore, arrestees do not
have latitude to strategically postpone their bail hearing to receive a more lenient
magistrate. The process from arrest to bail hearing has been described as a conveyor
belt: on average the time from arrest to the bail hearing is 17 hours and defendants
are seen as soon as Pretrial Services notifies the Arraignment Court that they are
ready (Clark et al., 2011). Thus the cases seen by each scheduled magistrate should
be statistically very close to identical. I confirm this empirically in Section 6.
Since the duties of the bail magistrate are so limited, there are few channels outside
of the setting of bail through which the magistrate could affect outcomes. One concern
would be a correlation between the schedules of the magistrates and the likelihood of
receiving a particular judge, prosecutor or defense attorney later on in the criminal
proceedings. However the peculiar schedule of the magistrates does not align with the
schedule of any other actors in the criminal justice system. For one, this is because the
other courts are not open on weekends. This is also because Philadelphia predominantly
operates on a horizontal system, meaning that a different prosecutor handles each
different stage of the criminal proceedings. Likewise, if the defendant is represented
by the Defender Association (∼60% of the sample), she will have a different defense
attorney at each stage.12 While attorneys often rotate duties, their rotations are based
on a Monday-Friday work week and not the ‘five days on, five days off’ schedule of the
magistrates.
While eligibility for public defense is another potential channel through which the
12

The most serious cases are not handled horizontally, however the choice of attorney to handle these
cases has nothing to do with the magistrate.

9

magistrate could affect outcomes, this is supposed to be a pro-forma action based on
the income of the defendant. I am not able to see whether the defendant is deemed
eligible, I only know the attorney type at the time of disposition. This is likely to be
affected by detention status, as the decision to hire a private attorney is based both on
having the money to pay one and expectations over case outcomes. While I cannot test
directly whether magistrates vary in the rates at which they grant eligibility, I conduct
robustness checks in which I control for the attorney type at the time of disposition.
This has only a trivial impact on the results.

4

Data and descriptive statistics
The data for this analysis comes from the court records of the Pennsylvania Unified

Judicial System; PDF files of case dockets and criminal histories are publicly available
online. The data covers all criminal cases which had a bail hearing between September
13, 2006 and February 18, 2013.13 Before September 13, 2006, Philadelphia used a
different data management system with much lower data quality. I do not look at
cases which began after February 18, 2013 both because I wanted to leave ample time
for all cases to resolve and because one of the magistrates was replaced by a new one
on that date.
Figure 1a shows a histogram of the number of days defendants are detained before
disposition, conditional on being detained more than three days and less than 600
days.14 The left tail of the distribution is omitted since the primary definition of
‘detainees’ used in this paper is being unable to make bail within three days; the long
right hand tail of the distribution is omitted for visual simplicity. The median number
of days detained for those who are unable to make bail within three days is 78, the
mean is 146.
Summary statistics for the released group, the detained group, and the whole sample
are shown in Table 1. Defendants are predominantly male, and, although race is missing
for 11% of the sample, largely African-American. Those detained tend to have longer
criminal histories and are facing more serious charges than those released. It should
be noted, however, that 25% of the detained sample are only facing misdemeanor
charges.15
Almost half the sample have their charges dropped, dismissed, or are placed in
13

Case outcome or bail data was missing for 0.44% of the sample. These observations were dropped.
Although the sixth amendment guarantees the right to a speedy trial this can be waived if the defendant
finds it beneficial to extend.
15
The offense information used in this paper is taken from the charge at the time of the bail hearing.
Many of those who were originally charged with felonies subsequently had the felony charge downgraded to
a misdemeanor.
14

10

some sort of diversion program.16 Almost everyone else was convicted, through plea
or at trial, on at least one charge. 90% of cases resolved at trial result in convictions,
suggesting that prosecutors will not bring a case to trial if they don’t believe they have
a strong chance of winning. If a detained defendant pleads quickly to avoid more time
waiting in jail, she may be pleading guilty on a case that otherwise would not have
proceeded to court.
One third of the sample is released without being required to pay bail and an
additional 26% are able to pay their way out within three days of the bail hearing.
Figure 1b shows the number of people detained and released at a variety of bail ranges.
The median amount of bail for the detained group is $10,000. Almost 40% of those who
were given bail amounts less than or equal to $2000 are detained for longer than three
days. Among this low-bail sample – 77% of whom are charged only with misdemeanors
– the average number of days detained pretrial is 28. This group would need to pay a
deposit $200 or less to secure their freedom.
Table 2 shows the most common lead charges. The table is organized in descending
order, where the top of the list represents crime types where evidence tends to be
strongest. The horizontal line separates those classified as ‘strong-evidence’ crimes
from those classified as ‘weak-evidence’ crimes. The ranking of charges by strength of
evidence was done in two ways. First, I surveyed a variety of Philadelphia lawyers and
professors who specialize in criminal law. I asked them to rank each crime on a scale
of one to five, where a one meant that the evidence in that crime type tended to be
ambiguous and subject to multiple interpretations and a five meant that the evidence
in that crime type tended to be clear and difficult to dispute. Second, I calculated the
average conviction rate per crime type under the assumption that the conviction rates
would be higher among crime types where the evidence was stronger. I ranked each
crime according to both measures of evidence strength; the order in which they are
presented here is the average of the two rankings. While there were differences across
the two methods in the exact ordering, the general placement was quite similar.17
Figures 1c shows the likelihood of pleading guilty at various levels of ‘sentence
exposure’. Sentence exposure is a measure of how serious the case is: the average
incarceration sentence that similarly situated defendants receive if they are found guilty
at trial. This is estimated by taking the fitted values from a regression of sentence
length on offense, criminal history, demographics, and time controls, using the sample
of cases in which the defendant was found guilty at trial. A log transformation is
16

Diversion programs are designed for those with low level misdemeanor charges; if the defendant agrees
to requirements such as paying restitution to victims, entering rehab, or performing community service, they
are generally able to avoid a formal adjudication of guilt and a criminal record.
17
The ranking based on the survey would have placed car theft in the ‘strong-evidence’ category and other
types of theft in the ‘weak-evidence’ category.

11

applied to the fitted values to compress the long right tail. Figure 1c shows that while
the likelihood of pleading guilty is relatively flat at low levels of sentence exposure, it
rises rapidly and then falls off at higher levels of sentence exposure. While no definitive
conclusions should be inferred from this descriptive graph, it is consistent with a story
in which defendants plead guilty out of fear of a sentence exposure at trial. While the
plea rate drops off at the highest levels of sentence exposure, this may be because the
guilty pleas themselves come with increasingly long prison sentences. A present-biased
defendant may underweight the difference between the 10 year sentence she is offered
via plea or the 25 year sentence she is at risk of receiving at trial.
Figure 1d shows conviction rates at various levels of sentence exposure. It shows
that throughout most of the support of the distribution, the likelihood of being convicted decreases with sentence exposure. Figure 1e shows the likelihood of detention
at various levels of sentence exposure. We see a monotonically increasing relationship
between sentence exposure and the likelihood of being detained. This is not surprising,
as those who face longer sentences will face increased incentive to flee, or may pose
greater public safety risks to the community.18
In Table 3 I test to see if there are socio-economic disparities in the likelihood of
being detained pretrial. I regress the log bail amount and a dummy for pretrial detention on race and the log of average income in the defendant’s zip code, controlling for
offense, criminal history, age and gender.19 I limit the sample to those for whom both
zip code and race is available.20 I find that those coming from wealthier neighborhoods
have bail set lower and are less likely to be detained pretrial even after controlling for
a wide range of characteristics. A 10% increase in zip code wealth is correlated with
a .4% decrease in bail and a .2 percentage point decrease in the likelihood of being
detained pretrial. Race does not correlate with the bail amount in a statistically significant manner, however African-Americans are three percentage points more likely to be
detained pretrial than Caucasian defendants with similar offense profile, age, gender,
and criminal history.

18

These three graphs, taken together, suggest that OLS regressions of case outcomes on detention status
may in fact be biased downwards. Since the seriousness of the case is not fully visible in the data (offense
can be controlled for, but severity can vary within charge categories) this omitted variable will increase
the likelihood of detention but may actually decrease the likelihood of conviction. Other likely omitted
variables, such as wealth, lawyer quality, and community ties are more consistent with an upward bias of
OLS estimates.
19
I add one to all bail amounts to avoid taking the log of zero.
20
Average gross income per zip code in 2010 was downloaded from IRS.gov.

12

5

Identification test
Given the rotation of the work schedules, there may be slightly more imbalance

across the magistrates than there would be if the magistrate was randomly and independently drawn for each bail hearing. Since most parametric tests of randomness
assume an independent random draw, I conduct a permutation test to verify that the
sample of defendants seen by each magistrate are no more different from one another
than would be expected by chance, given a rotating work schedule such as the one in
use.
I generate a variety of “false” work schedules, keeping the basic parameters of the
work rotation fixed: five days on, five days off, rotating shifts, three eight hour shifts
per day, etc. However I vary both the day at which the five day rotation begins, the
hour at which the shifts begin, and the direction of rotation. (The actual magistrates
move from graveyard to morning to evening shift, a reverse rotation would move from
graveyard to evening to morning.) Since the schedule is five days long there are four
potential “false” start dates to choose from. Since the shift is eight hours long there
are three potential “false” start times, spaced two hours apart. I generate three false
rotations: one in which both groups work a reverse rotation, and two in which one
of the groups work a forward rotation while the other group works a reverse rotation.
Given five start dates (four false and one real), four start times (three false and one real)
and the three false rotations (I do not use the real rotation to minimize correlations
with the actual work schedule) I build 60 false work schedules.
Covi = α + β ∗ M agistratei + ψ ∗ T imei + ei

(1)

With each of these work schedules I run the regression specified in Equation 1,
where Covi is one of 66 different offense, criminal history and demographic covariates,
M agistratei is the dummy for the magistrate who was scheduled during the bail hearing
for case i under the false schedule, and T imei are the full set of controls for the time and
date of the bail hearing as described in Section 4.21 For each covariate and each false
21

Unless otherwise specified, all regressions shown in this paper control for the variables listed as follows.
Demographic variables include age, age squared, age cubed, gender, and dummies for being black or white.
Criminal history variables include the number of prior cases, prior felony cases, prior cases involving a serious
violent crime (robbery, assault, burglary, murder, rape), prior cases where the defendant was found guilty of
at least one charge, and dummies for having at least one prior case, having at least three prior cases, awaiting
trial on another charge, having a detainer placed on them, and having a prior arrest within five years of the
bail hearing. Offense variables include dummies for having a charge in the following category: murder, rape,
robbery, aggravated assault, burglary, theft, shoplifting, simple assault, drug possession, drug selling, drug
buying, possession of marijuana, F2 firearm, F3 firearm, possession of stolen property, vandalism, a nonfirearm weapon charge, prostitution, first offense DUI, second offense DUI, resisting arrest, stalking, motor
vehicle theft, indecent assault, arson, solicitation of prostitutes, disorderly conduct, pedophilia, intimidation
of witnesses, accident due to negligence, false reports to a police officer, fleeing an officer, and reckless

13

work schedule I collect the F statistic for the joint magistrate dummies. This creates
an empirical distribution of the F statistic under false schedules. I then compare the
F statistic from the Equation 1 regression in which the magistrate dummies are those
who were actually scheduled to the distribution of F statistics under false schedules.
For each covariate I build an ‘empirical p value’ which is the fraction of false-schedule
F statistics that are larger than the true-schedule F statistic.
I also perform this technique on three summary variables which are the fitted values
from a regression of a dummy for pretrial detention, a dummy for pleading guilty to
at least one charge, and a dummy for being convicted on at least one charge on all
66 of the offense, criminal history, demographic and time controls. These predicted
values are a weighted average of the characteristics that correlate most strongly with
the main dependent variables and the endogenous independent variable. I show the F
statistic and the empirical p value for these three summary statistics in Table 4. The F
statistics are 1.84, 2.59 and 1.91 respectively; as expected the covariates are somewhat
less balanced across the magistrates than they would be if the bail magistrate was
independently and randomly assigned to each bail hearing. However, at 0.96, 0.56
and 0.58 respectively, the empirical p values show that any imbalance in the predicted
likelihood of detention, guilt, or guilty pleas across the magistrates is no more than
would be expected by chance.
Figure 1f shows a histogram of the 69 empirical p values from the 66 covariates
and three summary statistics. As can be seen, the empirical p values are evenly spread
between zero and one. The mean p value is 0.56, the median is 0.58, and three p values
are less than or equal to 0.05. The mean F statistic is 2.21 and the median F statistic
is 1.84.
The ‘false’ F statistics comprising the empirical distribution will be correlated, since
there is overlap in the false schedules. This reduces the power on any single test (any
endangerment. Additional offense controls include dummies for being charged with a first, second or third
degree felony, an unclassified felony, a first, second or third degree misdemeanor, an unclassified misdemeanor,
or a summary offense. I also control for the total number of charges, the total number of felony charges,
the total number of misdemeanor charges, and the total ‘offense gravity score’ of the charges (the offense
gravity score is used by Philadelphia to measure the seriousness of a charge on a scale of 1-8). Time controls
include dummies for the day of the week that the bail hearing occurs, a dummy for graveyard, morning, and
evening shift, a cubic in day of the year (1-365), the bail date, and year dummies. Since magistrates tend
to leave and be replaced in the third week of February, I define the year dummies to align with the start
and end of a magistrate’s work period. The following dates serve as dividers: February 21, 2008, February
23, 2009, February 23, 2010, February 23, 2011, February 23, 2012, and February 18, 2013. I interact some
of the covariates with three time periods, as divided by February 23, 2009 and February 23, 2011. The
covariates that I allow to have differing impacts over these three time periods are the same as the ones that
I interact with the magistrate fixed effects in various specifications: murder, robbery, aggravated assault,
burglary, theft, shoplifting, simple assault, drug possession, drug sale, drug purchase, marijuana possession,
F2 firearm, F3 firearm, vandalism, prostitution, first offense DUI, motor vehicle theft, gender, a dummy for
being African-American, the number of prior cases, the number of prior violent crimes, a dummy for having
at least one prior and a dummy for having a detainer.

14

single covariate) since the 60 different false F statistics are not independent from one
another. However among the 69 different tests there is no evidence to suggest strategic
behavior that would undermine the credibility of the research design.

6

Empirical strategy and visual representations

of the research design
Instrumenting for sentencing outcomes using varying propensities of randomly assigned or rotating judges is a popular method of identifying the impact that the criminal justice system has on defendants (Kling, 2006; Aizer and Doyle, 2009; Loeffler,
2013; Tella and Schargrodsky, 2013; Mueller-Smith, 2015). Two concerns with this
methodology are that the judges may affect outcomes through channels other than the
primary sentencing characteristics the researcher is interested in, and that the monotonicity assumption may be violated. Concern about alternate channels of influence
are minimized in this setting due to the limited nature of the bail magistrates’ responsibilities and the fact that their work schedule does not align with the other actors in
the criminal justice system. However the assumption that the bail-setting propensities
for each magistrate do not vary according to defendant characteristics would be harder
to defend.
I follow Mueller-Smith (2015) in allowing the magistrates’ bail habits to vary across
time and according to defendant and case characteristics. Allowing the preferences of
the magistrates to update every two years relaxes the assumption that all magistrates
respond in the same way to changes in the criminal justice system: a new district
attorney, an update to the bail guidelines, a change in capacity constraints at the local
jail. Allowing the preferences of the magistrates to vary across case and defendant
characteristics relaxes the assumption that a strict magistrate must be equally strict
on all crimes and all defendants. In addition to minimizing non-monotonicity bias,
exploiting within-magistrate variation in bail-setting propensities increases power.
Empirically, allowing the magistrates’ bail preferences to vary is accomplished by
interacting the magistrate fixed effects with time period fixed effects and a subset of
the covariates in the first stage regression, as shown in Equation 2. Detentioni is a
dummy which is equal to one if the defendant is detained for more than three days in
case i, M agistratei is a dummy for the magistrate who was scheduled to work during
the bail hearing for case i, and Covi1 are the subset of the covariates across which I
allow the magistrate’s preferences to vary: the 16 most common crime types, gender,
race, and criminal history. Covi2 are the remainder of the offense, demographic and
criminal history controls, as listed in Footnote 21. Ti3 is a dummy for the time period
of the bail hearing for case i (there are three time periods as divided by February 23,

15

2009 and February 23, 2011) and T imei are the full set of controls for the time and
date of the bail hearing (which include Ti3 as a subset). I use a jackknife instrumental
variables technique to avoid the bias induced by many instruments (Angrist et al.,
1999).
Detentioni = α2 + M agistratei ∗ Ti3 ∗ ω2 + M agistratei ∗ Covi1
∗ φ2 + Covi1 ∗ Ti3 ∗ δ2 + Covi2 ∗ γ2 + T imei ∗ ψ2 + ei

(2)

The second stage is shown in Equation 3 where Case Outcomei represents a variety
of case outcomes, Detentioni is the fitted value from the jackknifed first stage, and
Covi1 , Covi2 , Ti3 and T imei are as described above. I present results from a linear
regression in both stages; logit regressions yield similar results.

Case Outcomei = α3 + Detentioni ∗β3 +Covi1 ∗Ti3 ∗δ3 +Covi2 ∗γ3 +T imei ∗ψ3 +

i

(3)

Figures 2 and 3 show visual representations of the first and second stage of the
IV regression. Figure 2 shows how pretrial detention rates vary by magistrate for
different crime types. The y axes show residuals from a regression of pretrial detention
on controls for the time and date of the bail hearing. The eight bars represent the
average detention residuals for the eight magistrates, ordered so that Magistrate 1 has
the lowest overall rate of pretrial detention, Magistrate 2 has the second lowest, and
so forth. The order of the magistrates remains the same in all subplots. Figure 2a
shows the magistrate means across the entire sample, Figures 2b-f show the means
for different crime types. In each offense category, n refers to the number of cases
charged with that offense and F refers to the joint significance of the magistrates in a
regression of detention status on magistrate dummies and time controls on that subsample. The figures demonstrate that while magistrates vary in their overall detention
rates, there is quite a bit of variance across the different crime types. For example,
although the overall detention rate of Magistrates 2 is quite low, this magistrate has the
highest detention rates for prostitution and drug possession. This within-magistrate
variation in detention rates suggest that the monotonicity assumption will be violated if
detention status is instrumented for with the overall detention rate of the magistrates.
Figure 3 shows that defendants whose bail hearing is presided over by high-bailsetting magistrates are more likely to be convicted or to plead guilty. In Figure 3a the
y and x axes show residuals from a regression of conviction and detention dummies
respectively on the set of time controls described by T ime. Figure 3b is similar except
the dummies are residualized over Cov 1 ∗ T 3 , Cov 2 and T ime. Each circle represents
the average detention and conviction residuals of the eight magistrates; the size of
the circle is proportional to the total number of cases seen by each magistrate. As

16

can be seen there is a clear positive correlation between the conviction and detention
residuals which changes very little once the effect of covariates have been removed. This
provides visual evidence that the slight differences in the groups of defendants seen by
each magistrate are not the cause of the positive relationship between detention and
conviction.
Figures 3c and d show the relationship between pretrial detention and pleading
guilty for those who are charged with weak-evidence crimes and those charged with
strong-evidence crimes respectively. In this graphic, the magistrate dummies have
been interacted with T 3 , dummies for the three time periods. The x and y axes in
both figures show residuals from a regression of pretrial detention and guilty pleas
respectively on covariates, time controls, and time-covariate interactions. The circles
show the average residuals per-magistrate-per-time period. A clear positive relationship
can be seen between the likelihood of being detained and the likelihood of pleading
guilty in weak-evidence cases. There is no visually discernible relationship among
strong-evidence cases.

7

Full sample results
In Table 5 I show results from a variety of jackknife instrumental variables spec-

ifications where the outcome variable is a dummy which equals one if the defendant
is convicted on at least one charge in Panel A, and a dummy variable which equals
one if the defendant pled guilty to at least one charge in Panel B. In Column 1 the
only instruments are the eight magistrate dummies and in Column 2 the magistrate
dummies are interacted with T 3 , the three time period dummies. The only controls
in the first two columns are T ime. The standard errors decrease between the first
and second column by about 10%, suggesting that allowing the magistrates to respond
differently to the various changes that occur during the period of my analysis increases
the power of the instrument. Covariates are added in Column 3 and the effect sizes
either increase (Panel A) or remain constant (Panel B). Columns 4-6 allow the bail
setting habits of the magistrates to vary according to offense, criminal history and
demographics of the defendant. In Column 4, the magistrate dummies are interacted
with dummies for the five most common lead charges: drug possession, first offense
DUI, robbery, selling drugs, and aggravated assault. In Column 5 I add interactions
between magistrate dummies and the number of prior cases/prior violent crimes, dummies for having at least one prior case, having a detainer, and being African-American
or female. In Column 6 I add interactions between the magistrate dummies and the
other most common crime types as listed in Table 2. The number of instruments per
specification as well as the F statistic of joint significance on the first stage instruments

17

are included in the bottom panel.
Both effect sizes and standard errors decrease as instruments are added. This
suggests two things: that allowing the bail-setting habits of the magistrates to vary
across defendant characteristics both increases the power and reduces non-monotonicity
bias in the results. In particular, if treatment effects are smaller among crime types
for which the monotonicity assumption is violated, then the estimates in Columns 1-3
will be biased upwards. It should be noted, however, that non-monotonicity bias will
not generate spurious results if no treatment effects exist. Under the null hypothesis
it would be very unlikely to see effect sizes as large as those shown in Table 5.
My preferred specification is the one where magistrate’s preferences are allowed to
vary across all 16 of the most common crime types, across the criminal history, race,
and gender of the defendant, and over the three time periods. I estimate that pretrial
detention leads to a 6.6 percentage point increase in the likelihood of being convicted
and a 5.3 percentage point increase in the likelihood of pleading guilty. Compared
to the means for each dependent variable, that converts into a 13% increase in the
probability of conviction and a 21% increase in the likelihood of pleading guilty.
Table 6 shows how pretrial detention affects conviction rates, guilty pleas, court
fees, the likelihood of being incarcerated, and both the maximum and minimum incarceration sentence.22 Panel A shows results from the jackknife instrumental variables
method with the most fully interacted specification and Panel B shows results from an
OLS regression controlling for the full set of offense, criminal history, demographic and
time controls. With the exception of court fees and the incarceration dummy, results
do not vary substantially between IV and OLS. This suggests that researchers who are
interested in estimating the effects of pretrial detention in other jurisdictions may be
able to achieve reasonable results with standard court data even in the absence of a
natural experiment.
The IV estimates suggest that pretrial detention leads to an average increase of
$128 in non-bail court fees owed. Conditional on being convicted, court fees average at
$775, and $1250 if convicted of a felony. For the tens of thousands of people convicted
as a result of pretrial detention – many of whom were unable to pay even fairly small
amounts of bail – these court fees may pose a significant burden. The IV results for the
likelihood of being incarcerated and the maximum incarceration sentence are positive
but noisy, however the estimates on the minimum incarceration sentence are more
precise. Pretrial detention leads to an expected increase of 140 days in sentence length
before being eligible for parole. Conditional on receiving a sentence of incarceration,
defendants spend an average of 22 months in jail before being eligible for parole.
In results not shown here, I test to see if pretrial detention affects post-disposition
22

Sentence length is coded as zero for individuals who do not receive an incarceration sentence.

18

crime. The estimates are noisy but are generally negative, suggesting that some crimes
may be averted as a result of increased incapacitation. Forthcoming work will provide
a cost-benefit analysis of the impact pretrial detention has on pretrial crime, including
novel estimates of the cost of incarceration to the incarcerated (Mayson and Stevenson,
2016).

8

Results for various subsamples
In Table 7 I show results for misdemeanors and felonies using both IV and OLS

techniques; the discussion below focuses on the IV results.2324 The effect sizes of the
felony sample are similar in magnitude to the full sample, but are noisy. The effects
among misdemeanors are more precisely measured and are slightly larger than the full
sample estimates, especially in relation to the means of the dependent variables. In
fact, pretrial detention among misdemeanor defendants leads to a statistically significant increase in all outcomes. The effects on punishment are particularly large: those
detained will be 8 percentage points more likely to receive a sentence of incarceration
over a mean 16 percentage point incarceration rate. While the expected increase in
sentence length is only a month or two, this represents more than a 100% increase relative to the mean.25 Those who are given an incarceration sentence for a misdemeanor
crime will spend an average of 9 months in jail before being eligible for parole.
In Table 8 I compare effect sizes across weak-evidence crimes and strong-evidence
crimes. Many of the strong-evidence crimes are possession crimes, where drugs or illegal
firearms were found on the defendant’s body or in her home or car. Shoplifting crimes
usually entail store merchandise found on the defendant’s person as they walk out of
the store. In DUI’s, the defendant was found behind the wheel with an elevated blood
alcohol level, and prostitutes are usually arrested after soliciting from an undercover
officer. The main piece of evidence among strong-evidence crimes is the police officer’s
statement and/or a lab report; these types of charges can be difficult to contest.
Most weak-evidence crimes are complainant offenses (an arrest was made as a result
of a complaint). Evidence in these types of crimes are much more likely to consist of
victim testimony, eyewitness accounts, surveillance/bystander video, alibis, character
23

The felony sample is defined as those who were charged with at least one felony at the time of the bail
hearing; many of these had their charges downgraded to misdemeanors only by the time of the arraignment.
24
The IV specifications allow the magistrate preferences to vary across time and across defendant characteristics, as shown in Column 6 of Table 5.
25
This increase in sentence length could not be explained solely by detained defendants being released
with time-served. If all of the defendants who were convicted as a result of detention were given time served
this would result in an average increase of only 12.24 days. (The average number of days detained for those
unable to make bail within 3 days is 144, I estimated that detention leads to a 8.5 percentage point increase
in conviction. 144 ∗ 0.085 = 12.24)

19

testimonials, identification from police lineups, etc. With this type of evidence it
can be harder to discern the facts of the case. Assaultive crimes in particular may
have multiple conflicting accounts of what occurred and why. Witness testimony can
be inconsistent, videos blurry, alibis hard to verify, and police lineup identification is
notoriously unreliable. While a number of theft crimes are included in this category,
possession of a stolen object does not automatically imply culpability. For example,
passengers in a stolen car may not be aware that the car was stolen.
My hypothesis is that extra-legal factors such as detention status will have less
effect on cases where the evidence is strong than they will in cases where the facts are
difficult to discern. Prosecutors are unlikely to drop charges if the evidence is strong,
nor will they lose if the case is brought to trial. While a detained defendant may
plead guilty sooner or to more unfavorable terms, the effect on conviction should be
minimal. In a weak-evidence case, however, the defendant’s willingness to wait may
prove important. Cases where the evidence is weak are much more likely to be dropped,
or to result in acquittal at trial. Furthermore, such cases may rely on the testimony of
witnesses who are reluctant to cooperate or whose memory fades over time. In fact, if
the prosecution’s key witnesses fail to appear four times in a row, the case is dropped.
If detention leads defendants to plead guilty or move to trial quicker than they would
otherwise this could be a significant disadvantage.
Detention status may also directly affect the evidence available among weak-evidence
crimes. 75% of the sample is represented by a public or court appointed attorney; the
high case volumes handled by these attorneys suggest that they may not have the time
to do as much investigative work as is necessary. A released defendant can contact
eyewitnesses, secure surveillance video, take photos of the crime scene, and otherwise
collect exculpatory evidence. A released defendant can also pressure witnesses, destroy
evidence, or otherwise impede the investigation.
While the standard errors in the IV estimates are large enough that definitive
conclusions can’t be drawn, the results generally suggest that effect sizes are larger
among weak-evidence crimes. The difference is particularly striking for sentence length:
those detained on a weak-evidence crime can expect to be sentenced to an additional
18 months in prison before being eligible for parole. With the exception of court fees,
the IV effects for strong-evidence crimes are close to zero and are not statistically
significant.
The OLS results also support the hypothesis that effect sizes are larger among weakevidence crimes. The OLS estimates of the effect pretrial detention has on conviction
are 0.098 among weak-evidence crimes and 0.007 among strong-evidence crimes. Figure
4 shows OLS effects by crime category (the IV results are too noisy for such small
samples). Strong-evidence crimes are at the top and weak-evidence crimes are at the

20

bottom. The coefficient plot, which shows the estimated effect that pretrial detention
has on guilty pleas, again suggests that effect sizes are larger among weak-evidence
crimes.
Table 9 shows IV results for blacks, whites, young defendants, older defendants,
those with one or no prior arrests, and those with more extensive criminal history.
Overall, there is nothing to suggest that effect sizes differ across race. The point
estimates are generally quite similar, although the subsample is small enough that
many are not statistically significant. Results are also similar among younger and older
defendants. The point estimates for sentence outcomes are greater among the younger
defendants, but the standard errors are large as well. We do, however, see suggestive
evidence that effect sizes are larger for those with limited prior interactions with the
criminal justice system. Among first or second time arrestees, pretrial detention leads
to a 12 percentage point increase the likelihood of pleading guilty and a 17 percentage
point increase in the likelihood of being convicted.26

9

Robustness checks and effect sizes for varying

definitions of pretrial detention
In Table 10 I present several robustness checks for the full sample results. Panel A
is identical to Panel A of Table 6 except that magistrate dummies are also included as
controls in the second stage regression. Controlling for the eight magistrate dummies
implies that the impact pretrial detention has on case outcomes is being identified solely
off of within-magistrate variations in detention rates. In particular, these controls
will absorb any other fixed aspects of the magistrates that may affect the results.
For example, the most lenient magistrate may also be particularly encouraging or
supportive during the bail hearing. If this affects the defendant’s expectations of success
at trial – and thus their willingness to accept a guilty plea – this would undermine
the exogeneity assumption. If the effect sizes change greatly as a result of including
magistrate fixed effects as controls this would raise concerns that the magistrates are
affecting case outcomes through channels other than pretrial detention. Panel A shows
that although the inclusion of magistrate fixed effects increases the standard errors,
the effect sizes are not changed dramatically.
Panel B is identical to Panel A of Table 6 except that controls for attorney type are
added. While attorney type is likely to be endogenous to both the bail amount and
detention status, a large change in effect size as a result of these controls would raise
concerns that the effects are being driven by variations in the magistrate’s willingness
26

The sample of first time arrestees is so small that the IV results are hard to interpret.

21

to grant public defense. Alternatively, it could suggest that the effects seen are not
as a result of pretrial detention per se, but rather the wealth affects of bail and the
ability to hire a quality lawyer. However the inclusion of controls for having a public
defender, a court-appointed attorney, or a private attorney have only trivial results on
the estimates.
In Figure 5 I look at how effect sizes differ if pretrial detention is defined as spending
at least one night in jail after the bail hearing, more than two weeks, more than thirty
days, or still being in jail at the time of disposition. The figures show regression
estimates from the IV specification where magistrate’s preferences are allowed to vary
over time and according to defendant characteristics. Figures 5a and b show results
for the full sample, Figure 5c shows the weak-evidence sample and Figure 5d shows
cases for first or second time arrestees. The outcome in Figure 5b is conviction and
the outcome for all other figures is guilty pleas.
In each subplot, the effect size increases as the number of days detained increases.
The standard errors increase as well. This is because the initial bail amount set by
the magistrate becomes less relevant to detention status as time goes on (future judges
may revise bail downward).27 However despite the noisy estimates, the lower bound
on the 95% confidence interval is far from zero in some of these specifications. Among
weak-evidence crimes, the lower bound of the effect of being detained throughout the
entire pretrial period on the likelihood of pleading guilty is almost 10 percentage points.

10

Conclusion

Using a natural experiment in Philadelphia where the likelihood of being detained
pretrial is exogenously affected by the magistrate who presides over the bail hearing,
I find that pretrial detention leads to an increase in the likelihood of being convicted,
mostly by increasing the likelihood that defendants, who otherwise would have been
acquitted or had their charges dropped, plead guilty. The effects are larger among first
or second time arrestees and among crime types where evidence tends to be weak.
In Philadelphia, almost 80 percent of arrestees have bail set at $10,000 or less.
These arrestees would only need to pay up to a $1000 deposit to secure their release,
an amount that is likely to be had in savings or available to borrow by most middle
and upper middle class Americans. Yet 60% of arrestees with a $10,000 bail are unable
to pay this amount within three days and 34% remain in jail at the time of disposition.
Some argue that money bail is unconstitutional since it is so difficult to ensure that
the price of bail is set proportional to one’s means in a way that precludes detention
27

I chose ‘detained more than three days’ as my preferred measure since it seemed a reasonable balance
between strength of instrument and size of the effect.

22

based on wealth. In fact, many jurisdictions do not take ability to pay into account
at all when setting bail, a practice that is clearly a violation of the Equal Protection
Clause.28 A nonprofit legal organization called Equal Justice Under Law has been filing
a series of lawsuits to ensure that the defendant’s ability to pay is taken into account
in the setting of bail.
Yet this is not the only important legal/policy question which hinges, at least
partly, on the full costs of pretrial detention to the defendant. Another as-of-yetunanswered question is whether defendants have a right to counsel at the bail hearing.
The Supreme Court has ruled that the state must pay for indigent defendants to have an
attorney at all ‘critical stages’ of the criminal proceedings, but what exactly constitutes
a critical stage is not completely clear. Many jurisdictions, including Philadelphia, do
not provide counsel to indigent clients at the bail hearing.
Another outstanding question is whether detention places undue coercive pressure
on defendants to plead guilty. Guilty pleas are required to be voluntary by the Due
Process Clauses of the Fifth and Fourteenth Amendments and if the ‘punishment’
of waiting in jail until trial is worse than the penalties involved with pleading guilty
then a plea may might not be considered voluntary. Finally, the Eighth Amendment
prohibits ‘excessive bail’. Since the full costs of bail include its effect on case outcomes,
bail amounts that might have been considered reasonable when only weighing the costs
of a pretrial loss of liberty against the benefits of averted crime may seem excessive
when including all the downstream effects of pretrial detention.29
The findings presented in this paper, taken in context of these four highly policyrelevant questions, suggest several things. First they suggest that jurisdictions should
move away from a means-based method of determining who is detained pretrial, as
the socio-economic disparities of such a system ripple far beyond the pretrial period.
Second, they suggest that the bail hearing is indeed a critical stage, and indigent defendants should have the right to counsel. Third, they suggest that low risk defendants –
for whom the ‘punishment’ of pretrial detention is worse than the expected punishment
for the crime – should not be detained as detention may result in coerced guilty pleas.
Finally, while determinations of ‘excessiveness’ are beyond the scope of this paper, the
evidence suggests that the costs of detention are high and, if justified, the benefits must
be high as well.

28

The Department of Justice issued a Statement of Interest in Varden v. City of Clanton (February
2015) declaring that fixed bail bond schedules that do not take indigence into account are a violation of the
Fourteenth Amendment.
29
For a more detailed discussion of the constitutional questions which hinge at least partly on how pretrial
detention affects case outcomes see (Heaton et al., 2016).

23

References
Abrams, David S. and Chris Rohlfs, “Optimal Bail and the Value of Freedom:
Evidence from the Philadelphia Bail Experiment,” Economic Inquiry, 2011, 49 (3),
750–770.
Aizer, Anna and Joseph J. Jr. Doyle, “Juvenile Incarceration, Human Caoital
and Future Crime: Evidence from Randomly-Assigned Judges,” National Bureau of
Economic Research Working Paper, 2009.
Allen, Ronald J. and Larry Laudan, “Deadly Dilemmas II: Bail and Crime,”
Northwestern Public Law Research Paper, December 2008, (08-44).
Anderson, James M. and Paul Heaton, “How Much Difference Does the Lawyer
Make? The Effect of Defense Counsel on Murder Case Outcomes,” Yale Law Journal,
2012, 122 (1).
Angrist, Joshua D., Guido W. Imbens, and Alan B. Krueger, “Jackknife
Instrumental Variables Estimation,” Journal of Applied Econometrics, 1999, 14 (1),
57–67.
Ares, Charles E., Anne Rankin, and Herbert Sturtz, “The Manhattan Bail
Experiment: An Interim Report on the Use of Pretrial Parole,” New York University
Law Review, 1963, 38.
Bibas, Stephanos, “Plea Bargaining Outside the Shadow of the Trial,” Harvard Law
Review, June 2004, 117 (8).
Clark, John, Daniel Peterca, and Stuart Cameron, “Assessment of Pretrial
Services in Philadelphia,” Technical Report, Pretrial Justice Institute February 2011.
Cohen, Thomas H. and Brian A. Reaves, “Pre-Trial Release of Felony Defendants in State Court,” Technical Report, Bureau of Justice Statistics Special Report
November 2007.
Devers, Lindsey, “Plea and Charge Bargaining: Research Summary,” BJA Report
January 2011.
Free, Marvin D. Jr., “Bail and Pretrial Release Decisions,” Journal of Ethnicity in
Criminal Justice, 2004, 2 (4), 23–44.
Goldkamp, John S., “The Effects of Detention on Judicial Decisions: A Closer
Look,” The Justice System Journal, 1980, 5 (3), 234–257.

24

Heaton, Paul, Sandra Mayson, and Megan Stevenson, “The Downstream Criminal Justice Consequences of Pretrial Detention,” Working Paper 2016.
Kahneman, Amos Tversky Daniel, “Prospect Theory: An Analysis of Decision
under Risk,” Econometrica, 1979, 47 (2), 263–291.
Kling, Jeffrey R., “Incarceration Length, Employment, and Earnings,” American
Economic Review, June 2006, 96 (3), 863–876.
Loeffler, Charles E., “Does Imprisonment Alter the Life Course? Evidence on Crime
and Employment From a Natural Experiment,” Criminology, 2013, 51 (1), 137–166.
Lowenkamp,

Christopher

T.,

Marie

VanNostrand,

and

Alexander

Holsinger, “Investigating the Impact of Pretrial Detention on Sentencing Outcomes,” Technical Report, Laura and John Arnold Foundation 2013.
Mayson, Sandra and Megan Stevenson, “Cost of Crime Versus the Cost of Doing
Time: Is Pretrial Detention ‘Cost-Justified’ When the Loss To Legally Innocent
Detainees is Included?,” Working Paper 2016.
Minton, Todd D. and Zhen Zeng, “Jail Inmates at Midyear 2014,” Technical
Report, Bureau of Justice Statistics Bulletin June 2015.
Mueller-Smith, Michael, “The Criminal and Labor Market Impacts of Incarceration: Identifying Mechanisms and Estimating Household Spillovers,” Working Paper
August 2015.
Oleson, J.C., Christopher T. Lowenkamp, Timothy P. Cadigan, Marie VanNostrand, and John Wooldredge, “The Effect of Pretrial Detention on Sentencing in Two Federal Districts,” Justice Quarterly, 2014.
ONDC, “”FAFSA Facts”,” Technical Report, Office of National Drug Control Policy
U.S. Department of Education.
Phillips, Mary T., “Pretrial Detention and Case Outcomes, Part 1: Nonfelony
Cases,” Final Report, New York Criminal Justice Agency, Inc. November 2007.
, “Bail, Detention and Felony Case Outcomes,” Research Brief, New York Criminal
Justice Agency, Inc. September 2008.
Rankin, Anne, “The Effect of Pretrial Detention,” New York University Law Review,
1964, 39.

25

Sacks, Meghan and Alissa R. Ackerman, “Pretrial detention and guilty pleas: if
they cannot afford bail they must be guilty,” Criminal Justice Studies, 2012, 25 (3),
265–278.
Schlesinger, Traci, “Racial and ethnic disparity in pretrial criminal processing,”
Justice Quarterly, 2005, 22 (2), 170–192.
SHRM, “”Society for Human Resource Management Survey Findings: Background
Checking - The Use of Criminal Background Checks in Hiring Decisions”,” Technical
Report, The Sentencing Project July 2012.
Shubik-Richards, Claire and Don Stemen, “Philadelphia’s Crowded, Costly Jails:
The Search for Safe Solutions,” Technical Report, Pew Charitable Trusts Philadelphia Research Inititiative May 2010.
Tartaro, Christine and Christopher M. Sedelmaier, “A tale of two counties: the
impact of pretrial release, race, and ethnicity upon sentencing decisions,” Criminal
Justice Studies, 2009, 22 (2), 203–221.
Tella, Rafael Di and Ernesto Schargrodsky, “Criminal Recidivism after Prison
and Electronic Monitoring,” Journal of Political Economy, 2013, 121 (1), 28 – 73.
Thacher, David, “The Rise of Criminal Background Screening in Rental Housing,”
Law & Social Inquiry, 2008, 33 (1), 5–30.
Uggen, Christopher, Sarah Shannon, and Jeff Manza, “”State-Level Estimates
of Felon Disenfranchisement in the United States, 2010”,” Technical Report, The
Sentencing Project July 2012.
Williams, Marian R., “The Effect of Pretrial Detention on Imprisonment Decisions,”
Criminal Justice Review, 2003, 28 (2), 299–316.
Wooldredge, John, “Distinguishing Race Effects on Pre-Trial Release and Sentencing
Decisions,” Justice Quarterly, 2012, 29 (1), 41–75.

26

Table 1: Summary statistics

Age
Male
Caucasian
African-American
Missing race
Number of prior cases
Has felony charge
Number of charges
Bail
Bail=$0
Detained>3 days
Never released
All charges dropped
Case went to trial
Not guilty on all charges
Guilty of at least one charge
Pled guilty to at least one charge
Court fees charged
Sentenced to incarceration
Days of incarc. sentence
Days before elig. for parole
Observations

Released
32.9
0.79
0.30
0.53
0.15
4.14
0.36
4.95
$3,345
0.54
0
0.00
0.48
0.32
0.03
0.49
0.21
$386
0.18
94
42
197,775

Detained Total
32.0
32.5
0.89
0.83
0.26
0.29
0.65
0.57
0.05
0.11
6.25
4.97
0.75
0.51
6.71
5.64
$63,336 $26,877
0.01
0.33
1
0.41
0.58
0.23
0.48
0.48
0.19
0.27
0.03
0.03
0.49
0.49
0.33
0.26
$211
$317
0.34
0.24
585
292
331
159
133,840 331,615

Note: Summary statistics are presented for those who are released within
three days of the bail hearing (Column 1), those who are detained for more
than three days after the bail hearing (Column 2) and the entire sample
(Column 3). All offense variables refer to the charges present at the time
of the bail hearing. The statistic shown is the mean and, unless otherwise
indicated, variables are dummies where 1 indicates the presence of a characteristic. Age is measured in years, those marked “Number...” are count
variables, and those expressed in dollar amounts are currency. The bottom
two rows refer to the maximum number of days of the incarceration sentence and the minimum number of days before the defendant is eligible for
parole. The sentence is coded as zero if the defendant did not receive an
incarceration sentence.

27

Table 2: Main offenses
DUI, 1st offense
Prostitution
Shoplifting
Small amount marijuana
Illegal firearms (F2 and F3)
Drug possession
Buying drugs
Selling drugs
Car theft
Theft
Robbery
Burglary
Murder
Vandalism
Simple assault
Aggravated assault
Observations

0.065
0.020
0.042
0.022
0.036
0.14
0.053
0.13
0.021
0.042
0.073
0.046
0.021
0.011
0.065
0.091
331615

Note: The statistic shown is the fraction
of the overall sample whose lead charge is
as listed. Crime types are listed according to strength of evidence; crime types
at the top of the list tend to have stronger
evidence than those at the top of the list.
Strength of evidence is measured both by
a poll of criminal justice lawyers and by
average conviction rate. The horizontal
line separates those placed in the ‘weakevidence’ category and those placed in the
‘strong-evidence’ category

28

Table 3: How do race and neighborhood wealth relate to bail amount and pretrial detention?

African-American
Log income (average per zip code)
Observations
R2
Demographic controls
Criminal history controls
Offense controls

(1)
Log bail amount
0.00197
(0.0125)
-0.0451∗∗∗
(0.0144)
251236
0.584
Y
Y
Y

(2)
Pretrial detention
0.0278∗∗∗∗
(0.00178)
-0.0217∗∗∗∗
(0.00205)
251236
0.336
Y
Y
Y

Standard errors in parentheses
Heteroskedastic-Robust Standard Errors
∗
p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01, ∗∗∗∗ p < 0.001

Note: This table shows the how race and neigborhood wealth correlate with bail
amount and pretrial detention status, after controlling for offense, criminal history,
age, gender and time. Only the subset of defendants for whom zipcode and race information was available were included

Table 4: Covariate balance across magistrates
Summary
Predicted
Predicted
Predicted

statistics for defendant characteristics
likelihood of pretrial detention
likelihood of pleading guilty
likelihood of conviction

F statistic
1.84
2.59
1.91

Empirical p value
0.96
0.56
0.58

Note: The predicted likelihoods described in the left-most columns are the predicted values
from a regression of pretrial detention, guilt, and conviction, respectively, on offense, criminal
history, demographics and time controls. The F statistics are the F statistics in a test of joint
significance of eight magistrate dummies when regressing the predicted values on the magistrate dummies and controls for the time and date of the bail hearing. The empirical p values
show the likelihood of seeing an F statistic as big or bigger if the magistrate seen was due to
chance; this is calculated as the fraction of ‘false’ F statistics as big or bigger in a permutation
test.

29

Table 5: How does pretrial detention affect conviction rates and guilty pleas?
Panel A: Full sample (IV)
Pretrial detention

(1)
0.166∗∗
(0.0734)

Conviction (mean dep. var.= 0.49)
(2)
(3)
(4)
(5)
0.181∗∗∗ 0.252∗∗∗ 0.122∗∗∗ 0.0871∗∗
(0.0653) (0.0794) (0.0411) (0.0369)

(6)
0.0665∗∗
(0.0293)

(1)
0.124∗∗
(0.0617)

Guilty pleas (mean dep. var.=0.25)
(2)
(3)
(4)
(5)
∗∗∗
∗∗
∗∗∗
0.174
0.175
0.104
0.0582∗
(0.0561) (0.0715) (0.0364) (0.0329)

(6)
0.0531∗∗
(0.0265)

Panel B: Full sample (IV)
Pretrial detention
Magistrate X 3 time periods
Magistrate X top 5 crimes
Magistrate X crim. history
Magistrate X demographics
Magistrate X top 16 crimes
Time controls
Covariates
Observations
First stage F
Number of instruments
Mean indep. var

Y

Y

Y
Y

Y
Y
Y
Y

Y

Y

331615
34.87
8
0.41

331615
19.53
8
0.41

Y
Y
331615
26.09
19
0.41

Y
Y
331615
21.87
59
0.41

Y
Y
331615
14.75
107
0.41

Y
Y
Y
Y
Y
Y
Y
331615
11.65
203
0.41

Standard errors in parentheses
Heteroskedastic-Robust Standard Errors
∗
p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01, ∗∗∗∗ p < 0.001

Note: The dependent variable in Panels A and B respectively is a dummy equal to one if the defendant is
convicted on at least one charge and a dummy equal to one if the defendant pled guilty on at least one
charge. The instruments in the first two columns are the eight magistrate dummies; in the subsequent
columns the instruments include interactions between the magistrate dummies and three time period fixed
effects, the five most common crime types, a variety of criminal history variables, defendant demographics,
and the remainder of the 16 most common crime types as shown in Table 2. The first two columns control
only for the time and date of the bail hearing, all subsequent columns include the full set of controls as
described in Footnote 21. The F statistic on the exogenous first stage instruments is listed at the bottom,
as are the number of instruments used in that specification and the mean of the independent variable. A
linear jackknife instrumental variables regression is used. The Rˆ2 is not reported due to difficulties of
interpreting this statistic in an IV regression.

30

Table 6: Full sample results - jackknife IV and OLS
(1)
Conviction

(2)
Guilty
plea

(3)
Court
Fees

(4)
Incarceration

(5)
Max
days

(6)
Min
days

Panel A: Full sample (IV)
Pretrial detention 0.0665∗∗
(0.0293)

0.0531∗∗
(0.0265)

128.7∗∗∗∗
(33.59)

0.0193
(0.0251)

119.3
(73.40)

140.6∗∗
(61.86)

Panel B: Full sample (OLS)
Pretrial detention 0.0355∗∗∗∗
(0.00197)
Observations
331615
First stage F
11.65
Mean dep. var.
0.49
Mean indep. var.
0.41

0.0558∗∗∗∗
(0.00181)
331615
11.65
0.25
0.41

-103.0∗∗∗∗
(2.621)
331615
11.65
$312
0.41

0.1000∗∗∗∗
(0.00166)
331615
11.65
0.24
0.41

136.9∗∗∗∗
(3.430)
331613
11.65
292
0.41

69.88∗∗∗∗
(2.518)
331613
11.65
155
0.41

Standard errors in parentheses
Heteroskedastic-Robust Standard Errors
∗
p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01, ∗∗∗∗ p < 0.001

Note: This table shows how pretrial detention affects various case outcomes using both a jackknife IV
regression (Panel A) and an OLS regression (Panel B). The outcome variables are dummies for being
convicted/pleading guilty, total non-bail court fees in dollars, a dummy for whether or not the defendant receives an incarceration sentence, the maximum days of that incarceration sentence and the
minimum days the defendant must serve before being eligible for parole. In all of the IV specifications
magistrate preferences are allowed to vary across three time periods and according to offense, criminal history and demographics of defendants. The F statistic on the first stage of the jackknife IV are
shown in the sub-panel, as are the means of the dependent and independent variables. All regressions
include the full set of controls as described in Section 6.

31

Table 7: Comparing results for misdemeanors and felonies
(1)
Conviction
Panel A: Misdemeanors (IV)
Pretrial detention 0.0850∗∗
(0.0366)

(2)
Guilty
plea

(3)
Court
Fees

(4)
Incarceration

(5)
Max
days

(6)
Min
days

0.0684∗∗
(0.0300)

93.48∗∗
(37.85)

0.0851∗∗∗
(0.0279)

66.00∗∗∗
(21.64)

30.68∗∗
(12.13)

-13.34∗∗∗∗
(3.081)

0.0513∗∗∗∗
(0.00213)

39.23∗∗∗∗
(2.087)

19.62∗∗∗∗
(1.401)

Panel B: Misdemeanors (OLS)
Pretrial detention 0.0190∗∗∗∗ 0.0515∗∗∗∗
(0.00298) (0.00249)
Observations
First stage F
Mean dep. var.
Mean indep. var.

163125
12.82
0.50
0.23

163125
12.82
0.16
0.23

163125
12.82
$351
0.23

163125
12.82
0.16
0.23

163124
12.82
48
0.23

163124
12.82
19
0.23

Panel C: Felonies (IV)
Pretrial detention
0.0598
(0.0433)

0.0545
(0.0414)

136.7∗∗
(53.66)

-0.0214
(0.0398)

147.5
(135.3)

179.0
(115.2)

Panel D: Felonies (OLS)
Pretrial detention 0.0514∗∗∗∗
(0.00266)

0.0589∗∗∗∗
(0.00259)

-174.2∗∗∗∗
(4.023)

0.134∗∗∗∗
(0.00244)

194.7∗∗∗∗
(5.583)

98.10∗∗∗∗
(4.053)

168490
7.92
0.35
0.58

168490
7.92
$274
0.58

168490
7.92
0.32
0.58

168489
7.92
528
0.58

168489
7.92
294
0.58

Observations
First stage F
Mean dep. var.
Mean indep. var.

168490
7.92
0.47
0.58

Standard errors in parentheses
Heteroskedastic-Robust Standard Errors
∗
p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01, ∗∗∗∗ p < 0.001

Note: This table shows effect sizes in misdemeanor crimes (Panels A and B) and felonies (Panel C and
D). The outcome variables are dummies for being convicted/pleading guilty, total non-bail court fees
in dollars, a dummy for whether or not the defendant receives an incarceration sentence, the maximum
days of that incarceration sentence and the minimum days the defendant must serve before being eligible for parole. In all IV specifications magistrate preferences are allowed to vary across three time
periods and according to offense, criminal history and demographics of defendants. The F statistic on
the first stage of the jackknife IV are shown in the sub-panel, as are the means of the dependent and
independent variables. All regressions include the full set of controls as described in Footnote 21.

32

Table 8: Effect sizes by strength of evidence
(1)
Conviction

(3)
Court
Fees

(4)
Incarceration

(5)
Max
days

(6)
Min
days

Panel A: Weak-evidence crimes (IV)
Pretrial detention
0.0415
0.0735∗
(0.0421)
(0.0383)

-42.86
(44.04)

-0.00348
(0.0354)

516.3∗∗∗∗
(141.7)

541.1∗∗∗∗
(124.1)

Panel B: Weak-evidence crimes (OLS)
Pretrial detention 0.0983∗∗∗∗ 0.0894∗∗∗∗
(0.00316) (0.00285)

-67.93∗∗∗∗
(3.797)

0.131∗∗∗∗
(0.00250)

171.3∗∗∗∗
(7.089)

83.71∗∗∗∗
(5.703)

122742
9.10
$158
0.56

122742
9.10
0.20
0.56

122741
9.10
466
0.56

122741
9.10
276
0.56

Panel C: Strong-evidence crimes (IV)
Pretrial detention
0.0308
0.0277
(0.0415)
(0.0369)

206.1∗∗∗∗
(49.26)

0.0132
(0.0357)

3.994
(37.54)

0.857
(17.61)

Panel D: Strong-evidence crimes (OLS)
Pretrial detention 0.00707∗∗ 0.0379∗∗∗∗
(0.00289) (0.00273)

-111.3∗∗∗∗
(4.167)

0.0832∗∗∗∗
(0.00262)

98.36∗∗∗∗
(3.568)

47.45∗∗∗∗
(1.809)

165488
13.29
$435
0.27

165488
13.29
0.27
0.27

165488
13.29
187
0.27

165488
13.29
88
0.27

Observations
First stage F
Mean dep. var.
Mean indep. var.

Observations
First stage F
Mean dep. var.
Mean indep. var.

122742
9.10
0.35
0.56

165488
13.29
0.60
0.27

(2)
Guilty
plea

122742
9.10
0.25
0.56

165488
13.29
0.27
0.27

Standard errors in parentheses
Heteroskedastic-Robust Standard Errors
∗
p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01, ∗∗∗∗ p < 0.001

Note: Panels A and B show effect sizes in case types where evidence tends to be relatively weak (murder, aggravated assault, simple assault, vandalism, burglary, theft, robbery, car theft) and Panels C
and D show case types where evidence tends to be relatively strong (DUI, drug possession, illegal
firearms, drug sale, drug purchase, prostitution, shoplifting). The outcome variables are dummies
for being convicted/pleading guilty, total non-bail court fees in dollars, a dummy for whether or not
the defendant receives an incarceration sentence, the maximum days of that incarceration sentence
and the minimum days the defendant must serve before being eligible for parole. In all specifications
magistrate preferences are allowed to vary across three time periods and according to offense, criminal history and demographics of defendants. The F statistic on the first stage of the jackknife IV are
shown in the sub-panel, as are the means of the dependent and independent variables. All regressions
include the full set of controls as described in Footnote 21.

33

Table 9: Comparing results across defendant characteristics
(1)
Conviction

(2)
Guilty
plea

(3)
Court
Fees

(4)
Incarceration

(5)
Max
days

(6)
Min
days

Panel A: White defendants (IV)
Pretrial detention 0.0700
0.0265
(0.0596) (0.0557)
Observations
93937
93937
Mean dep. var.
0.55
0.29

52.43
(75.81)
93937
$361

-0.0253
(0.0541)
93937
0.27

146.7
(129.6)
93937
254

186.0∗
(102.1)
93937
124

Panel B: Black defendants (IV)
Pretrial detention 0.0698∗
0.0355
(0.0400) (0.0359)
Observations
191200
191200
Mean dep. var.
0.49
0.25

136.4∗∗∗
(45.19)
191200
$296

-0.00125
(0.0342)
191200
0.25

75.48
(111.3)
191199
357

116.2
(94.53)
191199
196

Panel C: Defendants under 30 (IV)
Pretrial detention 0.0487
0.0909
(0.0658) (0.0598)
Observations
167392
167392
Mean dep. var.
0.47
0.27

87.38
(79.66)
167392
$304

-0.0106
(0.0573)
167392
0.24

296.2
(214.3)
167391
348

286.8
(187.3)
167391
193

Panel D: Defendants over 30 (IV)
Pretrial detention 0.0748∗∗ 0.0553∗ 174.0∗∗∗∗
(0.0359) (0.0326) (40.90)
Observations
164194
164194
164194
Mean dep. var.
0.51
0.25
$320

0.0254
(0.0308)
164194
0.24

18.56
(74.48)
164193
235

55.91
(60.45)
164193
117

Panel E: First or second time
Pretrial detention 0.175∗∗
(0.0751)
Observations
113932
Mean dep. var.
0.41

-0.0333
(0.0588)
113932
0.17

194.0
(225.8)
113930
202

300.5
(192.3)
113930
108

arrestees (IV)
0.122∗
-40.23
(0.0686) (95.65)
113932
113932
0.22
$313

Panel F: Defendants with two or more prior arrests (IV)
Pretrial detention 0.0541∗
0.0387 170.6∗∗∗∗ 0.0736∗∗∗
(0.0317) (0.0286) (36.03)
(0.0282)
Observations
217683
217683
217683
217683
Mean dep. var.
0.53
0.28
$312
0.27

185.8∗∗ 180.6∗∗∗
(75.66) (63.08)
217683 217683
339
180

Standard errors in parentheses
Heteroskedastic-Robust Standard Errors
∗
p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01, ∗∗∗∗ p < 0.001

Note: This table shows effect sizes among black defendants (Panel A), white defendants (Panel
B), defendants under 30 (Panel C), defendants over 30 (Panel D), first or second time arrestees (Panel E) and defendants with two or more prior arrests (Panel F). In all specifications magistrate preferences are allowed to vary across three time periods and according to
offense, criminal history and demographics of defendants.

34

Table 10: Robustness checks
Panel A: Full sample, controlling for magistrate fixed effects (IV)
(1)
(2)
(3)
(4)
(5)
ConvGuilty
Court
IncarcMax
iction
plea
Fees
eration
days
∗∗∗
Pretrial detention 0.0371
0.0413
106.0
0.00242
85.32
(0.0326) (0.0298) (37.28) (0.0283) (92.73)
Observations
331615
331615
331615
331615 331613

(6)
Min
days
160.0∗∗
(79.78)
331613

Panel B: Full sample, controlling for attorney type
(1)
(2)
(3)
ConvGuilty
Court
iction
plea
Fees
∗∗
∗∗
Pretrial detention 0.0731
0.0587
124.3∗∗∗∗
(0.0287) (0.0260) (33.56)
Observations
331615
331615
331615

(6)
Min
days
134.9∗∗
(61.70)
331613

(IV)
(4)
Incarceration
0.0258
(0.0247)
331615

(5)
Max
days
114.4
(73.05)
331613

Standard errors in parentheses
Heteroskedastic-Robust Standard Errors
∗
p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01, ∗∗∗∗ p < 0.001

Note: This table presents robustness checks for the main results. Panel A includes magistrate fixed effects as controls; the effects of pretrial detention are thus identified solely off of
within-magistrate variation in detention rates. Panel B includes controls for attorney type:
public defender, court-appointed attorney, and private. The outcome variables are dummies for being convicted/pleading guilty, total non-bail court fees in dollars, a dummy for
whether or not the defendant receives an incarceration sentence, the maximum days of that
incarceration sentence and the minimum days the defendant must serve before being eligible for parole. In all specifications magistrate preferences are allowed to vary across three
time periods and according to offense, criminal history and demographics of defendants.
All regressions include the full set of controls as described in Footnote 21.

35

Figure 1: Descriptive graphs
(a) Days detained pretrial, conditional on being detained more than three days

(b) Bail amounts and detention status

3.0e+04

Median Mean
$0
$1-$2000

2.0e+04

$10001-$25000
$25001-$50000

1.0e+04

People detained

$2001-$5000
$5001-$10000

$50001-$100000
$100001-$500000
$500000+

0

0
0

100

200

300

400

500

25,000

50,000

75,000

100000

125000

600
Detained

Days detained

Released

(c) Likelihood of pleading guilty at different
levels of sentence exposure

(d) Likelihood of conviction at different levels
of sentence exposure

(e) Likelihood of being detained pretrial at
different levels of sentence exposure

(f) Distribution of 69 ‘empirical p values’
from permutation test

Note: Figure 1a shows the number of people detained/released at various levels of bail. Figure 1b shows the
number of days detained pretrial for those who are detained for more than three days. Figures 1c, d and
e are a binned scatterplots showing the fraction pleading guilty, convicted, and detained pretrial at various
levels of sentence exposure. ‘Sentence exposure’ is a log transform of the predicted value from a regression
of days of the incarceration sentence on offense, criminal history, demographics and time controls, with the
sample limited to those who were found guilty at trial. Figure 1f is a histogram of the ‘empirical p values’
of 69 different permutation tests to evaluate covariate balance across magistrates.

36

Figure 2: Average detention rates by magistrate for different offense types
(a) All cases (n=331,615, F=38.56)

(b) Shoplifting (n=15,775, F=31.82)

(c) DUI, 1st offense (n=25,850, F=25.88)

(d) Simple assault (n=85,396, F=24.93)

(e) Prostitution (n=6,529, 42.14)

(f) Drug possession (n=109,042, F=18.61)

Note: This figure shows pretrial detention rates by magistrate over the whole sample (Figure 2a), and for
different offense categories (Figure 2b-f). The numbers 1 through 8 delineate the different magistrates. The
y axes show the residuals from a regression of pretrial detention on time controls. The error bars indicate
95% confidence intervals for the mean. n indicates the number of observations per category, and the F
statistic refers to a joint F statistic on the eight magistrate dummies when regressing pretrial detention
on the magistrate dummies and time controls. The numbering of the magistrates is consistent across all
samples.

37

Figure 3: Visual IV
(a) Full sample – conviction rates and pretrial
detention are residualized over time controls

(b) Full sample – conviction rates and pretrial
detention are residualized over time controls,
offense, criminal history and demographics

(c) Weak-evidence crimes – guilty plea rate
and pretrial detention are residualized over
time controls, offense, criminal history and
demographics

(d) Strong-evidence crimes – guilty plea rate
and pretrial detention are residualized over
time controls, offense, criminal history and
demographics

Note: The y and x axes in Figure 3a show the residuals from a regression of a dummy for conviction and
pretrial detention (respectively) on controls for the time and date of the bail hearing. Figure 3b is the
same, except conviction and detention have been residualized over offense, criminal history and demographic
covariates as well as time controls. The circles in Figures 3a-b show the average detention and conviction
residuals for each magistrate; the size of the circle is proportional to the number of cases seen by that
magistrate. The y and x axes in Figures 3c-d are residuals from a regression of pleading guilty on offense,
criminal history, demographic and time controls. Figure 3c shows the weak-evidence sample and Figure 3d
shows the strong-evidence sample. Here the circles represent the average detention and guilty plea residuals
per-magistrate-per-time period. There are three time periods as separated by February 23, 2009 and February
23, 2011.

38

Figure 4: OLS estimates of the impact pretrial detention has on guilty pleas within different
offense categories, ordered by strength of evidence

Note: The above coefficient plots show the OLS estimates of the impact pretrial detention has on guilty pleas
for different offense, as labeled on the left. The offenses are ordered according to the strength of evidence
that tends to be present for different case types; the offenses on the top tend to have the strongest evidence.
Each dot represents the estimated coefficient on pretrial detention, the line represents the 95% confidence
interval.

39

Figure 5: Coefficient plots showing the impact of pretrial detention on case outcomes using
various definitions of ‘pretrial detention’
(a) Full sample

(b) Full sample

(c) Weak-evidence crimes

(d) First or second arrest

Note: The above coefficient plots show jackknife IV regression results where the endogenous independent
variable is defined as being detained at least one day, greater than three days, greater than 14 days, greater
than 30 days, or until the time of disposition. The top two plots show results from the full sample, the
bottom left plot shows results from the ‘weak-evidence’ sample and the bottom right shows results from the
sample for first or second time arrestees. The instruments in all specifications are the magistrate dummies
interacted with offense, criminal history, race, gender, and three time periods; the full set of controls are
included. The dot shows the magnitude of the coefficient estimate as indicated on the x axis and the line
indicates a 95% confidence interval.

40

 

 

The Habeas Citebook Ineffective Counsel Side
Advertise here
BCI - 90 Day Campaign - 1 for 1 Match