1  

Gender-Based Favoritism Among Criminal Prosecutors

Stephanie Holmes Didwania1

Abstract

Criminal prosecutors enjoy wide discretion in the decisions they make but are largely unstudied

by empirical scholars. This paper explores gender bias in prosecutorial decision-making. I find

that defendants are charged more leniently when they are the same gender as their prosecutor as

opposed to when the defendant and prosecutor are different genders. Such favoritism in charging

ultimately translates into significantly lower sentences for defendants who are paired with a same-

gender prosecutor. Gender-based leniency, however, is not uniform across all geographic areas,

all types of cases, and all defendant characteristics. It is more pronounced in states with less

prevalent sexism, cases in which gender is likely to be more salient, and on same-gender

prosecutorial teams. However, gender match in defendant-prosecutor pairs is not strongly

associated with differences in cooperation and bargaining. I conclude that prosecutors’ social

preferences are more likely to explain gender-based leniency than differences in how male and

female prosecutors work.

                                                            1 Assistant Professor of Law, Temple University Beasley School of Law. I am grateful to Emily Buss, Adam Chilton,

Travis Crum, Dhammika Dharmapala, Vikas Didwania, Nate Ela, Craig Garthwaite, J.B. Heaton, William Hubbard, Benjamin Jones, Cree Jones, Emma Kaufman, Amanda Kleintop, Genevieve Lakier, Dorothy Lund, Anup Malani, Jonathan Masur, Tom Miles, Margot Moinester, Rachel Montgomery, Meghan Morris, Jeffrey Omari, Manisha Padi, Asad Rahim, Weijia Rao, Kyle Rozema, CJ Ryan, Max Schanzenbach, Roseanna Sommers, Lior Strahilevitz, and participants at the Conference on Empirical Legal Studies, the Annual Meeting of the American Law and Economics Association, the Marquette University Law School Faculty Workshop; the Midwestern Law and Economics Association, the UC Berkeley Law, Economics, and Business Workshop; and the University of Chicago Law School Faculty Works-in-Progress Workshop for valuable comments and suggestions. Comments are welcome at didwania@temple.edu.

2  

Introduction

Observers have long commented on the immense power our criminal system bestows upon

its prosecutors. Nearly 80 years ago, then-U.S. Attorney General Robert H. Jackson famously

remarked, “The prosecutor has more control over life, liberty, and reputation, than any other person

in America. His discretion is tremendous”2 (Jackson 1940). Prosecutors alone have the power to

initiate criminal cases, and they exercise important decision-making authority throughout the

criminal process, for example by selecting which charges to bring against defendants and by

engaging in plea bargaining.

Although prosecutors exercise vast discretion in criminal cases, their behavior is largely

unstudied by quantitative empirical researchers.3 Like all people, however, prosecutors are

susceptible to bias in decision-making. In fact, because prosecutors enjoy expansive and largely

unreviewable discretion, one might expect the prosecutorial setting to be one in which bias is

especially likely to affect behavior.

This paper focuses on one important kind of bias: in-group favoritism. In-group favoritism

occurs when a decision-maker gives preferential treatment to those who share a salient trait with

the decision-maker, such as being a member of their gender, racial, ethnic, or religious group

(Everett, Faber, and Crockett 2015).4 Scholars have documented evidence of gender-based in-

group favoritism in a variety of laboratory and experimental settings, and a growing consensus

suggests that the majority of discrimination in the United States takes the form of in-group

favoritism rather than outright hostility towards people who are different (Greenwald and

Pettigrew 2014, Krieger 1998).

This paper explores in-group favoritism on the basis of gender.5 Researchers have shown

that in laboratory experiments, both men and women exhibit favoritism towards members of their

                                                            2 At the time of this statement, Robert H. Jackson was not yet a Supreme Court justice. 3 On the other hand, there is considerable theory on the promise and pitfalls of the way prosecutor offices self-

police, incentives that prosecutors face, and prosecutorial ethics (Bibas 2009, Leonetti 2012, Meares 1995, Ouziel 2017, Pfaff 2017, Sklansky 2017).

4 In this paper, I use the term in-group favoritism to represent a phenomenon: that decision-makers treat people relatively more favorably when the person is in their in-group It is important to emphasize that I do not take a view about the underlying causes of this differential favoritism. For example, I do not take a position on whether the effect that I document is conscious or subconscious, nor do I have a view about the extent to which it is driven by out-group disfavoritism versus in-group favoritism.

5 Because of the way the data is coded, this paper classifies people as having a binary gender that is either female or male. Unfortunately, neither of the data sources used in the paper include information about transgendered defendants or defendants who do not identify as having binary gender. Rosenblum (2000) provides a thoughtful discussion of the legal issues facing transgendered prisoners.

3  

own gender (Hoyt, Simon, and Reid 2009; Lindeman & Sundvik 1995), although some have found

that women demonstrate stronger in-group favoritism then men (Rudman and Goodwin 2004),

while others have found the opposite (Vial et al. 2017). Outside of lab settings, in-group favoritism

on the basis of gender has been documented in educational, professional, and financial settings.

For example, one recent working paper finds evidence of this phenomenon in attorney disciplinary

hearings: reporting that female attorneys are treated less harshly in such hearings when the panel

contains more female appellate judges (Kennedy, McDonnell, and Stephens 2017). Others have

documented in-group favoritism in the financial industry, finding, among other things, that male

analysts tend to under-forecast the earnings of firms led by female CEOs (Jannati et al. 2016).

This paper quantifies gender-based in-group favoritism among federal prosecutors with an

expansive data set that identifies—among other variables—the gender of the lead prosecutor

assigned to each case. Using data that covers nearly 150,000 federal criminal defendants sentenced

in fiscal years 2002 through 2016, I find that while male and female prosecutors exhibit small and

statistically insignificant differences in their treatment of defendants overall, they show relative

favoritism towards defendants of their own gender. The leniency associated with gender match in

defendant-prosecutor pairs ultimately translates into significant differences in sentence length for

defendants who match with their prosecutor on gender compare to those who do not.

To make sense of the findings, the paper also considers heterogeneity in the results.

Gender-based leniency is not uniform across all geographic areas, case types, and defendants. For

example, I show that the results are most pronounced in states with below-median levels of sexism.

The results are also stronger in cases in which gender is likely to be more salient, and among same-

gender teams of prosecutors (relative to prosecutors working alone and mixed-gender teams). On

the other hand, prosecutor-defendant gender match does not seem to influence the intensity of any

bargaining between defendants and prosecutors. Similarly, prosecutor-defendant gender match

does not appear to affect the extent to which the defendant cooperates with the prosecutor except

among those defendants for whom cooperation is likely to be incredibly valuable, where the

relationship is marginally significant. I conclude that this evidence suggests that the results are

more likely to be generated by preference-based, rather than process-based, channels.

The research setting presents a few important challenges. First, as is common in many

studies of in-group favoritism, the research design provides a relative—but not absolute—result.

Put another way, the research design cannot disentangle in-group favoritism (special treatment

4  

towards one’s own in-group) from out-group disfavoritism (hostile treatment towards one’s out-

group). As discussed in Section 7, one natural consequence of the research design is that the results

do not suggest an unambiguous policy solution.

A second challenge stems from the fact that prosecutors are not randomly assigned to cases,

which could bias the results. I discuss the identification challenges in Subsection 4.2 and present

several robustness checks in Section 6 that suggest that the non-random assignment of cases to

prosecutors is very unlikely to be generating the results of the paper.

Examining bias in prosecutorial decision-making is important for several reasons. First—

and perhaps most obviously—our criminal system is meant to shield defendants from bias. The

Equal Protection Clause of the Fourteenth Amendment to the U.S. Constitution, for example,

broadly dictates that the government should treat people equally,6 and gender is a category upon

which discrimination is thought to be especially invidious.7 Mitigating bias in the federal criminal

system is also a priority of the U.S. Department of Justice, which announced in 2016 that all federal

law enforcement officers and prosecutors would receive training in how to recognize and address

implicit bias.8

It is also important to study bias among prosecutors because prosecutorial decisions carry

meaningful consequences, implicating a suspect or defendant’s liberty and risking far-reaching

collateral effects on their life and livelihood. Scholars have documented the effects of contact with

the criminal justice system and criminal convictions on, for example, high school completion

(Aizer and Doyle 2015); recidivism (Di Tella and Schargrodsky 2013, Mueller-Smith 2015);

employment (Mueller-Smith 2015, Pager 2003); dependence on public assistance (Mueller-Smith

2015); wages and wage growth (Western 2002); and family formation and support (Geller,

Garfinkel, and Western 2011).

The paper proceeds as follows. Part 1 situates this paper in the prior literature. Part 2

examines the empirical setting, describing the key responsibilities of federal prosecutors and

explaining how their offices are organized. Part 3 describes the data. Part 4 presents the empirical

                                                            6 U.S. CONST, amend. XIV. 7 Craig v. Boren, 429 U.S. 190 (1976). Laws that explicitly classify people on the basis of gender are “inherently

suspect,” Frontiero v. Richardson, 411 U.S. 677, 682 (1973), and are analyzed with intermediate scrutiny, Craig, 429 U.S. at 197.

8 Department of Justice Announces New Department-Wide Implicit Bias Training for Personnel, U.S. DEP’T OF

JUSTICE (June 27, 2016), https://www.justice.gov/opa/pr/department-justice-announces-new-department-wide-implicit-bias-training-personnel. Ross (2016) provides a description of the Justice Department’s implicit-bias training.

5  

strategy and main results. Part 5 sheds light on the mechanisms responsible for the main findings

by examining heterogeneity in the results along several dimensions. Part 6 examines the robustness

of the results. Part 7 discusses the findings, considers policy interventions, and suggests directions

for further research.

1. Relevant Prior Literature

Limited quantitative work examines the behavior of criminal prosecutors, and the existing

research—while insightful—leaves many unanswered questions. Very few papers leverage

individualized data on line prosecutors. One important exception is a recent working paper by

CarlyWill Sloan, which documents an increased likelihood of conviction for property-crime

misdemeanors when the defendant and prosecutor are of different races, but not for other types of

misdemeanors (Sloan 2019).

This work contributes to the literature in several important ways. This paper is among the

first to examine bias in prosecutorial behavior using individualized data on line prosecutors. In

contrast to the majority of the literature on bias in the criminal system—which largely focuses on

race—this work considers gender. Gender disparities in federal criminal cases are large, highly

statistically significant, and difficult to explain. Assessing in-group favoritism by prosecutors

might help policymakers better understand and address gender disparity in the criminal system.

Earlier work, though largely lacking individualized data on line prosecutors, sheds

important light on the relationship between prosecutorial decision-making and career concerns by

analyzing charging and trial decisions.9 For example, Boylan and Long (2005) find that

prosecutors are more likely to take cases to trial in federal districts in which the local labor market

for attorneys has above-average salaries.10 Boylan (2005) examines the career paths of 570 United

States Attorneys—the attorneys who lead federal prosecutorial offices—over a thirty year period

and finds that the most important predictor of favorable career outcomes for U.S. Attorneys after

leaving their U.S. Attorney’s Office is the length of prison sentences imposed on defendants who

were prosecuted by the office. Boylan (2005) acknowledges, however, that the article does not use

                                                            9 One prior paper examines in-group favoritism among prosecutors on the basis of race, but only has racial

information aggregated to the federal district level (Ward, Farrell, and Rousseau 2009). The authors find, among other things, that districts with more black representation among prosecutors have smaller black-white disparities in sentencing outcomes.

10 The findings presented in Boylan and Long (2005), however, are also consistent with an effect that is driven by defense counsel facing the same kinds of incentives that the authors attribute to prosecutors. The decision to go to trial is not solely a prosecutorial decision. It is the defendant’s Sixth Amendment right to exercise, although prosecutors undoubtedly influence the decision during plea bargaining.

6  

data on the individual line prosecutors who carry out the work of the United States Attorney’s

Offices, who are the focus of this paper.

Other prior work has studied prosecutorial decision-making by analyzing charging

behavior. Bjerk (2005) found that prosecutors strategically charge mandatory minimums, for

example, by being more likely to reduce a charge to a misdemeanor when a felony charge would

trigger a mandatory minimum under a “three-strikes” law. Others have found that charging

decisions are a critical source of gender- and race-based sentencing disparity (Rehavi and Starr

2014, Starr 2015, Yang 2014). In particular, most race-based sentencing disparity begins at the

charging stage, in large part because prosecutors are more likely to charge crimes carrying

mandatory minimum sentences when the defendant is black (Rehavi and Starr 2014). This

prosecutorial tactic appears to respond to increased judicial discretion after the Supreme Court’s

2005 decision, United States v. Booker (Starr and Rehavi 2013, Yang 2014, Yang 2015).

Together the existing research illustrates that prosecutorial decision-making could be

influenced by bias.11 However, while this prior work suggests that prosecutorial offices treat

defendants in a biased way, it does not attempt to explain how bias enters the decision-making

process.12 This paper begins to fill this gap by examining behavior with individualized information

about line prosecutors—the people who enforce our criminal laws every day.

This paper is also related to two distinct bodies of literature, upon which the remainder of

this section focuses. First, prior work examines the behavior of other actors in the criminal system,

including judges, law enforcement officers, and jurors. Second, this paper is connected to prior

research on the effects of gender-based bias in organizations. The rest of this section discusses

these two areas of scholarship and describes the contributions of this paper.

                                                            11 There is mixed evidence that minority defendants also fare worse in plea bargaining. Kutateladze, Andiloro, and

Johnson (2014) and Kutateladze et al. (2014) find that black and Latino defendants are more likely to receive a custodial plea offer than white defendants, but Kutateladze et al. (2014) also finds that black and Latino defendants are more likely to have their cases dismissed. Both of these papers use data from New York City that is comprised primarily of misdemeanor defendants.

12 One important exception is a recent working paper in which the authors carried out a randomized, controlled experiment in which they surveyed prosecutors by presenting them with vignettes and asking them how they would charge in each situation. The vignettes were manipulated to change the race and class of the fictional perpetrator. The authors generally did not find evidence of bias in prosecutorial decision-making (Robertson, Baughman, and Wright 2019).

7  

1.1. Studies of Judge, Law Enforcement Officer, and Juror Behavior and Bias

Empirical scholars have long examined race- and gender-based bias by actors other than

prosecutors in the criminal system. Most of this scholarship has examined bias with respect to race

and ethnicity rather than gender, and the literature that investigates gender favoritism has produced

mixed results. One recent article finds no strong evidence that judges either favor or disfavor

defendants of the same gender (Lim, Silveira, and Snyder 2016). Earlier research reports that the

presence of more female judges on a district court reduces the court’s gender gap in sentence

length, but because this analysis is aggregated to the district court level, it does not constitute

evidence of in-group favoritism (Schanzenbach 2005). Other earlier work finds that female judges

show more gender disparity than male judges, but this study does not report whether this difference

are statistically significant (Steffensmeier and Hebert 1999).13

Considerably more work studies race-based favoritism in the criminal system, although

evidence related to in-group favoritism is mixed. For example, one recent article finds that judges

exhibit in-group disfavoritism14 towards juvenile defendants of the same race (Depew, Eren, and

Mocan 2017).15 Another recent paper, however, finds evidence that judges might slightly favor

adult defendants of their own race (Lim, Silveira, and Snyder 2016). Others similarly find that

African American judges exhibit smaller racial disparities in sentencing than their white

counterparts, which is consistent with race-based in-group favoritism (Abrams, Bertrand, and

Mullainathan 2012). In-group favoritism also presents—to some extent—along ethnic lines in

pretrial detention decisions: one study finds that Arab and Jewish judges in Israel are less likely to

detain defendants who share their ethnicity, but in-group favoritism does not affect the length of

detention ordered (Gazal-Ayal and Sulitzeanu-Kenan 2010).

In the case of juror behavior, it appears that the diversity environment—and not just the

race of the individual actor—affects outcomes. Researchers have found that juries formed from

all-white jury pools convict black defendants more often than white defendants, and that this gap

is entirely eliminated when the jury pool includes at least one black member, even when the black

                                                            13 The study also does not include fixed effects for geography or time, nor are standard errors clustered so it is hard

to draw firm inferences from the results. 14 Depew, Eren, and Mocan (2017) refer to this phenomenon as “negative in-group favoritism” in their article. For

clarity, this paper uses the term “in-group disfavoritism” to mean the same thing: that a decision-maker treats in-group members worse than out-group members.

15Another recent contribution finds that judges treat more harshly defendants who share the same first initial of their name, which the author describes as an example of implicit egotism (Chen 2017).

8  

jury pool members are not seated (Anwar, Bayer, and Hjalmarsson 2012). Recent work has also

documented evidence of favoritism on the basis of race and gender in jurors’ decisions to convict

(Flanagan 2018). The effects of racial diversity among judges are more nuanced, but do not

constitute strong evidence of race-based in-group favoritism (Schanzenbach 2005, Schanzenbach

2015).

Work that examines in-group favoritism among law enforcement officers has largely

focused on racial bias,16 and—like the work on in-group favoritism among judges—has produced

mixed results. Recent work finds that black officers are more likely than white officers to shoot

unarmed white civilians (Fryer 2016), consistent with earlier work, aggregated to the police-force

level, that police departments with more minority officers are more likely to arrest white suspects,

with little impact on the arrests of non-white suspects (Donohue and Levitt 2001). Others similarly

find that officers are more likely to conduct a search of a driver of a different race, which they

attribute to preference-based discrimination after ruling out the possibility that officers are better

at searching members of their own racial group or that the results are driven by the non-random

assignment of officers to neighborhoods (Antonovics and Knight 2009). Others, on the other hand,

do not find race-based differences in officers’ propensities to arrest non-white suspects (Brown

and Frank 2006). An empirical study of traffic stops in Oakland, California offers similarly

muddled evidence: officers appear to exhibit race-based in-group favoritism in some

neighborhoods, and the opposite in other neighborhoods (Sanga 2014).

Outside the criminal system, there is also literature that seeks to estimate in-group

favoritism in civil and extralegal settings. A study of in-group favoritism in professional basketball

famously finds that NBA referees demonstrate race-based in-group favoritism towards players

(Price and Wolfers 2010). In the civil context, Israeli judges favor defendants that share their

ethnicity in small claims cases, and in-group favoritism increases at times when ethnicity is more

salient (operationalized as years in which there was a recent terrorist attack in the vicinity of the

court) (Shayo and Zussman 2011). Earlier work has similarly found that judges’ race and gender

affect decision-making in cases in which race and gender are salient, such as affirmative action,

                                                            16 One exception is an unpublished paper that examines ticketing behavior on the basis of gender. The author finds

that, compared to male police officers, female officers are less likely overall to ticket, but relatively more likely to ticket female drivers (Rowe 2009).

9  

discrimination, sexual harassment, and voting rights cases (Boyd, Epstein, and Martin 2010; Cox

and Miles 2008; Farhang and Wawro 2004; Peresie 2005).

1.2.Gender Differences in Organizations

A rich literature in social psychology investigates differences in the behavior and treatment

of men and women in organizations. This section highlights a few findings from this literature that

are most relevant to this work, with particular emphasis on studies involving the legal profession.

This section is not intended to be exhaustive, but rather, to help understand why one might or

might not expect to find gender-based differences in how prosecutors work.

Prevailing wisdom proposes that women face different expectations than men in the

workplace. Much scholarship has focused on expectations that women exhibit warmth (Fiske et

al. 2002, Rudman and Glick 1999). These findings could drive in-group favoritism if female

prosecutors are expected to show more mercy towards defendants in particularly sympathetic

circumstances, whom are often female.17 But recent research also suggests that female attorneys

face higher standards of ethical behavior, which suggests female prosecutors might exhibit less in-

group favoritism than male prosecutors (Kennedy, McDonnell, and Stephens 2017). Relevant to

this paper, Kennedy, McDonnell, and Stephens (2017) provides evidence of in-group favoritism:

finding that the harsher treatment of female attorneys diminishes as the share of female judges on

the appellate panel increases.18

Other work considers how female managers influence hiring and pay for male versus

female employees. In the context of legal employment, for example, one study examines hiring at

large U.S. law firms in the 1990s and finds evidence of in-group favoritism in hiring: when women

make entry-level hiring decisions they are more likely to fill vacancies with women than men

(Gorman 2005).

                                                            17 Relative to male defendants, female defendants are, for example: (a) less likely to commit violence; (b) more

likely to be custodial parents; and (c) more likely to be prior victims of abuse (Starr 2015). 18 On the other hand, some anecdotal accounts assert that federal prosecutors are unique among the legal profession

in that their offices do not exhibit gender discrimination. For example, in an article published nearly thirty years ago, Judge Reena Raggi described prosecutor offices as workplaces “where gender is irrelevant,” and declared that during the seven years she spent as a prosecutor at the United States Attorney’s Office for the Eastern District of New York, she “encountered virtually no gender discrimination, and that “women have succeeded in making gender an obsolete issue” in criminal prosecution (Raggi 1989).

10  

Although much prior literature considers in-group favoritism in the employment context,

it is not clear that these findings will translate to the criminal setting. Part 2 describes the research

setting studied here—the federal criminal system.

2. Research Setting

This paper studies individual prosecutors in the federal criminal system. This section

focuses on the institutional setting, describing the responsibilities of federal prosecutors and how

they carry out their work. Of course, as described in more detail below, U.S. Attorney’s Offices

(USAOs) vary on many dimensions, including the size and organization of the offices and the

types of cases that are prosecuted. This section does not attempt to provide comprehensive,

individualized information about each office. Instead, it describes the rules and procedures that are

common to all USAOs and highlights instances in which heterogeneity is particularly prominent.

2.1.The Responsibilities of Federal Prosecutors

There are 94 geographically distinct federal district courts that cover the United States.

Each federal district court is associated with precisely one USAO, with one exception.19 Figure 1

labels and depicts the boundaries of the federal district courts. The shaded districts indicate those

districts that are represented in the data used in this paper. Each USAO is led by a United States

Attorney. The lawyers who work in USAOs are called Assistant United States Attorneys (AUSAs).

USAOs represent the United States as a party in both civil and criminal federal cases in their

districts, and typically include separate criminal and civil divisions.20

In handling criminal cases, USAOs are broadly tasked with prosecuting violations of

federal criminal law in their jurisdictions. USAOs enjoy wide discretion in how they carry out this

work, and most prosecutorial decisions are unreviewable by courts except in limited

circumstances. 21 For example, a prosecutor’s decision to initiate a prosecution and, if so, which

                                                            19 The District of Guam and the District of the Northern Mariana Islands share a USAO. 20 I collected this information by hand by visiting the website of each USAO. The majority of USAO websites

indicated that they have separate civil and criminal sections. I did not find any instances in which a USAO website stated that the USAO’s civil and criminal sections were combined. There were a handful USAOs for which it was unclear whether the civil and criminal sections were combined or separate.

21 Some argue that the role of prosecutors raises separation-of-powers concerns by endowing prosecutors with what is essentially adjudicatory power (Barkow 2009).

11  

criminal violation(s) to charge are largely unreviewable by courts. Similarly, defendants typically

cannot challenge a prosecutor’s choices during plea bargaining.

Federal prosecutors also make recommendations to the court at various points during a

criminal prosecution. Most notably, AUSAs advise the court as to whether they believe the

defendant should be detained pending trial and recommend what they think is an appropriate

sentence. Prosecutors can also support or oppose reductions or enhancements to a defendant’s

sentence. For example, under Federal Rule of Criminal Procedure 35(b) and U.S. Sentencing

Guideline § 5K1.1, a prosecutor can ask the court to reduce a defendant’s sentence if the defendant

provides substantial assistance to the government.22

Despite limited oversight from the courts, however, a prosecutor’s decision-making

authority is not absolute. Individual AUSAs are subject to both formal and informal supervision

from several sources. First, like most employees, individual prosecutors are supervised within their

organization. For the time period covered by this paper, all AUSAs were governed by the United

States Attorneys’ Manual (the “Manual”) which laid out detailed guidelines for how USAOs

should be organized and how individual prosecutors should exercise discretion.23 For example, the

Manual dictated that “[t]o ensure consistency and accountability, charging and plea agreement

decisions must be reviewed by a supervisory attorney. All but the most routine indictments should

be accompanied by a prosecution memorandum that identifies the charging options supported by

the evidence and the law and explains the charging decisions therein.”24

Second, the Manual includes instructions for deciding which charges to file. During the

time period covered by this paper, the Manual expressed a policy that federal prosecutors should

usually charge “the most serious offense that is consistent with the nature of the defendant’s

conduct, and that will probably be sufficient to sustain a conviction.”25 The Manual, however, left

room for an AUSA to deviate from this policy by conducting an “individualized assessment” of

                                                            22 The Guidelines allows a sentencing judge to give the defendant a reduced sentence (including a sentence below

the mandatory minimum) “upon motion of the government stating that the defendant has provided substantial assistance in the investigation or prosecution of another person who has committed an offense.” U.S. SENTENCING

GUIDELINES MANUAL § 5K1.1 (2016). 23 As noted in footnote 9, the United States Attorney’s Manual was replaced by the Justice Manual in September

2018. Because AUSAs were governed by the United States Attorneys’ Manual for the time period covered by this paper, I cite this version of the Manual.

24 UNITED STATES ATTORNEYS’ MANUAL § 9-27.300. 25 UNITED STATES ATTORNEYS’ MANUAL § 9-27.300.

12  

the defendant’s offense conduct and history.26 In theory, these directives ought to promote

uniformity in prosecutorial charging behavior and potentially mitigate bias.

Third, although courts are “hesitant to examine the decision to prosecute,”27 they exercise

oversight during plea bargaining and sentencing. For example, judges have the power to reject

certain kinds of plea agreements reached by the prosecutor and defendant.28 In federal court—the

setting studied here—defendants and prosecutors rarely bargain to a sentence, and judges enjoy

considerable discretion in choosing a defendant’s sentence, especially after the Supreme Court’s

decision in United States v. Booker.29 It is possible, then, that judges sentence in a way that

neutralizes any prosecutorial bias. In theory, a defendant can also challenge his prosecution on the

grounds that it was brought selectively—that is, based on a prohibited consideration such as the

defendant’s race or religion.30 In practice, however, selective prosecution challenges virtually

never succeed (McAdams 1998, Bibas 2009).

Finally, prosecutors are subject to outside pressures from the legislative branch, media,

advocacy organizations, and the general public. However, prosecutors tend to play an outsized role

in criminal justice policy-making (Barkow 2013). In the federal setting studied in this work, are

not subject to electoral pressures, and are in some ways insulated from political pressures.31 In

contrast, nearly all states elect the District Attorneys that lead their prosecutorial offices (Lantigua-

Williams 2016).

2.2.How U.S. Attorney’s Offices are Organized and Cases Assigned

USAOs largely work autonomously, although federal prosecutors occasionally collaborate

with other USAOs or federal agencies on cases. Individual AUSAs work under the control of their

                                                            26 UNITED STATES ATTORNEYS’ MANUAL § 9-27.300. 27 Wayte v. United States, 470 U.S. 598, 608 (1985). 28 Fed. R. Crim. P. 11(c)(3)(A), 11(c)(5). 29 543 U.S. 220 (2005). In Booker, the Supreme Court held that the U.S. Sentencing Guidelines are advisory; that

is, that Congress may not require district judges to sentence defendants within the Guidelines range. However, a district judge must always calculate a criminal defendant’s Guidelines range before sentencing and articulate any reasons for deviating from that advisory range. 543 U.S. at 259-60; see also Peugh v. United States, 569 U.S. 530, 530 (2013) (“District courts are required to being their sentencing analysis by looking at the ranges in the guidelines; a judge must have a good reason for deviating from those ranges”).

30 Oyler v. Boyles, 368 U.S. 448, 456 (1962). 31 Historically, U.S Attorneys have been insulated from political pressure, although some argue that the positions

have become increasingly politicized over the last decade. In 2006, President George W. Bush fired seven U.S. Attorneys, and in 2017, President Donald Trump dismissed all U.S. Attorneys that had been appointed by President Barack Obama.

13  

district’s United States Attorney (through a system of supervisory oversight) and United States

Attorneys can only be removed by the President.32 As one might expect, USAOs substantially vary

in size and organizational structure. For example, the U.S. Attorney’s Office for the District of

Hawaii employs 27 AUSAs,33 and the criminal division of the office is divided into three sections:

Drug and Organized Crime, Fraud and Financial Crimes, and Special Crime. In contrast, the U.S.

Attorney’s Office for the Central District of California—based in Los Angeles—employs 264

AUSAs, and has a criminal division that is split into 10 sections: Asset Forfeiture, Criminal

Appeals, Cyber and Intellectual Property Crimes, Public Corruption and Civil Rights, Major

Frauds, Organized Crime Drug Enforcement Task Force (OCDETF), Public Integrity and

Environmental Crimes, and Violent and Organized Crime.34 On average, a USAO is divided into

roughly four sections.35

AUSAs are typically assigned cases by their supervisors. One USAO describes the process

as follows: “Typically, cases are brought for initial review to the Criminal Division Chief by

federal agents responsible for investigating crimes, then assigned to AUSAs within units to

work.”36 USAOs appear to vary in the extent to which AUSAs specialize in certain types of cases.

For example, the District of Wyoming explains that “[d]ue to the small size of the staff, criminal

[AUSAs] must each be prepared to handle a variety of matters. Specialization is encouraged in

order for AUSAs to develop deeper expertise in important areas of the law; however, every

criminal AUSA must be prepared to handle almost any criminal matter on short notice, in the event

other AUSAs are unavailable.37 Another USAO explains, “At any given time, any one of our

AUSAs has a caseload that includes violent crime, complex financial institution fraud, health care

fraud, financial crimes, computer fraud, environmental crime, public corruption, organized crime,

complex drug and money laundering activities, and cases involving multiple defendants and

international organizations.”38 While the majority of prosecutors in the data (74 percent) have

                                                            32 28 U.S.3. § 541(c). 33 Of these 27 AUSAs, 20 are assigned to the criminal division. 34 In addition to these sections, the office also includes a national security division and a tax division, each of which

handle both civil and criminal cases. 35 I collected this data by hand by visiting the website of each U.S. Attorney’s Office. It is on file with the author. 36 UNITED STATES ATTORNEY’S OFFICE FOR THE EASTERN DISTRICT OF TENNESSEE, “Criminal Division,” available

at: https://www.justice.gov/usao-edtn/criminal-division. 37 UNITED STATES ATTORNEY’S OFFICE FOR THE DISTRICT OF WYOMING, “Criminal Division,” available at:

https://www.justice.gov/usao-wy/criminal-division. 38 UNITED STATES ATTORNEY’S OFFICE FOR THE DISTRICT OF RHODE ISLAND, “About Us,” available at:

https://www.justice.gov/usao-ri/about-us.

14  

worked on more than one offense type, around one-third (32 percent) have worked significantly

on multiple types of cases (defined as having prosecuted at least 20 cases of the offense type).

3. Data

This paper uses data from two distinct sources: (1) the United States Sentencing

Commission’s annual sentencing data files (the “Commission data”); and (2) federal case data

from the Legal Information Office Network System, which is published by the Executive Office

of United States Attorneys in response to Freedom of Information Act requests (the “LIONS

data”). This section describes the two data sources and explains how they were merged.39

3.1. United States Sentencing Commission Data

The United States Sentencing Commission (the “Commission”) is an independent agency

of the federal judiciary. One of the Commission’s primary responsibilities is to promulgate the

Federal Sentencing Guidelines—a manual that all federal judges use in sentencing criminal

defendants.40 The Commission also publishes annual data files containing detailed information

about defendants sentenced in federal district courts.41 The Commission creates these data files by

compiling information from sentencing documents submitted by the federal district courts. The

Commission extensively reviews the data it receives for completeness and quality before making

the data files available to the public on its website.42

The Commission’s individual-level data files include detailed case information, including:

demographic characteristics of the defendant; the defendant’s criminal history; the statutes under

which the defendant was convicted; the defendant’s recommended Guidelines range; the sentence

imposed; and any reasons for an out-of-range sentence.

                                                            39 The data set used in this paper was previously used in work examining the consequences of pretrial detention on

case outcomes and is described in similar terms there (Didwania 2018). 40 In Booker, the Supreme Court held that the Guidelines are advisory; that is, that Congress may not require district

judges to sentence defendants within the Guidelines range. However, a district judge must always calculate a criminal defendant’s Guidelines range before sentencing and articulate any reasons for deviating from that advisory range. 543 U.S. at 264 (“The district courts, while not bound to apply the Guidelines, must consult those Guidelines and take them into account when sentencing.”).

41 The Commission reports data for defendants convicted of felonies and Class A misdemeanors. Notably, the Sentencing Commission data files do not include information about cases involving: juvenile offenders, defendants convicted of Class B and C misdemeanors, and death penalty cases (Reedt, Semisch, and Blackwell 2013).

42 The Commission cross-checks its data for completeness and quality with data from another source—the Administrative Office of U.S. Courts. It pays special attention to cases with out of range values, logical inconsistencies, and sentences outside the Guidelines range.

15  

This paper uses Commission data for defendants sentenced in fiscal years 2002 through

2015. The primary outcome variable used in the analysis is the defendant’s base offense level,

which reflects the severity of the offense with which the defendant was charged. Other outcome

variables used include: whether the defendant faced a mandatory minimum sentence; whether the

defendant received a substantial assistance reduction; whether the defendant received less than the

mandatory minimum sentence (if facing one); the defendant’s ultimate sentence; and whether the

defendant received a sentence below the recommended Guidelines range. The paper also uses the

following defendant- and case-specific information, both as control variables and in matching

defendants between the two data sets: the defendant’s race, Hispanic ethnicity, gender, age, highest

level of education, and criminal history points; the months of incarceration and amount of

probation, fine, and supervised release imposed; the date sentenced; the lead statute of conviction;

and the types and quantities of controlled substances involved in the case, if any.

The Commission data is insufficient for purposes of this project, however, for two reasons.

Most critically, the Commission data does not contain any information about the prosecutor(s)

associated with each case. A second limitation is that the Commission data does not identify the

courthouse in which the defendant was sentenced; it simply identifies the federal district court.

Roughly 80 percent of federal district courts comprise more than one courthouse, to which

defendants, judges, and prosecutors are non-randomly assigned. It is therefore impossible for an

empirical researcher to fully account for intra-courthouse correlation using Commission data.

Another important feature of the Commission data is that it necessarily is a selected sample

because it only includes defendants who were ultimately sentenced. Issues related to sample

selection are discussed in more detail in Appendix B. The LIONS data—discussed in the next

subsection—fills these gaps.

3.2. LIONS National Caseload Statistical Data

The Executive Office for United States Attorneys (EOUSA) regularly publishes case data

from the 93 USAOs located throughout the United States. This data originates from the Legal

Information Office Network System (LIONS), which is the computer program that the EOUSA

uses to track cases. The breadth of the LIONS data is substantial—the data covers all cases in

which a USAO was involved, including those cases that are the primary responsibility of another

agency and those in which the USAO declines to prosecute. The LIONS data consists of many

16  

discrete text files that can be linked using case and participant identification numbers that are

provided in the data.

The LIONS data includes abundant case-specific information, but little defendant

information. For example, the LIONS data includes: the lead charge, the type of offense,

sentencing information including the length of incarceration ordered and whether the defendant

received a term of supervised release, probation, community service, or a fine; the sentencing

judge; the sentencing date; and the drug type(s) and quantity involved in the case. On the other

hand, the LIONS data includes little demographic information about defendants, only identifying

a defendant’s gender and country of citizenship.

Crucially, the LIONS data also provides some information about the USAO staff members

who work on each case. LIONS anonymously identifies the staff members associated with each

case, and labels their role in the case as, for example, lead attorney, co-counsel, paralegal, or victim

witness coordinator. Roughly 84 percent of staff assignments in the data are coded as “lead

attorneys.” The LIONS data also includes a variable that indicates the staff member’s salutation—

such as “Mr.” or “Ms.”—which I use to deduce the staff member’s gender. A complete list of all

salutations and how each was coded is produced in Appendix A.

It is also important to note two shortcomings of the LIONS data. First, unlike the

Commission data, there is no evidence that the LIONS data is cross-checked for accuracy with

court documents. Second, the LIONS data contains many more missing values than the

Commission data.

3.3. Merged Sample

The data used in this paper comprises a sample that is merged between the Commission

and LIONS data sources of federal defendants sentenced in fiscal years 2002 through 2016. The

sample is restricted to U.S.-citizen defendants in district courts that do not border Mexico.43 In

                                                            43 I remove non-citizens and defendants in border districts for three reasons. First, non-citizens who are charged

with immigration offenses—as most non-citizen defendants are—are often eligible for reduced sentences as part of the Department of Justice’s Fast Track program (Cole 2012), so the extent to which prosecutors exercise charging discretion in these cases is very different from cases involving U.S. citizen-defendants. Second, the districts that border Mexico face unique caseload pressures compared to the other federal district courts. The five districts that border Mexico are: the District of Arizona, the Southern District of California, the District of New Mexico, the Southern District of Texas, and the Western District of Texas. Together, these five districts account for 34 percent of federal criminal defendants in the Commission data. It is plausible that prosecutors’ behavior towards all defendants in border districts is affected by their immigration-related workload, as in Huang (2011). Third, and consequently, because the border districts include so many immigration defendants receiving identical sentences, these defendants are

17  

total, 73.6 percent of such defendants from the Commission data are identified in the LIONS data.

After merging, the sample is reduced in several ways. The most important restriction is removing

defendants whose lead prosecutor is of unknown gender—a little more than half of defendants.44

In addition to this restriction, I also remove defendants from four federal district courts whose

courthouse identification variables appear to be unreliable45 and those whose prosecutor

identification variable appears to be unreliable;46 defendants who are coded as different genders in

the LIONS and Commission data sets (1.6 percent of defendants); and defendants who are missing

entries in any of the following variables in the Commission data: gender, race, Hispanic ethnicity,

age, educational attainment, criminal history points, offense type, sentencing fiscal year, and base

offense level.

Because I control for the year in which a case was commenced and the Guidelines version

used at sentencing, I exclude defendants who cases commenced before 1999 and after 2014

because there are few defendants in these early and late years. Because I include courthouse-

section fixed effects in the main regression analysis, I exclude courthouse-section cells that have

fewer than 25 observations and courthouse-section cells in which less than five percent of

defendants are female. Finally, I exclude offenders charged with sex offenses or offenses involving

child exploitation, a decision that is discussed in subsection 4.1.

The sample used in the main analysis includes 143,457 defendants sentenced in 65 of the

94 federal district courts in the United States, comprising 182 unique courthouses and 503 unique

courthouse-sections. Table 1 presents summary statistics of the data. Roughly 72 percent of cases

have a male lead prosecutor. Eighty-three percent of defendants are male and 70 percent have

graduated from high school. Around 40 percent of defendants are black, 12 percent are Hispanic

ethnicity, and 43 percent are non-Hispanic white. The average defendant age in the sample is 36

years old.

                                                            intrinsically hard to merge between the Commission and LIONS data sets. The results, however, are largely robust to including citizen-defendants from the border districts (results on file with the author).

44 A lead prosecutor’s gender is unknown if the salutation field in the data is empty, or if the salutation entry does not convey gender information (for example, if it is “AUSA”). Appendix A contains a complete list of how salutations are coded. Missingness in the salutation field is largely driven by missingness between—rather than within—USAOs. Among those USAOs with any gender information, most have gender information for at least 75 percent of AUSAs.

45 Data from the Northern District of Alabama, the Northern District of Iowa, the District of Nebraska, and the District of South Carolina contain many more courthouses than actually exist in those districts.

46 This restriction removes 1.8 percent of defendants.

18  

In terms of case outcomes, the vast majority of defendants—roughly 86 percent—receive

sentences of some incarceration, and the mean sentence among all defendants is 60 months.

Average sentence length is considerably lower than the mean Guidelines range, which is 84

months. This is because roughly half of defendants receive a sentence below their recommended

Guidelines range, while just two percent receive a sentence above their recommended Guidelines

range.

Appendix Table A.1 assesses the representativeness of the sample by comparing the

variable means of defendant and case characteristics presented in Table 1 to those: among all

Commission defendants in federal districts represented in the data (column (2)); among all

defendants in the merged sample (column (3)); and among the full universe of Commission

defendants (column (4)). Table A.1 allows readers to not only assess how well the data used in this

work represents the full universe (column (1) versus column (4)), but also to assess how well the

matching process worked (column (3) versus column (4)). In all columns, I restrict to U.S. citizen-

defendants and exclude districts that border Mexico so as to compare equivalent populations.

Table A.1 demonstrates that data used in this paper is extremely representative of the districts

represented in the data, the matched sample as a whole, and the universe of all Commission

defendants.

3.4. Graphical Description of the Results

The defendant’s base offense level is the primary dependent variable used in the analysis.

The base offense level is a natural number between 0 (least serious) and 43 (most serious).47 In

federal criminal cases, a defendant’s sentencing advisory Guidelines range is determined based on

two factors: their final offense level and their criminal history score. The final offense level is

computed by adjusting the base offense level to reflect the defendant’s offense conduct and their

behavior during the case. For example, under U.S. Sentencing Guideline § 3E1.1, a defendant can

earn up to a three-point reduction to their base offense level by quickly pleading guilty.48 This

paper uses the defendant’s base offense level because it reflects the severity of the charged offense

                                                            47 In the data, 22 out of the 143,454 defendants have base offense levels that are above 43 (ranging from 44-47). I

assume these are erroneous entries. 48 The three-point reduction is only available to defendants whose offense level is 16 or larger. Other defendants are

only eligible for a two-point reduction.

19  

before any adjustments occur as the case proceeds. I therefore consider it to be the variable in the

data that most closely quantifies the prosecutorial charging decision.49

Increasing by one base offense level corresponds with a roughly ten percent increase in the

Guidelines-recommended sentencing range. Figure 2 presents a histogram of the base offense level

variable in the data. Due to the way the Guidelines assign base offense levels to offenses, base

offense levels that are even numbers are more prevalent than odd numbers, which Figure 2

illustrates.

Before turning to regression estimates, Figure 3 depicts the average base offense level for

the four possible gender pairings of prosecutor and defendant relative to the average across all

defendants within the same courthouse. In Figure 3, cases prosecuted by male prosecutors are

depicted with solid columns, and cases prosecuted by female prosecutors are depicted with striped

columns. Figure 3 constitutes suggestive evidence of gender-based in-group favoritism: female

defendants earn lower base offense levels (i.e. lesser charges) when assigned a female rather than

male prosecutor, while male defendants earn a lower base offense level when assigned a male

rather than female prosecutor. When all defendants are pooled, the overall difference between male

and female prosecutors are smaller than when one compares the averages within defendant gender.

4. Empirical Strategy and Main Results

4.1. Regression Design

While Figure 3 presents suggestive evidence of gender-based in-group favoritism in

charging decisions, it is important to control for defendant and case characteristics, changes over

time, and geographic differences. For the main analysis, I obtain estimates of gender-based

favoritism using the following empirical specification:

yipct = β0 + β1MaleDi + β2MalePip + β3MaleDi*MalePip + θXi + γfc + δt + εipfct (1)

where yipct is an outcome for defendant i, assigned prosecutor p, charged with an offense of type f,

in courthouse c, in year t; MaleDi is an indicator variable that equals one if defendant i is male;

MalePip is an indicator variable that equals one if defendant i's prosecutor p is male; Xi is a vector

                                                            49 More precisely, the base offense level will capture charging severity at sentencing—which will reflect any

bargaining between prosecutor and defendant. The results presented in the main analysis in section 4.3, however, are robust to alternative measures of charging severity. These results are presented in Appendix Table A.1.

20  

of observable characteristics of defendant i and their case; and γfc and δt are courthouse-section

type and time fixed effects, respectively.

In this formulation, β1 captures average differences in outcomes for female and male

defendants when the prosecutor is female, β2 captures average differences in outcomes for female

and male prosecutors when the defendant is female, and β3 captures in-group favoritism: average

differences in how female and male prosecutors treat female defendants relative to male

defendants.50 In other words, Equation (1) is a difference-in-differences design51 in which, holding

all else equal, the predicted differences in base offense level for each defendant-prosecutor gender

combination relative to female-female pairs are:

Male Prosecutor Female Prosecutor Difference (M-F) Male Defendant β1+β2+β3 β1 β2+ β3 Female Defendant β2 - β2 Difference (M-F) β1+ β3 β1 β3

4.2. Estimation Challenges

A central challenge in estimating equation (1) is that whether a prosecutor and defendant

match on gender might be correlated with unobservable defendant or case characteristics that also

affect case outcomes, even after controlling for many observable defendant and case characteristics

and relevant fixed effects. In this case, a regression of equation (1) will produce biased coefficient

estimates.

There are several ways in which the assignment of cases to prosecutors might be non-

random in a way that biases the coefficient estimate of β3. Most obviously, the estimate of β3 will

be biased upward (downward) if supervisors are likely to assign cases with less serious (more

serious) defendants to prosecutors who share the defendant’s gender.

There are a few reasons why it is unlikely that cases are assigned in this way. For one thing,

while it might be possible for a supervisor to consider characteristics of the defendant in assigning

cases involving a solo defendant to prosecutors that share the defendant’s gender, this task

becomes unrealistically complicated when there are co-defendants of different genders. In the data,

however, more than half of all female defendants have at least one male co-defendant but all

defendants within a case are assigned the same lead prosecutor. In subsection 6.1, I show that the

                                                            50 The paper uses ordinary least squares regression throughout for interpretive clarity. All regressions in which the

outcome variable is binary are robust to using logit regression. 51 This empirical specification follows Shayo and Zussman (2011).

21  

estimate of in-group favoritism is largely robust to restricting the sample to cases in which there

is at least one male and at least one female defendant, a restriction that removes more than 80

percent of the data.

For another thing, it is plausible that the factors that influence an AUSA’s charging

decision are not known by a supervisor at the time of case assignment. If supervisors do not have

a nuanced picture of a defendant’s social history and specific offense conduct at the time case

assignments are made, it would be difficult for such assignments to be made strategically. For this

reason, I rely on the first prosecutor assigned to a case under the assumption that the first attorney

assignment is made when a supervisor has the least information. Prosecutors who join a case later

might be more strategically assigned, but they are not part of the main analysis.52 In subsection

5.3, I define the prosecutor gender variable based on the gender composition of all lead prosecutors

assigned to a case, and find that the results are very similar to those in the main analysis.

A second concern is that female prosecutors might be disproportionately represented in

areas of federal criminal law in which women are treated more leniently. Courthouse-section fixed

effects in the regression equations should largely address this concern, which seems to materialize

in one setting: cases involving sexual offenses and child exploitation. These cases exhibit above-

average gender disparities, involve very few female defendants, and are more likely than other

types of cases to be staffed by female prosecutors. Therefore, I exclude this offense category from

the main analysis, although the results are robust to including these cases. I do not find evidence

of gender disproportionality in the share of prosecutors in any other offense type. Among the four

offense types remaining in the data after sex offenses are removed, the percentage of lead

prosecutors that are female ranges from 28 to 30 percent. In contrast, among sex offense cases,

roughly 44 percent of lead prosecutors are female.

I also look for quantitative evidence to shed light on whether prosecutors are assigned

systematically different defendants by gender. Table 2 presents results of regressing the

prosecutor’s gender on defendant and case characteristics that are fixed at the time of case

including: the defendant’s age, number of dependents, criminal history category, the prosecutor’s

                                                            52 In the data, 99.8 percent of cases have just one prosecutor assigned on the first day. For the few cases in which

more than one prosecutor is assigned to a case on the first day, I code the prosecutor as female if all prosecutors assigned on the first day are female, male if all prosecutors assigned on the first day are male, and remove the case from the data if both male and female prosecutors are assigned on the first day. In subsection 5.3 below, I examine the behavior of same-gender and mixed-gender teams of prosecutors.

22  

workload, and indicators for the defendant being female, black, Hispanic, another race other than

white, a high school graduate, a college graduate, whether the case involved more than one

defendant, whether the case involved more than one prosecutor, and the type of offense.53 Columns

(1) through (3) present results in the whole sample, while column (4) restricts attention to female

defendants and column (5) restricts attention to male defendants.

In columns (1) and (2), prosecutor gender is regressed on defendant and case characteristics

without any geographic controls. In these regressions, male prosecutors are associated with an

increased likelihood of prosecuting male defendants and a decreased likelihood of prosecuting

defendants of color. Defendants’ educational attainment and criminal history are not statistically

different for male and female prosecutors, although male prosecutors are assigned defendants that

are statistically, but not meaningfully, younger than those prosecuted by female prosecutors when

no time controls are included. The coefficient estimates across columns (1) and (2) are also very

similar—suggesting that there are not large differences in caseload composition over time that are

correlated with prosecutor gender. Once courthouse-section fixed effects are included in column

(3), male and female prosecutors no longer appear to differ in the racial compositions of their

caseloads. This change demonstrates the crucial role of courthouse-section fixed effects—it

suggests that female prosecutors are more prevalent in geographic regions and offense types in

which non-white defendants are more prevalent.

Even after including courthouse-section fixed effects in column (3), the coefficient

estimate on defendant gender remains positive and highly significant. The point estimate—

0.022—suggests that a male defendant’s probability of being assigned a male prosecutor is about

2.2 percentage points larger than a female defendant’s probability of being assigned a male

prosecutor. Thus, it appears that supervisors might be influenced (consciously or not) by gender

in making case assignments, which is not surprising. Researchers studying the assignment of law

enforcement officers to neighborhoods similarly report that officers are more likely to be assigned

to areas in which the majority racial group is the officer’s racial group (Antonovics and Knight

2009, Donohue and Levitt 2001).

Given that defendants are slightly but statistically significantly more likely to be assigned

a prosecutor of the same gender, it is crucial to check whether prosecutor gender seems to be

                                                            53 The offense type categories are: drug offenses, fraud and other white-collar offenses, regulatory offenses, and

violent offenses.

23  

correlated with any defendant and case characteristics that are fixed at the time of the case,

conditional on defendant gender. Table 2, column (4) reports results of regressing prosecutor

gender on all the covariates listed above among male defendants, while column (5) reports results

among female defendants. The point estimates presented in columns (4) and (5) are all close to

zero and all but one are statistically insignificant, suggesting that female and male prosecutors see

similar mixes of defendants within each gender group.

It is also important to remember that USAOs and sections within USAOs likely vary in

their assignment procedures. In many courthouse-sections, prosecutor gender is uncorrelated with

defendant gender. Subsection 6.2 restricts attention to courthouse-sections cells in which

prosecutor gender is not significantly related to any defendant characteristics that are fixed at the

time of the case—including gender—and finds that the results are quite similar (if not slightly

stronger) than those in the main analysis, offering more evidence that the results are not driven by

differences in prosecutors’ caseloads.

4.3. Initial Results

As described in the previous section, the base offense level is the primary outcome variable

used in the analysis. Table 3 presents results of estimating equation (1) with ordinary least squares

regression. Standard errors are clustered at the courthouse level to account for likely intra-

courthouse correlation in the error term (Bertrand, Duflo, and Mallainathan 2004).

The coefficient on the interaction term—in the top row of Table 3—represents gender-

based in-group favoritism. The regression with no controls—presented in column (1)—appears to

overstate gender-based favoritism. In the regression with demographic and case controls, year

fixed effects, and courthouse-section fixed effects presented in column (4), the coefficient estimate

of the interaction term is -0.542 and is statistically significant at the one-percent level. This result

can be interpreted as the differential effect of prosecutor gender on male defendants relative to

female defendants. Column (3) uses district-section fixed effects instead of courthouse-section

fixed effects and finds results that are extremely similar to those reported in column (4), suggesting

the additional geographic information contained in the courthouse variable is not critical. The

results in column (5) shows that using courthouse-section-year fixed effects produces a stronger

coefficient estimate on the interaction term, although this regression excludes roughly twenty

24  

percent of the data from the full sample and is less precisely estimated.54 To maximize statistical

power, the remaining specifications include courthouse-section and year fixed effects separately,

as in column (4).

Consistent with in-group favoritism, the second row of Table 3 demonstrates that

prosecutor gender is a significant predictor of charging severity when the defendant is female

(columns (1)-(5)). However, in column (6), which does not include the interaction term, the

coefficient on prosecutor gender is quite small and is statistically insignificant—indicating that

prosecutor gender does not influence prosecutorial behavior on average.

The third row of results in Table 3 indicates that male defendants are subject to significantly

larger base offense levels than female defendants when the prosecutor is female, even after the

inclusion of many control variables and fixed effects.55 After controlling for demographic and case

characteristics, year fixed effects, and courthouse-section fixed effects, the coefficient estimates

are around a two-point increase in base offense level for male defendants in columns (3) through

(5). Column (6), which does not include a prosecutor-defendant interaction term (and therefore

represents average effects) estimates a 1.6-point gender gap in base offense level after controlling

for case and demographic characteristics.

Appendix Table A.2 reports similar results when equation (1) is estimated with alternative

measures of charging severity.56 The first such alternative is the defendant’s statutory minimum

across all counts of conviction. While the statutory minimum reflects charging severity untethered

from judicial interference, it is a coarse measure. Roughly two-thirds of defendants have no

statutory minimum, and some of these defendants are still sentenced to life in prison (suggesting

that they committed very serious offenses despite not facing a mandatory minimum). The second

alternative measure of charging severity is the leave-one-out average sentence for all defendants

                                                            54 As described in the notes to Table 3, the regression that included courthouse-section-year fixed effects removed

courthouse-section-year cells with fewer than 25 observations, reducing the sample size from 143,457 observations to 103,858 observations.

55 Because the regression equation includes an interaction term, the coefficients for the male defendant and male prosecutor variables must be interpreted as conditional effects rather than as main effects in columns (1) through (5) in Table 3.

56 Some might object that the base offense level is an imperfect measure of charging severity because it can include adjustments for facts that are found by the court. This concern is largely mitigated by the fact that the paper uses the value of the base offense level before Chapter Two adjustments for specific offense characteristics. The Commission data reports both the base offense level (inclusive of Chapter Two adjustments) and the Chapter Two adjustments themselves, so it is straightforward to measure the base offense level prior to Chapter Two adjustments, which is the outcome variable used in Table 3.

25  

within the same courthouse who were sentenced under the same Guideline as the defendant,

excluding the defendant themselves. This is likely a noisy measure of charging severity given that

sentences are highly dependent on other factors (such as criminal history). Both of these alternative

measures of charging severity, however, produce estimates of gender-based favoritism that are

statistically significant.

4.4. Additional Outcome Variables

The prior analysis uses the defendant’s base offense level as the outcome of interest. While

the base offense level plausibly captures charging severity, it is not a case outcome—it is simply

an input into a defendant’s ultimate sentence. Thus, it is useful to think about additional outcome

variables. I organize additional outcome variables into three general categories. The first are

outcomes that are largely a function of prosecutorial discretion. As explained above, I believe the

defendant’s base offense level almost purely captures a prosecutorial decision.57 Other variables

of this nature include: whether the defendant was charged with a crime carrying a mandatory

minimum penalty and whether the government recommended a reduction to the defendant’s

offense level. The second type of outcome variable is one that is a function of both judicial and

prosecutorial discretion. The defendant’s ultimate sentence is an example of such an outcome: a

prosecutor can influence the sentence through the charging decision, the application of

enhancements and reductions, and their own recommendation, but, ultimately, the district judge

sentences the defendant. Finally, there are outcome variables over which the prosecutor has

relatively little control. For example, judges largely determine whether the defendant receives a

sentence below the recommended Guidelines range.58

                                                            57 To be more precise, the base offense level will also reflect any bargaining over the charges between the defendant

and the prosecutor. Section 5.4 further discusses the bargaining dynamics between prosecutors and defendants. The Justice Manual (formerly the U.S. Attorneys Manual), however, attempts to limit charge-bargaining by instructing prosecutors to “charge and pursue the most serious, readily provable offenses,” which are defined as “those that carry the most substantial guidelines sentence, including mandatory minimum sentences.” Justice Manual 9-27.300. The Manual does not completely foreclose charge bargaining, making clear that “there will be circumstances in which good judgment would lead a prosecutor to conclude that a strict application of the above charging policy is not warranted.” Id. In such a situation, the individual AUSA must get approval from a supervisor to bring a lesser charge. Id.

58 Of course, a prosecutor can request an out-of-range sentence, but such requests are rare in the data. Prosecutors request a below-Guidelines sentence for roughly four percent of defendants, while 46 percent of defendants ultimately receive sentences below the recommended Guidelines range. Upward departures requested by the prosecutor are not reported in the data, but only two percent of defendants receive sentences above their recommended Guidelines range, so it is likely that such requests by prosecutors are similarly rare.

26  

Table 4 presents regression results estimating equation (1) with eight additional outcome

variables. Panel A considers early outcomes—variables capturing decisions that occur while a case

is going on and over which the prosecutor exercises large influence. These early outcomes include:

(1) the final offense level; (2) whether the defendant was charged with a mandatory minimum; (3)

the defendant’s final Guidelines Range (represented by its mean, transformed by the inverse

hyperbolic sine); and (4) whether the defendant received a substantial assistance reduction. Panel

B considers outcomes over which the prosecutor’s influence is more limited: (5) whether the

defendant was released pending trial; (6) the defendant’s sentence in months (transformed by its

inverse hyperbolic sine); (7) same as (6) but with the base offense level added as a control variable;

and (8) whether the defendant received a sentence below their recommended Guidelines range. If

gender-based favoritism shapes prosecutorial decision-making, one would expect coefficient

estimates of the interaction term to be significant for outcome variables that are under the

prosecutor’s control, but insignificant for outcome variables that are largely determined by the

defendant’s judge.

As predicted, the results in Panel A suggests that gender match—represented by the

coefficient on the interaction term—is associated with a reduction in the defendant’s final offense

level, a reduction in the probability that the prosecutor charges a mandatory minimum offense, and

a roughly six-percent reduction in the defendant’s Guidelines range. Defendant-prosecutor gender

match does not appear to affect the probability that a defendant receives a substantial assistance

reduction. This finding suggests that favoritism is concentrated at the beginning of the case,

perhaps because there is less involvement from other actors, like defense counsel, or because they

are simply because early decisions are made with less information about the defendant.59

Panel B, which considers outcomes in which a judge exercises decision-making authority,

finds that the prosecutor-defendant gender interaction match plays a smaller role. Sentence length

reported in column (5)—which is a function of both prosecutorial and judicial discretion—is the

only outcome that appears to be influenced by gender match. The coefficient estimate on the

interaction term, -0.102, suggests that female defendants can expect to receive a ten percent shorter

                                                            59 For example, Marianne Bertrand and Esther Duflo explain, “Implicit biases are more likely to drive behavior

under conditions of ambiguity, high time pressures and cognitive loads, or inattentiveness to the task.” (Bertrand and Duflo 2017).

27  

sentence relative to male defendants when they are assigned a female prosecutor. These findings

suggest that judicial decision-making does not entirely mitigate in-group favoritism in charging.60

This sentencing reduction appears to enter through the charging channel: when the defendant’s

base offense level is included as a control variable in column (7), the point estimate is attenuated

and not statistically significant. Also as predicted, estimates of gender-based favoritism are

statistically insignificant predictors of whether the defendant received a sentence below the

recommended Guidelines range—an outcome that is largely within the judge’s purview. Nor is

gender match a statistically significant predictor of whether a defendant is released pending trial—

a decision that the prosecutor might influence, but that is ultimately decided by a judge.

Finally, it is worth also highlighting that being a male defendant is significantly predictive

of worse outcomes on every measure, consistent with the findings in Starr (2015) of large gender

disparities at all stages of a criminal case. Specifically, relative to female defendants, a male

defendant suffers: an increased probability of being charged with a mandatory minimum, a reduced

probability of receiving a substantial assistance reduction, a reduced probability of being released

pending trial, a longer sentence, an increased probability of receiving at least the mandatory

minimum when one is charged, and a reduced probability of being sentenced below the

recommended Guidelines range. Moreover, the estimated gender gap in sentence length reported

in Table 4 is extremely close in magnitude to the estimate in Starr (2015).61

4.5. Selection into the Sample

Because this paper uses data on defendants sentenced in federal district courts, it necessarily

excludes defendants who are not ultimately convicted. In this subsection, I present suggestive

evidence that it is unlikely that sample selection biases the results of the paper. All quantitative

work on the criminal system must grapple with the selected sample problem: researchers typically

only observe defendants who formally enter into the criminal system (or, in the case of many data

sources, are eventually sentenced). It is plausible that selection into the sample could bias the

                                                            60 This explanation is consistent with Joshua Fischman and Max Schanzenbach’s finding that judicial discretion

does not contribute to racial disparities and might mitigate racial disparity (Fischman and Schanzenbach 2012). 61 The coefficient on defendant gender in the regression using the inverse hyperbolic since of sentence length as a

dependent variable is 0.848, which is equivalent to a roughly 57 percent increase in sentence length for male defendants relative to female defendants. This point estimate is very close to the finding in Starr (2015) of a roughly 60 percent gender gap in sentence length in federal criminal cases.

28  

results if there is in-group favoritism on the margin of selection.62 On the other hand, one might

not expect selection into the sample to pose a serious threat in the federal setting because nearly

all federal felony defendants—around 94 percent—are ultimately convicted. Only 0.4 percent are

acquitted after trial. The remaining 5.6 percent of cases that do not result in conviction are those

voluntarily dismissed by the government (Motivans 2017).

I am able to make progress on the sample selection issue because the LIONS data includes

defendants whose charges the government declines to prosecute, voluntarily dismisses, or of which

the defendant is acquitted. Therefore, I can investigate whether there appears to be gender-based

favoritism on the margin of whether a defendant is convicted. This subsection takes on this task

by using the LIONS data on its own, without matching it to the Commission data (which only

includes sentenced defendants).

Using the LIONS data alone has several drawbacks, which is why it is not the default for

the rest of this paper. First, the LIONS data does not include any of the outcome variables used in

this paper except the defendant’s sentence length. Second, the LIONS data lacks most of the

control variables used in the paper, such as the defendant’s race and ethnicity, level of education,

age, and criminal history. And although LIONS has a variable that indicates the defendant’s

gender, it contains many missing values. Third, the LIONS data does not appear to be quality-

checked for completeness and accuracy the way the Commission data has been. Despite these

shortcomings, I verify the LIONS-only sample’s reliability, as described below.

To create the LIONS-only subsample I apply the same restrictions to the data that I used

when constructing the sample used in the paper, as described in subsection 3.3. The LIONS-only

sample contains 89,117 defendants, of whom 83,790 are sentenced. This translates into a 6.0

percent charged-but-not-convicted rate, which is very similar to the Bureau of Justice Statistics

reporting that in 2014, 6.4 percent of federal felony defendants were not ultimately convicted

(Motivans 2017). Appendix Table A.3 presents the results similar to those in Table 2 by regressing

prosecutor gender on defendant and case characteristics. In the LIONS-only sample, however,

there are many fewer control variables than in Table 2. Even though the LIONS-only sample

contains many fewer observations and control variables, it produces results in Appendix Table A.3

                                                            62 In particular, if prosecutors display gender-based favoritism in the decision to decline or dismiss cases, the paper’s

estimates of gender favoritism would understate the phenomenon. On the other hand, if there is in-group disfavoritism in the decision to decline or dismiss cases, the paper’s estimates could overstate the magnitude.

29  

that are very similar in magnitude and statistical significant to those produced in the full data.

While male prosecutors are slightly (around two percentage points) but statistically significantly

more likely to be assigned to cases with male defendants, the case characteristics that are available

in the LIONS data—whether the case involved multiple defendants, whether the case involved

multiple prosecutors, and the offense type—are not statistically different for male relative to

female prosecutors.

Table 5 looks for in-group favoritism on the margin of whether a person is ultimately

sentenced (i.e., enters the data used in the paper). The outcome variable is a binary variable that

equals one if a defendant’s charges were acquitted, declined or dismissed and zero if the defendant

was convicted and sentenced. The coefficient estimates on the interaction term are statistically

insignificant in all specifications. The point estimates on the interaction term, although statistically

insignificant, are all greater than zero, which suggests that if anything, in-group favoritism might

exist on the margin of selection into the sample, which would likely attenuate (rather than bolster)

the findings of in-group favoritism later in the case that are documented in this section.

5. Heterogeneity and the Sources of Gender-Based Favoritism

The previous section demonstrates that prosecutors appear to charge defendants more

leniently when the defendant is the same gender as the prosecutor. In theory, there are many

underlying mechanisms that could generate this finding. This section broadly considers two. First,

gender-based leniency might derive from social preferences, or, taste-based discrimination

(Becker 1957). I call this the preference-based explanation for the findings. Second, the

prosecutorial process might generate gender-based leniency. Prosecutors might be able to more

(or less) effectively prosecute defendants with whom they have more in common. The paper refers

to mechanisms in this category as process-based explanations for the findings. For example, if

prosecutors are better able to evaluate the risk of recidivism among defendants of their gender than

opposite-gender defendants, differential treatment could constitute statistical discrimination, along

the lines of Cornell and Welch (1996). Conversely, defendant response to their prosecutor’s gender

could drive the results if, for example, defendants are more likely to trust and therefore cooperate

with an own-gender prosecutor.

This section presents evidence of heterogeneity in the results. Heterogeneity is not only

descriptively interesting in its own right, it also has the potential to shed light on the mechanisms

30  

responsible for in-group favoritism, which is the focus of this section. In particular, this section

evaluates whether preference- or process-based mechanisms are likely to be responsible for

gender-based favoritism by examining how, if at all, the results are heterogeneous across many

dimensions, including geography and sexism (subsection 5.1); offense type (subsection 5.2); and

prosecutorial team composition (subsection 5.3). These subsample comparisons help discern

possible explanations for the gender-based favoritism documented in Section 4. Subsection 5.4

further probes the process-based explanation by examining whether favoritism affects the

defendant’s level of cooperation, operationalized in several different ways.

Broadly speaking, I find substantially more support for the preference-based explanation

of gender-based favoritism. As described in more detail below, the preference-based explanation

is bolstered by findings that in-group favoritism is stronger in states with below-median measures

of sexist attitudes (compared to states with above-median sexism), and in cases in which gender is

more salient. I also find that favoritism is stronger in same-gender prosecutorial teams (compared

to mixed-gender teams or solo prosecutors), but I argue that depending on the nature of group

decision-making, this result could be consistent with either preference- or process-based

explanations (or both). I fail, however, to find any concrete evidence to support the process-based

explanation: defendant-prosecutor gender match is not a significant predictor of whether the

defendant receives a substantial assistance reduction, a safety valve reduction, an acceptance of

responsibility reduction, or takes their case to trial.

5.1. Geography, Sexism, and Gender-Based Gender Favoritism

Gender-based favoritism might vary across geographic regions. In the United States, there

is significant regional variation in the prevalence of sexist attitudes. It is plausible that a locality’s

level of background sexism could affect the extent to which prosecutors demonstrate gender-based

leniency, particularly if such favoritism is preference-based. On the other hand, because criminal

processes are similar across geographic regions within the federal criminal system, one might not

expect gender-based favoritism to vary across geographic regions if such favoritism is purely

process-based.

A new working paper by Kerwin Kofi Charles, Jonathan Guryan, and Jessica Pan quantifies

regional variation in sexism across the United States, and I exploit this state-level variation to

explore the possibility that gender-based leniency in charging is related to the intensity of sexist

31  

attitudes in the locality.63 As a preliminary matter, Figure 5 depicts district-level estimates of in-

group favoritism by estimating equation (1) separately for each federal district court in the sample.

Darker shading indicates that the federal district court demonstrates stronger gender-based

favoritism. Figure 5 presents evidence that gender-based favoritism is geographically

heterogeneous, but demonstrates no clear geographic pattern.

I next examine whether the role of prosecutor-defendant gender match is moderated by the

background level of sexism in the state in which the district is situated. It is worth emphasizing

that sexism can manifest in many different ways. For example, sexism can be either conscious or

implicit. No matter its form, however, one threshold prediction—which I verify below—is that

female defendants will be treated more harshly relative to male defendants in states with higher

levels of sexism than in states with less prevalent sexism.

In Table 6, I divide states in the sample by whether they exhibit either above- or below-

median sexism according to Charles, Guryan, and Pan (2018).64 Table 6 examines how the

interactions of defendant gender, prosecutor gender, and local sexism relate to charging decisions.

As above, the defendant’s base offense level is the outcome variable. As a preliminary matter,

women appear to be charged more harshly relative to men in states with above-median sexism

(row 5 in column (8)), providing validation that the sexism variable captures negative attitudes

toward women.

The results in Table 6 also present several important clues about the results presented in

Section 4. First, as columns (1) and (2) demonstrate, gender-based favoritism in the data is largely

driven by differential treatment of female defendants. For female defendants, the increased

charging severity associated with having a male prosecutor (0.398 in the base offense level) is

nearly eight times larger than the decreased charging severity for male defendants assigned a male

prosecutor (-0.053). In other words, female defendants earn a more important benefit from being

assigned a gender-matching prosecutor than male defendants do.

This finding could be explained by the fact that gender is likely to be more salient for female

than male defendants. Around 83 percent of criminal defendants are men, so a defendant’s gender

is likely to be more noteworthy—and thus more salient—when the defendant is female. Typically,

                                                            63 Charles, Guryan, and Pan (2018) are interested in the labor market and familial status effects of sexism. They find

that background sexism affects women’s wages, labor force participation, marriage age, and childbearing. 64 Six states and territories—Washington DC, Guam, Hawaii, Idaho, Nevada, and Puerto Rico—are removed from

the sample because Charles, Guryan, and Pan (2019) do not report sexism scores for these states and territories.

32  

in-group preferences are more pronounced for in-groups that have higher salience. (Everett, Faber,

and Crockett 2015; McLeish and Oxoby 2011).

Second, although matching with one’s prosecutor on gender is important for female

defendants, it is significantly more important in states with below-median sexism scores than in

states with more prevalent sexism (columns (3)-(8)). This suggests that in-group favoritism might

work against, rather than in tandem with, sexism. Overall, the results seem to reject the idea that

sexism cultivates in-group favoritism. Instead, the results are consistent with gender favoritism

that is generated by positive feelings toward women.

It is plausible that inter-group empathy bias is responsible for the findings. As others have

shown, people are more likely to feel empathy towards in-group members (Cikara et. al 2011).

And, perhaps most relevant to the prosecutorial setting,65 in experimental settings, people are more

likely to help in-group than out-group members (Everett, Faber, and Crockett 2015), and to feel

pain and empathy when observing the pain of an in-group member compared to an out-group

member (Xu et al. 2009, Gutsell and Inzlicht 2012).66 Others have found that people with higher

implicit bias tend to have reduced empathy in response to out-group pain (Avenanti, Sirigu, &

Algiuoti 2010). If sexism reduces in-group favoritism, it might break these empathies.

5.2. Heterogeneity by Offense Type

This subsection considers heterogeneity by offense type. Cases in the data are categorized

into four discrete offense types: drug offenses, regulatory and other offenses, violent offenses, and

white collar offenses. A little over half of offenders are prosecuted for drug-related offenses. If

gender-based favoritism is preference-based, I predict that it will be the most intense among

offense types for which gender is most salient.

Table 7 presents results of estimating equation (1) by offense type. While the estimates of

in-group favoritism are statistically significant at the ten-percent level for all offense types, they

are the strongest among violent offenders and, to a lesser extent, drug offenders, and weaker for

regulatory and white collar offenders. These findings are consistent with in-group favoritism being

                                                            65 Prosecutors are both witnesses to and agents of the pain felt and costs imposed on criminal defendants when they

are prosecuted, convicted, and ultimately sentenced. 66 While people also experience the opposite—pleasure in response to an out-group member’s adversity (a

phenomenon called schadenfreude)—a review of the literature concluded that “positive social preferences of in-group love may play a stronger role than negative social preferences for outgroup derogation.” (Everett, Faber, and Crockett 2015).

33  

strongest in offense types in which gender is more salient. Among violent offenders—for which

in-group favoritism is the strongest—only six percent of offenders are women. In contrast, among

the roughly equally-sized population of white collar offenders—for which in-group favoritism is

the weakest—roughly one-third of defendants are female. In fact, gender might be especially

salient for women in violent crime cases not only because female defendants are rare in such

prosecutions, but also due to prevailing social norms that women do not engage in violent behavior

(Russell 2014). These findings are consistent with the findings above that salience appears to

impact the intensity of gender-based favoritism exhibited by prosecutors.

The results in Table 7 are also notable because they suggest that gender favoritism is

heightened among more serious offenders, presenting a conflict with prior work that examines the

labor market concerns of federal prosecutors. As described in section 2, Boylan and Long (2005)

and Boylan (2005) find that better career outcomes in the private legal market are associated with

more aggressive tactics by federal prosecutors—longer prison sentences and increased

probabilities of taking a case to trial. Much labor literature suggests that male prosecutors are more

likely to benefit from these future, private-sector rewards than female prosecutors (Goldin 2014,

Wiswall and Zafar 2018). Yet, male prosecutors appear to show leniency in cases involving male

defendants—who are, on average, significantly more serious offenders. This unexpected finding

is consistent with a preference-based explanation for leniency by male prosecutors toward male

defendants because such leniency seems to work against what one would predict would be in

prosecutors’ self-interest.

I also consider two alternative potential explanations for these findings. First, if the

relationship between gender match and charging severity is concave rather than linear, one would

expect stronger results in offenses with higher average base offense levels, which might explain

why the levels of in-group favoritism are larger for violent and drug crimes. This explanation does

not entirely explain the results. Logging the base offense level variable still produces estimates of

in-group favoritism that are the larger among violent offenders than other offense types. Moreover,

drug offenders are charged with offenses carrying the highest average base offense levels, but

violent offenders generate the largest estimates of in-group favoritism.

A second possible explanation is that in-group favoritism could be more intense for

defendants of color, and these defendants tend to be over-represented in violent and drug cases

relative to regulatory and white collar cases. It is possible, of course, that the extent to which

34  

prosecutors behave with gender-based favoritism could vary depending on the defendant’s race

and ethnicity. For example, Starr (2015) finds that gender disparity in federal criminal cases is

significantly larger among black than white defendants.

To examine this possibility more directly, Appendix Table A.4 presents results of

estimating equation (1) by defendant race and ethnicity. The coefficient estimates of in-group

favoritism are statistically significant at conventional levels in all racial groups, although the point

estimates are more extreme in the subsamples including minority defendants. However, three-way

interactions between prosecutor gender, defendant gender, and race or ethnicity are not statistically

significant. I therefore cannot rule out the possibility of equal gender-based favoritism across racial

groups. Appendix Tables A.5 and A.6 present results by race and ethnicity further subdivided by

offense type (Appendix Table A.5) and for all additional outcomes (Appendix Table A.6). Table

A.5 provides more support for the hypothesis that gender-based favoritism is strongest in cases in

which gender is most salient and that this pattern persists across black, Hispanic, and white

defendants.

5.3. Prosecution in Teams

So far the analysis has only considered the interaction between one defendant and one

prosecutor—the first lead prosecutor assigned to each case. However, around one-third of the cases

in the data involve more than one lead prosecutor. This subsection compares levels of gender-

based favoritism between cases prosecuted by solo prosecutors, mixed-gender prosecutorial teams,

and same-gender prosecutorial teams. In this subsection, I first test whether prosecutors exhibit

these kinds of peer effects and, after finding that they do, attempt to untangle what this finding

suggests about the mechanisms responsible for gender-based leniency.

To carry out this analysis, I compare estimates of in-group favoritism in three kinds of

cases: (a) those prosecuted by solo prosecutors (62 percent of observations); (b) those prosecuted

by same-gender teams of prosecutors (22 percent of observations); and (c) those prosecuted by

mixed-gender teams of prosecutors (16 percent of observations).67 Although I use the word team

to refer to the situation in which a case in the data includes more than one lead prosecutor, it is

likely that at least some of these cases involve instances of reassignment, such as if an AUSA

                                                            67 While 68 percent of cases have a solo prosecutor, only 62 percent of observations in the sample have a solo

prosecutor because some cases include multiple defendants who appear in the data as separate observations.

35  

resigns from their job or takes temporary leave from work. In other words, the extent to which

these arrangements reflect group decision-making likely varies and, if anything, the results

probably understate peer effects.

When more than one prosecutor is ultimately assigned to a case, the prosecutor gender

variable can be defined in several different ways. The preceding analysis defined the prosecutor

gender variable for each case as a binary variable indicating whether the first prosecutor assigned

to the case was male. A second way to define the prosecutorial gender of a mixed-gender team is

to use the fraction of male prosecutors assigned to the case.68 A third way to define the

prosecutorial gender of a team is to define it as a binary variable that indicates whether there is at

least one female prosecutor on the team. This subsection explores in-group favoritism in

prosecutorial teams using all three formulations.

Table 8 presents results of estimating equation (1) among defendants prosecuted by solo

prosecutors, those prosecuted by a same-gender team of prosecutors, and those prosecuted by a

mixed-gender team of prosecutors. For reference, column (1) reports results from the full sample—

it is identical to column (4) in Table 3. The results in Table 8 suggest that peer effects influence

charging behavior. The estimates among same-gender teams of prosecutors (reported in column

(3)) represent the effect of switching from an all-female team of prosecutors to an all-male team.

The estimate of in-group favoritism (-1.136 in the base offense level) is statistically significant,

and nearly three times the magnitude of the estimate among solo prosecutors in column (2)

(-0.442). Column (4) represents the effect of switching from an any-female team to an all-male

team, and suggests that the presence of just one female prosecutor in the group matters (the point

estimate is -0.743 and statistically significant), but does not generate as much favoritism as moving

to an all-female team.

In contrast, the estimates of in-group favoritism among mixed-gender teams of prosecutors

(reported in columns (5) and (6)) are not statistically significant when defining the team’s gender

composition as the gender of the first prosecutor (column (5)) or the fraction of male prosecutors

(column (6)). In both columns (5) and (6), the point estimates are negative and the subsample

                                                            68 Appendix Table A.7 presents results in which prosecutor gender is defined as the fraction of prosecutors that are

male for cases in which there is a mixed-gender team. As in Table 4, gender match signficantly predicts all early outcome variables except the substantial assistance reduction. Among sentencing outcomes, only the estimate of in-group favoritism on sentence length is statistically significant when the regression does not control for charging severeity.

36  

includes many fewer observations than the full sample, suggesting that even these mixed-gender

teams might not completely mitigate gender-based favoritism. Overall, however, the results are

consistent with the peer effects hypothesis that mixed-gender teams moderate the effects of in-

group favoritism, while same-gender teams heighten it.

Without more information about the group dynamics of prosecutorial teams, peer effects

are consistent with both the preference- and process-based explanations. For example, if

prosecutorial teams reach decisions by simply aggregating the preferences of the teams’ individual

members one would expect to find evidence of peer effects in gender-based favoritism. If gender-

based leniency is preference-based, one might also expect same-gender teams to exhibit enhanced

leniency towards own-gender defendants, akin to the now well-established finding that

ideologically homogeneous panels of federal appellate judges exhibit stronger ideological

preferences in group decision-making than heterogeneous panels (Kim 2009; Sunstein, Schkade

and Ellman 2004).69

But peer effects could also be consistent with process-based explanations if one

conceptualizes group decision-making by prosecutors as highly deliberative. If prosecutors are

more skilled at prosecuting defendants with whom they have more in common, other prosecutors

in the group might defer to the prosecutor who has the most commonality with the defendant,

which would also generate results as in Table 8.

5.4. In-Group Favoritism, Bargaining, and Cooperation

Evidence contained in Table 4 and described in subsection 4.4 suggests that gender-based

leniency is concentrated at charging, which itself challenges the process-based hypothesis. This

subsection explores the process-based hypothesis more closely by examining whether prosecutor-

defendant gender match is related to defendant cooperation or other aspects of the bargaining

process.

Table 9 presents results of estimating equation (1) for seven outcome variables that are

meant to capture the extent to which the defendant places trust in and cooperates with their

prosecutor. Column (1) finds that gender match is not related to the probability that the case

involves a superseding indictment or information—a variable that plausibly captures the intensity

                                                            69 Judges also appear to be influenced by their panel colleagues’ genders in deciding cases in which gender is salient.

(Boyd, Epstein, and Martin 2010; Farhang and Wawro 2004; Peresie 2005).

37  

of charge-bargaining. The coefficient on the interaction term is near-zero (0.008) and not

statistically significant.

Columns (2) and (3) reconsider the substantial assistance reduction. As described in section

4.4, a prosecutor can request a sentencing reduction for a defendant who provides substantial

assistance to the government.70 Column (2) (reproduced from column (4) in Table 4) shows that

across the whole defendant population, gender match between prosecutor and defendant is not a

statistically or economically significant predictor of whether the defendant receives a substantial

assistance reduction.

The substantial assistance reduction, however, is not equally important for all defendants.

Column (3) restricts attention to defendants for whom the substantial assistance reduction is

especially valuable—defendants facing a mandatory minimum who are ineligible for safety-valve

relief.71 For these defendants, a substantial assistance is especially beneficial because it is the only

way that they can receive a sentence below their mandatory minimum. Defendants in this group

comprise roughly 23 percent of the sample, and gender match is a marginally significant predictor

of whether the defendant receives a substantial assistance reduction among these defendants

(p=0.095).

This result is consistent with both preference-based and process-based explanations. If the

likelihood of receiving a substantial assistance reduction is driven by the defendant’s willingness

to cooperate with the government, significance on the match term suggests that either defendants

feel more comfortable cooperating with an own-gender prosecutor or prosecutors are better at

eliciting cooperation from like-gender defendants (or both). These are process-based explanations

for the results. On the other hand, if the substantial assistance is driven by prosecutorial discretion,

which a prosecutor may use at their will to benefit some defendants but not others, the results

would suggest a preference-based explanation.

Column (4) estimates whether gender match is related to whether the defendant receives a

safety valve reduction. Critically, to earn a safety valve reduction, a defendant must make a truthful

proffer to the government of “all information and evidence the defendant has concerning the

offense[s]”72—an element that requires the defendant to place significant trust in their prosecutor.

                                                            70 See note 26. 71 I define a defendant as likely to be ineligible for a safety-valve reduction if their case does not involve drugs, they

are charged with a weapon, or they have more than one criminal history point. 72 U.S. SENTENCING GUIDELINES MANUAL § 5C1.2 (2016).

38  

Columns (5), (6), and (7) use other dependent variables that seek to capture the extent to which

the defendant cooperated with the prosecution by pleading guilty rather than going to trial (column

(5)) and by receiving an acceptance of responsibility reduction (columns (6) and (7)).

If gender match between prosecutor and defendant leads to increased cooperation, one

would expect the coefficient estimates on the interaction terms to be greater than zero and

statistically significant in all columns of Table 9. This is not borne out for most defendants—with

one exception, none of the coefficient estimates are statistically significant and the 95% confidence

intervals suggest little influence of gender match on any of the outcome variables. The one

exception in which a coefficient estimate is marginally significant is the estimate on the interaction

term when the outcome variable is the substantial assistance reduction and the sample is limited to

those defendants for whom the reduction is likely to be very valuable. This evidence thus provides

more support for the idea that in-group favoritism is driven by automatic or implicit preferences,

rather than process-based explanations, with the caveat that cooperation might play a role for

defendants for whom it is extremely valuable.

6. Robustness

This section presents three robustness checks of the results. As reported in Table 2, a

defendant is—on average—around 2.2 percentage-points more likely to be assigned a prosecutor

of their own gender than a prosecutor of the opposite gender, which suggests that the assignment

of cases could be influenced, either consciously or subconsciously, by defendant and prosecutor

characteristics. As described in subsection 4.2, researchers observe similar assignment patterns in

the policing context, where officers typically are not randomly assigned to the neighborhoods in

which they work. The first two checks—described in subsections 6.1 and 6.2—are designed to test

whether differential assignment practices could be generating the results. The third robustness

check, described in subsection 6.3, demonstrates that the results are robust to including individual

prosecutor fixed effects.

6.1. Mixed-Gender Cases

Because prosecutors are assigned at the case-level (rather than the defendant-level) in the

data, I check whether the results hold up in the subset of cases with at least one male and at least

39  

one female co-defendant. As one might imagine, this restriction significantly reduces the sample—

from 143,457 defendants in the full data to just 26,099 defendants in these mixed-gender cases.

Appendix Table A.8 presents regression results based on the defendant composition of the

case. For reference, column (1) reproduces results in the full sample and is identical to column (4)

in Table 3. The remaining results in Table A.8 include estimates among cases with solo defendants

(column (2)); cases with multiple defendants of different genders (column (3)); and cases with

multiple defendants of the same gender (column (4)). Most observations in the data are cases with

just one defendant (column (2)) and the estimate of gender-based favoritism there is similar to the

estimate in the full data.73 The results are robust, however, to restricting to defendants who have

an opposite-gender co-defendant (column (3)). The point estimate on the interaction term is similar

to that in the full sample and is statistically significant at the five-percent level. The results among

cases in which there are multiple defendants of the same gender are reported for completeness but

are not particularly insightful because virtually all of these cases involve male defendants—it is

extremely rare for a multi-defendant case to include only female defendants.

6.2. Prosecutor and Defendant Characteristics Uncorrelated

The extent to which prosecutor gender is correlated with defendant gender (and other

covariates) is likely to be heterogeneous across USAOs and USAO sections. In particular, some

USAOs or USAO sections might assign cases using random assignment, or in ways that more

closely approximate random assignment, than others. In this subsection, I examine the results in

courthouse-sections in which prosecutor gender is uncorrelated with defendant gender and other

defendant characteristics that are fixed at the time of case assignment.

To carry out this robustness check, I regress prosecutor gender on each of five defendant

characteristics: age, and indicators for being black, Hispanic ethnicity, male, and a high school

graduate. I do a separate regression for each characteristic in each courthouse-section. All

regressions include year fixed effects. I then store the p-values associated with the coefficient

estimate of the relationship between each defendant characteristic and prosecutor gender for each

courthouse-section.

                                                            73 Although the coefficient estimate on prosecutor gender is significant in cases with a solo defendant, it is important

to remember that this represents a conditional effect. The average effect of prosecutor gender—obtained from a regression without the interaction term—is 0.004 and highly insignificant (p=0.964) (a precise zero).

40  

Appendix Table A.9 presents regression results when the data is restricted to courthouse-

sections in which prosecutor gender is not related to any of the defendant characteristics at the

one-, five-, and ten-percent levels. I find that the main regression results are highly robust to and

virtually unchanged by these restrictions, even in the specification that removes courthouse-

sections in which the coefficient estimate of prosecutor gender is significant at the ten-percent

level for any of the five defendant characteristics (column (7)), a restriction that meaningfully

reduces the size of the sample. These results suggest that the findings presented in the main

analysis are not driven by non-random assignment of cases to prosecutors. The even-numbered

columns report results without interaction terms and show that the coefficient estimate on the

prosecutor gender variable (representing average gender differences between male and female

prosecutors) is small and statistically insignificant in all specifications.

6.3. Prosecutor Fixed Effects

The data includes an anonymous prosecutor id variable. As a result, the analysis can be

performed with prosecutor-specific fixed effects. In the data, there are roughly 1,500 unique

prosecutors who appear in at least 25 cases. Appendix Table A.10 presents these results. Prosecutor

fixed effects control for unobserved differences between prosecutors that might be correlated with

the prosecutor’s gender.74

In Table A.10, I restrict the sample to prosecutors with at least 25 cases in the data—this

restriction reduces the sample by about ten percent, from 143,457 to 127,296. Column (1) in Table

A.6 presents the results in the full data with courthouse-section fixed effects—it is analogous to

column (4) in Table 3. In column (2), I include both courthouse-section fixed effects and individual

prosecutor fixed effects. In column (3), I perform an even more restrictive regression that includes

prosecutor-courthouse-section fixed effects. I then restrict to prosecutor-courthouse-section cells

with at least 25 observations, which further reduces the sample by roughly twenty percent. The

estimates of in-group favoritism in columns (2) and (3), however, are still statistically significant

at the one-percent level and are of a similar magnitude to the other estimates.

                                                            74 The ability to include individual-specific fixed effects will depend on the extent to which variation in the data is

within rather than across observations. For example, Price & Wolfers (2010) and Depew, Eren, & Mocan (2017) both include individual fixed effects in their main specifications. For Price & Wolfers (2010), these are individual player and referee fixed effects; for Depew, Eren, & Mocan (2017), these are judge (but not defendant) fixed effects. On the other hand, Antonovics & Knight 2009) investigate the possibility of using officer-specific fixed effects but conclude that including these fixed effects would not be well identified, and would drop many observations from the data.

41  

7. Discussion

This paper uses novel federal criminal case data covering nearly 150,000 criminal

defendants sentenced between 2002 and 2015 in 65 of the 94 federal district courts in the United

States. By exploiting the salutation field of the LIONS data, I obtain a variable that captures

prosecutor gender. The paper finds that defendants receive more lenient charges when they match

with their prosecutor on gender relative to when they are assigned a prosecutor of the opposite

gender. Through this charging channel, prosecutor-defendant pairs that match on gender produce

lower sentences relative to prosecutor-defendant pairs that mismatch. On the other hand, being

assigned a prosecutor of the same gender does not affect the probability that a defendant will

receive a sentence below the recommended Guidelines range—a decision that largely falls to the

defendant’s sentencing judge.

By examining estimates in subsamples of the data, I suggest that the findings are more

consistent with preference-based favoritism than process-based mechanisms. Specifically, gender-

based favoritism is stronger in states with less prevalent sexist attitudes, in cases in which gender

is likely to be more salient, and among same-gender prosecutorial teams. I argue that these findings

suggest that gender-based leniency derives from prosecutor preferences, and, in particular, could

be generated by female prosecutors feeling stronger empathic ties to female defendants than male

prosecutors do. I find little evidence to support a process-based explanation for the results: across

many alternative measures, defendants do not appear more likely to cooperate or place trust in a

prosecutor with whom they match on gender except among defendants for whom cooperation is

likely to be extremely valuable.

This paper thus documents an important and perhaps troubling phenomenon. The U.S.

Supreme Court has repeatedly emphasized that a central goal in federal sentencing law is to

promote the uniformity of sentences across defendants.75 The extent to which defendants are

treated differently based on something they cannot control—whether they happen to be assigned

a prosecutor of the same gender—conflicts with this essential idea that the punishment imposed

upon similarly situated defendants must be similar.

                                                            75 See, e.g., Booker, 543 U.S. at 255 (“Congress enacted the sentencing statutes in major part to achieve greater

uniformity in sentencing”); Kimbrough v. United States, 552 U.S. 85, 107 (2007) (“[I]t is unquestioned that uniformity remains an important goal of sentencing”).

42  

Critically here, where there is no clear benchmark as to a defendant’s optimal sentence,

one can not know whether the decisions produced by defendant-prosecutor gender matches are

superior or inferior to decisions produced by mismatches. There are, however, a few reasons to

believe the outcomes produced by gender match might produce better results.

First, some might think that leniency is an inherently desirable outcome. There is a

widespread—but not universal—view that the federal criminal system is too punitive. As evidence

of this point, for example, nearly half of federal felony defendants are sentenced below their

recommended Guidelines range while only two percent are sentenced above.

Second, if in-group leniency derives from empathy, as I suggest in subsection 5.1, gender-

favoritism might be desirable because empathy is considered “prosocial,” or, beneficial for society

(Cikara et. al 2011). Moreover, the fact that gender-based favoritism is more intense in states with

less pervasive sexism also suggests that such favoritism could create desirable outcomes.

Third, it is plausible that prosecutors are incentivized to seek the highest charges possible—

both because they are directed to do so by the Justice Manual76 and because this might be a metric

that is used for advancement. If so, leniency could be seen as a careful, thoughtful choice that

should be encouraged.

Fourth, in other contexts that examine in-group favoritism where there is an objective

measure of outcome quality, prior work has found that in-group matches produce better results.

For example, in the financial context, prior work has found that hedge fund managers are able to

better predict the performance of firms led by CEOs of the manager’s gender (Jannati et al. 2016),

and that loans perform better when the lender and borrower are culturally proximate (Fisman,

Paravisini, and Vig 2017). Both of these papers suggest that gender match produces objectively

better results. Of course, it is not clear that these results generalize to the criminal system.

Unfortunately, the lack of available data hinders understanding these important dynamics

in the criminal setting. While scholars have made strides in understanding judicial decision-making

thanks to rich and publicly-available data about judges,77 there is little publicly-available data for

researchers to use in analyzing decision-making by individual prosecutors. With more detailed

data about prosecutors, researchers would be able to further explore other kinds of individual bias

                                                            76 Justice Manual § 9-27-300. 77 For example, researchers can easily download a database with detailed biographical information about all past

and current federal judges from the Federal Judicial Center website: BIOGRAPHICAL DIRECTORY OF FEDERAL JUDGES: EXPORT, https://www.fjc.gov/history/judges/biographical-directory-article-iii-federal-judges-export.

43  

in prosecutorial decision-making. With this suggestion in mind, the remainder of this paper

discusses two potential policy changes that prosecutorial offices might implement.

Because gender-matching is associated with leniency, some might be tempted to embrace

in-group favoritism and propose policies that promote explicit gender-matching in prosecutorial

offices. Implementing an express policy of matching defendants to prosecutors on the basis of

gender would undoubtedly raise both constitutional and practical concerns. Moreover, even if one

wanted to match defendants to proximate prosecutors in an attempt to promote leniency, without

having examined in-group favoritism on many other dimensions—like race or religion—it is not

clear that gender is the best variable on which to match.

Another class of solutions might attempt to address outgroup disfavoritism and therefore

generate more equal treatment across defendants. Implicit bias training could be promising in this

respect. In 2016, the U.S. Department of Justice announced that it would begin providing all

federal law enforcement officers and prosecutors with implicit bias training. If in-group favoritism

stems from preference-based discrimination, this directive could address the problem but might,

perversely, make things worse. There is little evidence about the extent to which such training

programs can reduce bias in the criminal system and over the long-term, although such programs

have some success in simulated settings in the short-term (Plant and Peruche 2005). In the

employment setting more generally, however, many scholars have found that antibias and diversity

training are often ineffective, and can even lead to decreased support for diversity among white

participants (Plaut et al. 2011).

Prosecutors’ offices might also experiment with prosecution by teams. As described in

subsection 5.3, it appears that the presence of an opposite-gender peer mitigates the kind of bias

that generates in-group favoritism. There are, however, reasons to be cautious of this approach.

First, while it is possible that prosecution by teams could reduce in-group favoritism if team

members have a moderating influence on each other, one could imagine the group dynamics

playing out in other ways, too. For example, one might imagine that group decision-making in

prosecutor offices could lead to harsher treatment of defendants if team members always agreed

to the path proposed by the most punitive team member.

Second, and more broadly, it is important to remember that interventions in the criminal

and employment settings that policymakers expect will reduce disparity can often have perverse

effects. The prosecutorial response to Booker—in which prosecutors responded to a decrease in

44  

their ability to affect sentencing outcomes (via mandatory Guidelines) by charging mandatory

minimums more frequently—is just one example. Particularly when interventions alter the balance

of information and power between actors in the criminal system, caution is warranted.

There are still many important questions about prosecutorial behavior that remain

unanswered. For example, examining in-group favoritism on the basis of race is critical,

particularly in light of recent evidence of race-based in-group disfavoritism among judges in

criminal cases. Further work examining the scope of prosecutorial discretion and quantifying

disparities stemming from the assignment of prosecutors to defendants would also be valuable in

defining the contours of prosecutorial discretion in the criminal system.

45  

References

Abrams, David S., Marianne Bertrand, and Sendhil Mullainathan. 2012. "Do Judges Vary in Their Treatment of Race?" Journal of Legal Studies 347-383.

Aizer, Anna, and Joseph J., Jr. Doyle. 2015. "Juvenile Incarceration, Human Capital, and Future Crime: Evidence from Randomly Assigned Judges." Quarterly Journal of Economics 130 (2): 759-803.

Antonovics, Kate L., and Brian G. Knight. 2009. "A New Look at Racial Profiling: Evidence from the Boston Police Department." Review of Economics and Statistics 91 (1): 163-175.

Anwar, Shamena, Patrick Bayer, and Randi Hjalmarsson. 2012. "The Impact of Jury Race in Criminal Trials." The Quarterly Journal of Economics 127 (2): 1017-1055.

Barkow, Rachel E. 2013. "Prosecutorial Administration: Prosecutor Bias and the Department of Justice." Virginia Law Review 99: 271-342.

Barkow, Rachel E. 2009. "Institutional Design and the Policing of Prosecutors: Lessons from Administrative Law." Stanford Law Review 61: 869-921.

Becker, Gary S. 1957. The Economics of Discrimination. University of Chicago Press.

Bertrand, Marianne, and Esther Duflo. 2017. "Field Experiments on Discrimination." In Handbook of Field Experiments, by Abhijit Banerjee and Esther Duflo. North Holland.

Bertrand, Marianne, Esther Duflo, and Sendhil Mullainathan. 2004. "How Much Should We Trust Differences-in-Differences Estimates?" Quarterly Journal of Economics 119: 249-.

Bibas, Stephanos. 2009. "Prosecutorial Regulations versus Prosecutorial Accountability." University of Pennsylvania Law Review 157: 959-1016.

Bjerk, David. 2005. "Making the Crime Fit the Penalty: The Role of Prosecutorial Discretion under Mandatory Minimum Sentencing." The Journal of Law and Economics 48: 591-625.

Boyd, Christina L., Lee Epstein, and Andrew D. Martin. 2010. "Untangling the Causal Effects of Sex on Judging." American Journal of Political Science 54 (2): 389-411.

Boylan, Richard T. 2005. "What Do Prosecutors Maximize? Evidence from the Careers of U.S. Attorneys." American Law and Economics Review 7 (2): 379-402.

Boylan, Richard T., and Cheryl X. Long. 2005. "Salaries, Plea Rates, and the Career Objectives of Federal Prosecutors." Journal of Law and Economics 48 (2).

Brown, Robert A., and James Frank. 2006. "Race and Officer Decision Making: Examining Differences in Arrest Outcomes between Black and White Officers." Justice Quarterly 23 (1): 96-126.

Charles, Kerwin Kofi, Jonathan Guryan, and Jessica Pan. 2018. "The Effects of Sexism on American Women: The Role of Norms vs. Discrimination." NBER Working Paper 24904.

Chen, Daniel L. 2017. "Implicit Egoism in Sentencing Decisions: First Letter Name Effects with Randomly Assigned Defendants." TSE Working Paper 16-726.

46  

Cikara, Mina, Emile G. Bruneau, and Rebecca R. Saxe. 2011. "Us and Them: Intergroup Failures of Empathy." Current Directions in Psychological Science 20 20 (3): 149-153.

Cornell, Bradford, and Ivo Welch. 1996. "Culture, Information, and Screening Discrimination." Journal of Political Economy 104: 542-.

Cox, Adam B., and Thomas J. Miles. 2008. "Judging the Voting Rights Act." Columbia Law Review 108 (1): 1-54.

Cross, Frank B. 2003. "Decisionmaking in the U.S. Circuit Courts of Appeals." California Law Review 91 (6): 1457-1515.

Depew, Briggs, Ozkan Eren, and Naci Mocan. 2017. "Judges, Juveniles, and In-Group Bias." The Journal of Law and Economics 60 (2): 209-239.

Di Tella, Rafael, and Ernesto Schargrodsky. 2013. "Criminal Recidivism after Prison and Electronic Monitoring." Journal of Political Economy 121 (1): 28-73.

Didwania, Stephanie Holmes. 2020 (forthcoming). "The Immediate Consequences of Federal Pretrial Detention." American Law and Economics Review.

Doleac, Jennifer L., and Benjamin Hansen. 2017. "The Unintended Consequences of "Ban the Box": Statistical Discrimination and Employment Outcomes when Criminal Histories are Hidden." Working Paper. http://jenniferdoleac.com/wp-content/uploads/2015/03/Doleac_Hansen_BanTheBox.pdf.

Donohue III, John J., and Steven D. Levitt. 2001. "The Impact of Race on Policing and Arrests." Journal of Law and Economics 44 (2): 367-394.

Everett, Jim A.C, Nadira S Faber, and Molly Crockett. 2015. "Preferences and Beliefs in Ingroup Favoritism." Frontiers in Behavioral Neuroscience 9 (15): 1-15.

Farhang, Sean, and Gregory Wawro. 2004. "Institutional Dynamics on the U.S. Court of Appeals: Minority Representation Under Panel Decision Making." Journal of Law, Economics, and Organization 20 (2): 299-330.

Fischman, Joshua B., and Max M. Schanzenbach. 2012. "Racial Disparities Under the Federal Sentencing Guidelines: The Role of Judicial Discretion and Mandatory Minimums." Journal of Legal Studies 9 (2): 729-764.

Fiske, Susan T., Amy J.C Cuddy, Peter Glick, and Jun Xu. 2002. "A Model of (Often Mixed) Stereotype Content: Competence and Warmth Respectively Follow From Perceived Status and Competition." Journal of Personality and Social Psychology 82 (6): 878-902.

Fisman, Raymond, Daniel Paravisini, and Vikrant Vig. 2017. "Cultural Proximity and Loan Outcomes." American Economic Review 107 (2): 457-92.

Flanagan, Francis X. 2018. "Race, Gender, and Juries: Evidence from North Carolina." Journal of Law and Economics 61 (2): 189-214.

47  

Fryer Jr., Roland G. 2016. "An Empirical Analysis of Racial Differences in Police Use of Force." NBER Working Paper 22399.

Gazal-Ayal, Oren, and Raanan Sulitzeanu-Kenan. 2010. "Let My People Go: Ethnic In-Group Bias in Judicial Decisions-Evidence from a Randomized Natural Experiment." Journal of Empirical Legal Studies 7 (3): 403-428.

Geller, Amanda, Irwin Garfinkel, and Bruce Western. 2011. "Paternal Incarceration and Support for Children in Fragile Families." Demography 48 (1): 25-47.

Goldin, Claudia. 2014. "A Grand Gender Convergence: Its Last Chapter." American Economic Review 104 (4): 1091-1119.

Gorman, Elizabeth H. 2005. "Gender Stereotypes, Same-Gender Preferences, and Organizational Variation in the Hiring of Women: Evidence from Law Firms." American Sociological Review 70 (4): 702-728.

Greenwald, Anthony G, and Thomas F Pettigrew. 2014. "With Malice Toward None and Charity for Some: Ingroup Favoritism Enables Discrimination." American Psychologist 69 (7): 669-84.

Gutsell, Jennifer N., and Michael Inzlicht. 2012. "Intergroup Differences in the Sharing of Emotive States: Neural Evidence of an Empathy Gap." Social Cognition and Affective Neuroscience 7 (5): 596-603.

Hoyt, Crystal L., Stefanie Simon, and Lindsey Reid. 2009. "Choosing the Best (Wo)Man for the Job: The Effects of Mortality Salience, Sex, and Gender Stereotypes on Leader Evaluations." The Leadership Quarterly 20 (2): 233-246.

Huang, Bert I. 2011. "Lightened Scrutiny." Harvard Law Review 124 (5): 1109-1152.

Jackson, Robert H. 1940-1941. "The Federal Prosecutor." American Institute of Criminal Law and Criminology 41: 3-6.

Jannati, Sima, Alok Kumar, Alexandra Niessen-Ruenzi, and Justin Wolfers. 2016. "In-Group Bias in Financial Markets." Working Paper. December 24.

Kennedy, Jessica, Mary-Hunter McDonnell, and Nicole Stephens. 2017. "Do Women Face a Higher Ethical Bar? Exploring Discrimination in the Punishment of Ethical Violations." Working Paper. July 13.

Kim, Pauline T. 2009. "Deliberation and Strategy on the United States Courts of Appeals: An Empirical Exploration of Panel Effects." University of Pennsylvania Law Review 157: 1319-1381.

Krieger, Linda Hamilton. 1998. "Civil Rights Perestroika: Intergroup Relations After Affirmative Action." California Law Review 86: 1251-.

Kutateladze, Besiki L, Nancy R Andiloro, Brian D Johnson, and Cassia C Spohn. 2014. "Cumulative Disadvantage: Examining Racial and Ethnic Disparity in Prosecution and Sentencing." Criminology 52 (3): 514-551.

Kutateladze, Besiki Luka, Andiloro, Nancy R., Johnson, Brian D. 2014. "Opening Pandora's Box: How Does Defendant Race Influence Plea Bargaining?" Justice Quarterly 33 (3): 398-426.

48  

Lantigua-Williams, Juleyka. 2016. "Are Prosecutors the Key to Justice Reform?" The Atlantic, May 18. https://www.theatlantic.com/politics/archive/2016/05/are-prosecutors-the-key-to-justice-reform/483252/.

Leonetti, Carrie. 2012. "When the Emperor Has No Clothes III: Personnel Policies and Conflicts of Interest in Prosecutors' Offices." Cornell Journal of Law and Public Policy 22: 53-.

Lewis, Amy C., and Steven J. Sherman. 2010. "Perceived Entitativity and the Black-Sheep Effect: When Will We Denigrate Negative Ingroup Members." Journal of Social Psychology 150: 211-.

Lim, Claire S.H., Bernardo S. Silveria, and James M. Snyder. 2016. "Do Judges' Characteristics Matter? Ethnicity, Gender, and Partisanship in Texas State Trial Courts." American Law and Economics Review 18 (2): 302-357.

Lindeman, Marjaana, and Lilli Sundvik. 1995. "Evaluative Bias and Self-Enhancement Among Gender Groups." European Journal of Social Psychology 25 (3): 269-280.

Marques, Jose M., Vincent Y. Yzerbyt, and Jacques-Philippe Leyens. 1988. "The 'Black Sheep Effect': Extremity of Judgments Towards In-group Members as a Function of Group Identification." European Journal of Social Psychology 18: 1-.

McAdams, Richard H. 1998. "Race and Selective Prosecution: Discovering the Pitfalls of Armstrong." Chicago-Kent Law Review 73: 605-.

McLeish, Kendra N., and Robert J. Oxoby. 2011. "Social Interactions and the Salience of Social Identity." Journal of Economic Psychology 32 (1): 172-178.

Meares, Tracy L. 1995. "Rewards for Good Behavior: Influencing Prosecutorial Discretion and Conduct with Financial Incentives." Fordham Law Review 64: 851-919.

Miles, Thomas J. 2014. "Does the "Community Prosecution" Strategy Reduce Crime? A Test of Chicago's Experience." American Law and Economics Review 16 (1): 117-43.

Motivans, Mark. 2017. "Federal Justice Statistics, 2014 -- Statistical Tables." Bureau of Justice Statistics. March 23. Accessed November 21, 2017. https://www.bjs.gov/index.cfm?ty=pbdetail&iid=5873.

Mueller-Smith, Michael. 2015. "The Criminal and Labor Market Impacts of Incarceration." Working Paper.

Ouziel, Lauren M. 2017. "Ambition and Fruition in Federal Criminal Law: A Case Study." Virginia Law Review 103: 1077-.

Pager, Devah. 2003. "The Mark of a Criminal Record." American Journal of Sociology 108 (5): 937-975.

Peresie, Jennifer L. 2005. "Female Judges Matter: Gender and Collegial Decisionmaking in the Federal Appellate Courts." Yale Law Journal 114: 1759-.

Pfaff, John F. 2017. Locked In: The True Causes of Mass Incarceration and How to Achieve Real Reform. New York: Basic Books.

Plant, E. Ashby, and B. Michelle Peruche. 2005. "The Consequences of Race for Police Officers' Responses to Criminal Suspects." Psychological Science 16: 180-.

49  

Plaut, Victoria C., Flannery G. Garnett, Laura E. Buffardi, and Jeffrey Sanchez-Burks. 2011. "What About Me? Perceptions of Exclusion and Whites' Reactions to Multiculturalism." Journal of Personality and Social Psychology 101 (2): 337-53.

Price, Joseph, and Justin Wolfers. 2010. "Racial Discrimination among NBA Referees." Quarterly Journal of Economics 1859-1887.

Raggi, Reena. 1989. "Prosecutors' Offices: Where Gender is Irrelevant." Fordham Law Review 57 (6): 975-979.

Reedt, Lou, Courtney Semisch, and Kevin Blackwell. 2013. Effective Use of Federal Sentencing Data. Presentation to the American Society of Criminology, United States Sentencing Commission.

Rehavi, M. Marit, and Sonja B. Starr. 2014. "Racial Disparity in Federal Criminal Sentences." Journal of Political Economy 122 (6): 1320-1354.

Revesz, Richard L. 1997. "Environmental Regulation, Ideology, and the D.C. Circuit." Virginia Law Review 83 (8): 1717-1772.

Robertson, Christopher, Shima Baradaran Baughman, and Megan S. Wright. 2019. "Race and Class: A Randomized Experiment with Prosecutors." Journal of Empirical Legal Studies 16 (4): 807-847.

Rosenblum, Darren. 2000. ""Trapped" in Sing Sing: Transgendered Prisoners Caught in the Gender Binarism." Michigan Journal of Gender and Law 6: 499.

Ross, Jannell. 2016. "A Look Inside the Training the White House Thinks Could Stop Police Shooting, Protest, Damning DOJ Report Cycle." Washington Post, August 16. https://www.washingtonpost.com/news/the-fix/wp/2016/08/10/a-look-inside-the-training-the-white-house-thinks-could-stop-the-police-shootingprotestdamning-doj-report-cycle/?utm_term=.dea997f9ddc2.

Rowe, Brian. 2009. "Gender Bias in the Enforcement of Traffic Laws: Evidence Based on a New Empirical Test." Unpublished Manuscript. September.

Rudman, Laurie A., and Peter Glick. 1999. "Feminized Management and Backlash Toward Agentic Women: The Hidden Costs to Women of a Kinder, Gentler Image of Middle Managers." Journal of Personality and Social Psychology 77 (5): 1004-1010.

Rudman, Laurie A., and Stephanie A. Goodwin. 2004. "Gender Differences in Automatic In-Group Bias: Why Do Women Like Women More than Men Like Men?" Journal of Personality and Social Psychology 87 (4): 494-509.

Russell, Brenda L. 2014. Perceptions of Female Offenders: How Stereotypes and Social Norms Affect Criminal Justice Responses. New York: Springer.

Sanga, Sarath. 2014. "Does Officer Race Matter?" American Law and Economics Review 16: 403-.

Schanzenbach, Max M. 2005. "Racial and Gender Disparities in Prison Sentences: The Effect of District-Level Judicial Demographics." Journal of Legal Studies 57-92.

50  

Schanzenbach, Max M. 2015. "Racial Disparities, Judge Characteristics, and Standards of Review in Sentencing." Journal of Institutional and Theoretical Economics 27-47.

Shayo, Moses, and Asaf Zussman. 2011. "Judicial Ingroup Bias in the Shadow of Terrorism." The Quarterly Journal of Economics 126 (3): 1447-1484.

Sklansky. 2017. "The Nature and Function of Prosecutorial Power." Journal of Criminal Law and Criminology 106: 473.

Sloan, CarlyWill. 2019. "Racial Bias by Prosecutors: Evidence from Random Assignment." Working Paper.

Spohn, Cassia. 1990. "The Sentencing Decisions of Black and White Judges: Expected and Unexpected Similarities." Law & Society Review 24 (5): 1197-1216.

Starr, Sonja B., and M. Marit Rehavi. 2013. "Mandatory Sentencing and Racial Disparity: Assessing the Role of Prosecutors and the Effects of Booker." Yale Law Journal 2-80.

Starr, Sonja. 2015. "Estimating Gender Disparities in Federal Criminal Cases." American Law and Economics Review 17 (1): 127-159.

Steffensmeier, Darrell, and Chris Hebert. 1999. "Women and Men Policymakers: Does the Judge's Gender Affect the Sentencing of Criminal Defendants?" Social Forces 77 (3): 1163-1196.

Sunstein, Cass R., David Schkade, and Lisa Michelle Ellman. 2004. "Ideological Voting on Federal Courts of Appeals: A Preliminary Investigation." Virginia Law Review 90 (1): 301-354.

Uhlman, Thomas M. 1978. "Black Elite Decision Making: The Case of Trial Judges." American Journal of Political Science 22 (4): 884-895.

Vial, Andrea C, Victoria L Brescoll, Jaime L Napier, John F Dovidio, and Tom R Tyler. 2017. "Differential Support for Female Supervisors Among Men and Women." Journal of Applied Psychology 103 (2): 215-227.

Ward, Geoff, Amy Farrell, and Danielle Rousseau. 2009. "Does Racial Balance in Workforce Representation Yield Equal Justice? Race Relations of Sentencing in Federal Court Organizations ." Law and Society Review 757-806.

Welch, Susan, Michael Combs, and John Gruhl. 1988. "Do Black Judges Make a Difference?" American Journal of Political Science 32 (1): 126-136.

Western, Bruce. 2002. "The Impact of Incarceration on Wage Mobility and Inequality." American Sociological Review 67 (4): 526-546.

Wiswall, Matthew, and Basit Zafar. 2018. "Preference for the Workplace, Investment in Human Capital, and Gender." The Quarterly Journal of Economics 133 (1): 457-507.

Xu, Xiaojing, Xiangyu Zuo, Xiaoying Wang, and Shihui Han. 2009. "Do You Feel My Pain? Racial Group Membership Modculates Empathic Neural Responses." The Journal of Neuroscience 29 (26): 8525-8529.

51  

Yang, Crystal S. 2015. "Free at Last? Judicial Discretion and Racial Disparities in Federal Sentencing." Journal of Legal Studies 44 (1): 75-111.

Yang, Crystal S. 2014. "Have Interjudge Disparities Increased in an Advisory Guidelines Regime? Evidence from Booker." New York University Law Review 89 (4): 1268-1342.

52  

Appendix A: Gender Coding

This Appendix lists all of the salutation codes in the LIONS data and indicates how they were coded in the paper. F indicates female; M indicates male; and U indicates unknown.

Salutation Frequency (pct) Coding Continued… AUS 0.33 U SAUSA 0.01 U

AUSA 1.98 U UFAP 0.04 U Al 0.01 M USA 0.00 U

Amy 0.01 F Vaughn 0.02 M Ausa 0.00 U a 0.02 U

CAPT 0.00 U aa 0.00 U Capt 0.01 U ch 0.00 U Capt. 0.02 U d 0.00 U

Captain 0.02 U db 0.04 U Captian 0.00 U ep 0.00 U Caption 0.00 U gh 0.00 U

Cpt 0.01 U js 0.00 U Cpt. 0.01 U kp 0.00 U

Diana 0.00 F mR. 0.05 M Edie 0.00 U mS. 0.00 F

FAUSA 0.01 U mr 0.16 M Gentry 0.02 U mr. 0.01 M Hello 0.04 U ms. 0.10 F Hon. 0.04 U s 0.05 U Janet 0.01 F ss 0.04 U Jeff 0.02 M wn 0.08 U Jim 0.01 M Joe 0.00 M

John 0.02 M Lt. 0.00 U M 0.02 U

MISS 0.00 F MR 1.34 M MR. 2.77 M MRS 0.04 F MRS. 0.05 F MS 0.89 F MS. 1.45 F Miss 0.07 F Mr 4.65 M Mr, 0.00 M Mr. 59.10 M Mr.s 0.00 F Mrs 0.13 F Mrs. 2.52 F Ms 2.10 F Ms, 0.00 F Ms. 21.71 F Ns. 0.00 U

Richard 0.00 M

53  

Figure 1. Federal Judicial Districts Appearing in the Data

Notes: Shaded districts are those that are represented in the paper.

54  

Figure 2. Histogram of Base Offense Level

Notes: This figure plots the distribution of base offense levels in the data (N=143,457).

55  

Figure 3. Base Offense Level Averages by Defendant and Prosecutor Gender

Notes: Average base offense level de-meaned by courthouse.

16

17

18

19

20

21

Male D Female D All

Male P Female P

56  

Figure 4. Share of Cases with a Male Lead Prosecutor, Over Time

Notes: This figure plots the mean share of male lead prosecutors by year of case initiation. Annual means are computed with one observation per defendant. Horizontal line at the mean among all years: 0.716.

57  

Figure 5. District-Level Estimates of In-Group Favoritism

Notes: Shades categorize the quartile in which each federal judicial district’s point estimate of in-group favoritism falls.

Top 25%

25-50%

50-75%

Bottom 25%

58  

Table 1. Summary Statistics

Female Defendants

Male Defendants

All Defendants Min Max

Defendant Characteristics Male - - 0.829 0 1 Black 0.317 0.419 0.402 0 1 Hispanic 0.103 0.127 0.123 0 1 White 0.518 0.407 0.426 0 1 Another Race 0.062 0.047 0.050 0 1 High School Graduate Only 0.682 0.610 0.622 0 1 College Graduate 0.081 0.071 0.072 0 1 Age (years) 37.4 36.2 36.4 16 92 Number of Dependents 1.322 1.482 1.455 0 28 Criminal History Points78 2.20 5.54 4.97 0 65 Case Characteristics Lead Male Prosecutor 0.706 0.718 0.716 0 1 Multiple Defendants 0.444 0.377 0.389 0 1 Multiple Prosecutors 0.354 0.387 0.381 0 1 Base Offense Level 15.4 20.4 19.5 0 44 Final Offense Level 16.7 21.3 20.5 1 43 Case Start Year 2006.8 2006.7 2006.7 1998 2014 Sentencing FY 2008.5 2008.4 2008.4 2002 2016 Intermediate Outcomes Facing Mandatory Minimum 0.223 0.367 0.342 0 1 Substantial Assistance Reduction 0.255 0.217 0.223 0 1 Mean Guidelines Range 47.1 91.9 84.3 0 470 Case Outcomes Any Incarceration 0.684 0.892 0.856 0 1 Sentence (months)79 26.2 67.2 60.2 0 470 Sentence / Mean Guidelines Range 0.455 0.702 0.660 0 32 Below-Guidelines (0/1) 0.603 0.481 0.502 0 1 Above-Guidelines (0/1) 0.011 0.024 0.022 0 1 In-Range (0/1) 0.385 0.495 0.476 0 1 Observations 24,514 118,943 143,457 143,457 143,457

                                                            78 Total criminal history points are unadjusted. 79 Not including alternative confinement, such as house arrest. Sentences are capped at 470 months—the

Commission’s value for life sentences.

59  

Table 2. Regressing Male Prosecutor on Defendant Characteristics All

Defendants All Defendants All Defendants

Female Defendants

Male Defendants

(1) (2) (3) (4) (5) Male 0.015**

(0.006) 0.015** (0.006)

0.022*** (0.005)

- -

Black -0.032*** (0.012)

-0.031*** (0.012)

-0.001 (0.007)

-0.012 (0.011)

-0.0001 (0.007)

Hispanic -0.065** (0.027)

-0.062** (0.027)

-0.0009 (0.007)

-0.004 (0.012)

-0.00007 (0.008)

Another Race -0.044* (0.026)

-0.043 (0.026)

-0.004 (0.013)

-0.003 (0.024)

-0.004 (0.013)

Age -0.0004** (0.0002)

-0.0003 (0.0002)

-0.0003** (0.0001)

0.00006 (0.0004)

-0.0003** (0.0001)

Dependents 0.00003 (0.001)

0.00003 (0.001)

0.0003 (0.0008)

-0.0006 (0.0004)

0.0006 (0.0008)

HS Only 0.002 (0.004)

0.004 (0.004)

-0.0002 (0.003)

-0.001 (0.006)

-0.0003 (0.003)

College Grad 0.008 (0.009)

0.012 (0.009)

0.009 (0.007)

0.012 (0.014)

0.005 (0.008)

Multiple Defendants 0.004 (0.013)

0.007 (0.013)

-0.001 (0.010)

-0.003 (0.011)

-0.002 (0.011)

Multiple Prosecutors -0.025 (0.015)

-0.024 (0.016)

-0.009 (0.011)

-0.003 (0.015)

-0.011 (0.011)

Prosecutor Workload 0.0003 (0.001)

0.0002 (0.001)

-0.0003 (0.0009)

0.0005 (0.0009)

-0.0004 (0.0009)

F-Stat: Criminal History

0.40 (p=0.81)

0.43 (p=0.78)

1.28 (p=0.28)

1.07 (p=0.37)

0.88 (p=0.48)

F-Stat: Offense Type 1.60 (p=0.19)

1.73 (p=0.16)

0.26 (p=0.85)

0.38 (p=0.77)

0.71 (p=0.55)

F-Stat: Race/Ethnicity 4.49*** (0.005)

4.24*** (p=0.006)

0.03 (p=0.99)

0.43 (p=0.73)

0.03 (p=0.99)

Month and Year FEs X X X X Courthouse-Section FEs X

X X

Observations 143,457 143,457 143,457 21,512 118,318 OLS regressions of the gender of the defendant’s first lead prosecutor (1=male; 0=female) on defendant and case characteristics. ***: p<0.01; **: p<0.05; *: p<0.10. Standard errors are clustered at the courthouse level. Prosecutor workload variable is the total number of cases for which that prosecutor was the lead prosecutor in that year. Month and year fixed effects include sentencing fiscal year, Guidelines version used, and month and year of case initiation. Observations in columns (4) and (5) do not add up to 143,457 because they exclude courthouse-sections with fewer than 25 observations.

60  

Table 3. Main Results (Base Offense Level) (1) (2) (3) (4) (5) (6) Interaction -0.727**

(0.294)

-0.533*** (0.147)

-0.529*** (0.148)

-0.542*** (0.149)

-0.661*** (0.173)

-

Male Defendant 5.432*** (0.249)

1.946*** (0.155)

1.927*** (0.160)

1.938*** (0.161)

2.014*** (0.199)

1.554*** (0.095)

Male Prosecutor 0.761

(0.495) 0.461** (0.191)

0.426** (0.182)

0.421** (0.185)

0.436** (0.219)

-0.027 (0.099)

Demographic Controls X X X X X Month and Year FEs X X X X District-Section FEs X Courthouse-Section FEs X X Courthouse-Section-Year FEs

X

Observations 143,457 143,457 143,457 143,457 103,858 143,457

OLS regressions of the defendant’s base offense level. ***: p<0.01; **: p<0.05; *: p<0.10. Standard errors are clustered at the courthouse level. Demographic controls include race, age at sentencing, number of dependents, criminal history quintile, prosecutor caseload, and indicators for the defendant being a high school graduate, a college graduate, the case having multiple defendants, and the case having multiple prosecutors. Month and year fixed effects include sentencing fiscal year, Guidelines version used, and month and year of case initiation. There are fewer observations in column (5) because courthouse-section-year cells with fewer than 25 observations are excluded.

61  

Table 4. Additional Outcome Variables Panel A: Prosecutor-Based Outcomes

Final Offense Level

(1)

Facing Mandatory Minimum (0/1)

(2)

Final Guidelines Range (IHS)

(3)

Substantial Assistance (0/1)

(4) Interaction -0.347**

(0.157)

-0.017*** (0.006)

-0.057** (0.024)

-0.0003 (0.008)

Male Defendant

2.452*** (0.168)

0.078*** (0.006)

0.370*** (0.024)

-0.041*** (0.008)

Male Prosecutor 0.276 (0.188)

0.009 (0.008)

0.055* (0.029)

0.012 (0.008)

Observations 143,457 143,307 143,456 143,246

Panel B: Judge- and Prosecutor-Based Outcomes Released Pretrial

(0/1) (5)

Sentence Length (IHS)

(6)

Sentence (IHS) (control for BOL)

(7)

Below-Guidelines Sentence (0/1)

(8) Interaction 0.013

(0.009) -0.102** (0.043)

-0.042 (0.034)

0.005 (0.008)

Male Defendant

-0.127*** (0.009)

0.848*** (0.041)

0.632*** (0.034)

-0.087*** (0.010)

Male Prosecutor -0.001 (0.008)

0.081* (0.048)

0.034 (0.035)

0.004 (0.007)

Observations 141,046 143,457 143,457 143,456 OLS regressions. ***: p<0.01; **: p<0.05; *: p<0.10. Standard errors are clustered at the courthouse level. All regressions include demographic controls listed in Table 3, column (4), month and year fixed effects, and courthouse-section fixed effects.

62  

Table 5. In-group Favoritism and Selection into the Sample (1) (2) (3) (4) Dependent Variable: Acquittal/Dismissal/Declination Interaction 0.003

(0.006) 0.004

(0.006) 0.002

(0.006) -

Male Defendant -0.014**

(0.006)

-0.015** (0.006)

-0.015** (0.006)

-0.013*** (0.003)

Male Prosecutor -0.002 (0.008)

-0.003 (0.008)

-0.006 (0.007)

-0.004 (0.003)

Demographic Controls X X X X Month and Year FEs X X X Courthouse-Section FEs X X Observations 89,117 89,117 87,183 87,183

OLS regressions of whether a defendant’s case was acquitted, declined, or dismissed (0/1). ***: p<0.01; **: p<0.05; *: p<0.10. Standard errors are clustered at the courthouse level. Month and year fixed effects include sentencing fiscal year and month and year of case initiation. Demographic controls are indicators for whether the defendant is male, whether the case involved multiple defendants, whether the case had multiple prosecutors, and offense type.

63  

Table 6. In-Group Favoritism and Sexism

Male Defendants

Female Defendants

Male Defendants

Female Defendants

Low Sexism States

High Sexism States

All Defendants All Defendants

(1) (2) (3) (4) (5) (6) (7) (8)

Dependent Variable: Base Offense Level

Male Defendant - - - - 2.233*** (0.166)

1.582*** (0.233)

1.682*** (0.104)

2.175*** (0.176)

Male Prosecutor -0.053 (0.116)

0.398** (0.180)

-0.021 (0.168)

0.709*** (0.251)

0.791*** (0.259)

0.151 (0.280)

-0.040 (0.104)

0.785*** (0.268)

High Sexism State (0/1) -0.057 (0.216)

-0.163 (0.239)

-0.007 (0.231)

0.270 (0.308)

- - -0.439 (0.272)

0.480 (0.353)

Male D * Male P - - - - -0.834*** (0.199)

-0.240 (0.204)

- -0.824*** (0.204)

Male D * High Sexism - - - - - - -0.336* (0.183)

-0.512* (0.288)

Male P * High Sexism - - -0.068 (0.227)

-0.606* (0.344)

- - - -0.635* (0.377)

Male D * Male P * High Sexism

- - - - - - - 0.577** (0.286)

Observations 112,943 23,366 112,943 23,366 67,595 68,714 136,309 136,309

OLS regressions of the the defendant’s base offense level. ***: p<0.01; **: p<0.05; *: p<0.10. Standard errors are clustered at the courthouse level. All regressions include demographic controls listed in Table 3, month and year fixed effects.

64  

Table 7. In-Group Favoritism by Offense Type Drug

Offenses Regulatory Offenses

Violent Offenses

White Collar Offenses

Dependent Variable: Base Offense Level Interaction -0.706***

(0.250) -0.481* (0.252)

-0.996** (0.432)

-0.223* (0.125)

Male Defendant 1.836***

(0.262) 2.604*** (0.260)

2.342*** (0.375)

1.021*** (0.100)

Male Prosecutor 0.434

(0.334) 0.303

(0.269) 0.915** (0.451)

0.369*** (0.100)

Male Defendants (share) 0.86 0.78 0.94 0.67 Male Prosecutors (share) 0.72 0.70 0.70 0.72 Observations 69,449 17,376 23,450 22,741

OLS regressions of the the defendant’s base offense level. ***: p<0.01; **: p<0.05; *: p<0.10. Standard errors are clustered at the courthouse level. All regressions include demographic controls listed in Table 3, month and year fixed effects, and courthouse-section fixed effects. Observations do not add up to 143,457 because courthouse-sections types with fewer than 25 observations are excluded.

65  

Table 8. In-Group Favoritism in Prosecutorial Teams All Cases Solo Prosecutors Multiple Ps of

Same Gender Multiple Ps: No

Female Ps of Mixed

Genders: First Ps of Mixed

Genders: Fraction (1) (2) (3) (4) (5) (6) Dependent Variable: Base Offense Level

Interaction -0.542*** (0.149)

-0.442*** (0.163)

-1.136*** (0.353)

-0.743*** (0.211)

-0.465 (0.300)

-0.877 (1.230)

Male Defendant 1.938*** (0.161)

1.696*** (0.177)

2.453*** (0.328)

2.128*** (0.172)

2.393*** (0.276)

2.586*** (0.655)

Male Prosecutor 0.421** (0.185)

0.355* (0.199)

0.820** (0.338)

0.639*** (0.228)

0.300 (0.394)

0.671 (1.435)

Observations 143,457 87,649 29,207 52,439 20,345 20,345 OLS regressions of the the defendant’s base offense level. ***: p<0.01; **: p<0.05; *: p<0.10. Standard errors are clustered at the courthouse level. All regressions include demographic controls listed in Table 3, month and year fixed effects, and courthouse-section fixed effects. All regressions restricted to courthouse-section cells with at least 25 observations. Columns (3)-(6) also control for the number of prosecutors assigned to the case.

66  

Table 9. Alternative Measures of Cooperation Dependent Variable: Supserseding

Info/Indictment (1)

Substantial Assistance

(2)

Sub Assist: MM & Safety Inelig.

(3)

Safety Valve

(4)

Pled Guilty (No Trial)

(5)

Any Acceptance of Responsibility

(6)

Max Acceptance of Responsibility

(7) Interaction 0.008

(0.006)

-0.0003 (0.008)

0.038* (0.022)

-0.012 (0.009)

0.003 (0.003)

0.002 (0.004)

0.001 (0.005)

Male Defendant 0.004 (0.005)

-0.041*** (0.008)

-0.153*** (0.021)

-0.055*** (0.008)

-0.018*** (0.003)

-0.014*** (0.004)

-0.022*** (0.004)

Male Prosecutor -0.010 (0.008)

0.012 (0.008)

-0.017 (0.024)

0.014 (0.009)

-0.002 (0.003)

-0.002 (0.004)

-0.0001 (0.005)

Observations 143,457 143,457 33,630 68,303 143,457 143,457 143,457 OLS regressions. ***: p<0.01; **: p<0.05; *: p<0.10. Standard errors are clustered at the courthouse level. All regressions include demographic controls listed in Table 3, month and year fixed effects, and courthouse-section fixed effects. Column (2) includes defendants who are facing a mandatory minimum greater than zero and are ineligible for a safety-valve reduction, either because their cases does not involve a drug offense, they are charged with a weapon, or they have more than one criminal history point.

67  

Table A.1. Representativeness of the Sample Data in this

paper (1)

Districts in this paper

(2)

Matched Defendants

(3)

Commission Defendants

(4) Defendant Characteristics Male 0.83 0.84 0.83 0.84 Black 0.40 0.39 0.38 0.39 Hispanic 0.12 0.13 0.13 0.13 White 0.43 0.43 0.44 0.43 Another Race 0.05 0.05 0.05 0.05 High School Graduate Only 0.62 0.62 0.62 0.62 College Graduate 0.07 0.08 0.08 0.08 Age (years) 36.4 36.7 36.9 36.6 Number of Dependents 1.5 1.4 1.4 1.4 Criminal History Points80 5.0 4.7 4.7 4.7 Case Characteristics Lead Male Prosecutor 0.72 - - - Multiple Defendants 0.39 - - - Multiple Prosecutors 0.38 - - - Base Offense Level 19.5 19.3 19.1 19.3 Final Offense Level 20.5 20.6 20.5 20.6 Case Start Year 2007 - - - Sentencing FY 2008 2009 2009 2008 Intermediate Outcomes Facing Mandatory Minimum 0.34 0.34 0.33 0.35 Substantial Assistance Reduction 0.22 0.21 0.20 0.20 Mean Guidelines Range 84.3 85.5 82.1 86.0 Case Outcomes Any Incarceration 0.86 0.83 0.84 0.83 Sentence (months)81 60.2 62.4 59.3 63.1 Sentence / Mean Guidelines Range 0.66 0.65 0.66 0.65 Below-Guidelines (0/1) 0.50 0.50 0.48 0.47 Above-Guidelines (0/1) 0.02 0.02 0.02 0.02 In-Range (0/1) 0.48 0.48 0.50 0.51 Observations 143,457 381,503 374,529 501,537

Districts in the sample are depicted in Figure 1.

                                                            80 Total criminal history points are unadjusted. 81 Not including alternative confinement, such as house arrest. Sentences were capped at 470 months—the

Commission’s value for life sentences.

68  

Table A.2. Alternative Measures of Charging Severity Base Offense Level Stat Min

(IHS) Average Sentence

(IHS) (1) (2) (3) Interaction -0.542***

(0.149)

-0.094*** (0.033)

-0.037** (0.018)

Male Defendant 1.938*** (0.161)

0.424*** (0.035)

0.143*** (0.018)

Male Prosecutor 0.421** (0.185)

0.054 (0.044)

0.033** (0.020)

Observations 143,457 143,307 132,585

OLS regressions. ***: p<0.01; **: p<0.05; *: p<0.10. Standard errors are clustered at the courthouse level. All regressions include demographic controls listed in Table 3, month and year fixed effects, and courthouse-section fixed effects. The statutory minimum is the total statutory minimum prison term for all counts of conviction transformed by its inverse hyperbolic sine.82 The average sentence is the leave-one-out-mean sentence for defendants convicted under the same Chapter Two Guideline in the defendant’s courthouse transformed by its inverse hyperbolic sine.

                                                            82 The statutory minimum is capped at 470 months; 410 defendants had statutory minimums above this value.

69  

Table A.3. Prosecutor Gender and Defendant Characteristics in LIONS Sample Data Used in

the Paper (All Controls)

Data Used in the Paper

(LIONS Controls)

Full LIONS (LIONS Controls)

Male 0.022*** (0.005)

0.020*** (0.005)

0.022*** (0.007)

Multiple Defendants -0.001

(0.010) -0.001 (0.010)

0.002 (0.012)

Multiple Prosecutors -0.009

(0.011) -0.009 (0.011)

-0.012 (0.016)

F-Stat: Offense Type 0.26

(p=0.85) 0.42

(p=0.74)

0.93 (p=0.43)

Month and Year FEs X X X Courthouse-Section FEs X X X Observations 143,457 143,457 82,021

OLS regressions of the gender of the defendant’s first lead prosecutor on defendant and case characteristics. ***: p<0.01; **: p<0.05; *: p<0.10. Standard errors are clustered at the courthouse level. Month and year fixed effects include sentencing fiscal year and month and year of case initiation. All Controls are demographic controls listed in Table 3. LIONS Controls are indicators for whether the defendant is male, whether the case involved multiple defendants, whether the case had multiple prosecutors, and offense type. All regressions restricted to courthouse-sections with at least 25 observations.

70  

Table A.4. Results by Defendant Race/Ethnicity All Cases Black

Defendants Non-Black Defendants

Hispanic Defendants

Non-Hispanic Defendants

White Defendants

Non-White Defendants

(1) (2) (3) (4) (5) (6) (7) Dependent Variable: Base Offense Level Male D * Male P -0.494**

(0.198) -0.560***

(0.213) -0.466***

(0.164) -0.924***

(0.289) -0.440***

(0.161) -0.357** (0.180)

-0.569*** (0.185)

Male Defendant 2.677*** (0.203)

2.441*** (0.222)

1.554*** (0.187)

2.659*** (0.280)

1.735*** (0.175)

1.231*** (0.180)

2.439*** (0.184)

Male Prosecutor 0.396** (0.192)

0.472** (0.228)

0.330 (0.204)

0.846** (0.363)

0.297 (0.206)

0.236 (0.238)

0.492** (0.202)

Male D * Male P * Hispanic -0.252 (0.307)

- - - - - -

Male D * Male P * White 0.179 (0.245)

- - - - - -

Number of Courthouses 180 133 158 85 174 151 153 Number of CH-Sections

494 304 438 137 481 414 359

Observations 136,203 55,467 77,610 15,174 118,277 59,598 73,468 OLS regressions of the the defendant’s base offense level. ***: p<0.01; **: p<0.05; *: p<0.10. Standard errors are clustered at the courthouse level. All regressions include demographic controls listed in Table 3, month and year fixed effects, and courthouse-section fixed effects. All regressions restricted to courthouse-sections cells with at least 25 observations. Regression coefficients for race, race*prosecutor gender and race*defendant gender are unreported but included in the regression reported in column (1).

71  

Table A.5. Results by Defendant Race/Ethnicity and Offense Type Drug Offenses Regulatory Offenses Violent Offenses White Collar Offenses Non-White White Non-White White Non-White White Non-White White Dependent Variable: Base Offense Level Interaction -0.748**

(0.318) -0.305 (0.299)

-0.351 (0.270)

-0.620 (0.381)

-0.721 (0.675)

-0.816* (0.480)

-0.069 (0.163)

-0.203 (0.194)

Male Defendant 2.982*** (0.284)

0.610** (0.294)

3.191*** (0.243)

2.782*** (0.385)

2.513*** (0.552)

1.707*** (0.398)

0.715*** (0.146)

1.275*** (0.162)

Male Prosecutor 0.664* (0.367)

0.059 (0.456)

0.356 (0.246)

0.179 (0.382)

0.636 (0.691)

0.866* (0.509)

0.177 (0.127)

0.304** (0.137)

Male Defendants (share) 0.905 0.791 0.780 0.788 0.950 0.932 0.635 0.694 Male Prosecutors (share)

0.716 0.742 0.672 0.715 0.681 0.743 0.689 0.738

Observations 42,548 26,518 8,201 10,358 14,249 10,253 10,285 13,931 OLS regressions of the the defendant’s base offense level. ***: p<0.01; **: p<0.05; *: p<0.10. Standard errors are clustered at the courthouse level. All regressions include demographic controls listed in Table 3, month and year fixed effects, and courthouse-section fixed effects. All regressions restricted to courthouse-section with at least 25 observations.

72  

Table A.6. Additional Outcome Variables—By Race/Ethnicity Panel A: Early Outcomes

Final Offense Level (1)

Facing Mandatory Min (0/1) (2)

Final Guidelines Range (IHS) (3)

Subst. Assistance (0/1) (4)

Non-White White Non-White White Non-White White Non-White White Interaction -0.412**

(0.208) -0.185 (0.185)

-0.019* (0.011)

-0.006 (0.008)

-0.070** (0.029)

-0.031 (0.028)

0.00006 (0.010)

0.0002 (0.010)

Male Defendant

2.810*** (0.212)

1.830*** (0.195)

0.104*** (0.011)

0.044*** (0.008)

0.430*** (0.030)

0.280*** (0.030)

-0.057*** (0.013)

-0.019** (0.009)

Male Prosecutor 0.348 (0.229)

0.115 (0.232)

0.006 (0.011)

0.004 (0.012)

0.068** (0.034)

0.033 (0.035)

0.009 (0.010)

0.015 (0.011)

Observations 73,468 59,598 73,381 59,551 73,467 59,598 73,468 59,598

Panel B: Sentencing Outcomes Pretrial Release (0/1)

(5)

Sentence (IHS)

(6)

Sentence Length (IHS) (control for BOL)

(7)

Below-Guidelines (0/1)

(8) Non-White White Non-White White Non-White White Non-White White Interaction -0.002

(0.011) 0.021

(0.013) -0.117** (0.054)

-0.065 (0.051)

-0.057 (0.048)

-0.023 (0.042)

-0.003 (0.011)

0.007 (0.010)

Male Defendant

-0.170*** (0.013)

-0.087*** (0.011)

1.015*** (0.057)

0.659*** (0.044)

0.761*** (0.051)

0.515*** (0.035)

-0.105*** (0.014)

-0.066*** (0.008)

Male Prosecutor 0.014 (0.011)

-0.010 (0.012)

0.108* (0.057)

0.034 (0.058)

0.057 (0.048)

0.007 (0.041)

0.009 (0.010)

0.005 (0.008)

Observations 72,594 58,348 73,468 59,598 73,468 59,598 73,469 59,598 OLS regressions. ***: p<0.01; **: p<0.05; *: p<0.10. Standard errors are clustered at the courthouse level. All regressions include demographic controls listed in Table 3, column (4), month and year fixed effects, and courthouse-section fixed effects.

73  

Table A.7. All Outcome Variables with Gender Share Panel A: Early Outcomes

Base Offense Level

(1)

Final Offense Level

(2)

Facing Mandatory

Minimum (0/1) (3)

Final Guidelines Range (IHS)

(4)

Substantial Assistance (0/1)

(5)

Interaction -0.610*** (0.165)

-0.471*** (0.172)

-0.019** (0.007)

-0.068*** (0.026)

0.003 (0.008)

Male Def (0/1)

1.982*** (0.169)

2.538*** (0.177)

0.079*** (0.007)

0.378*** (0.025)

-0.043*** (0.008)

Male Pros (fraction)

0.454** (0.194)

0.284 (0.192)

0.011 (0.009)

0.053* (0.030)

0.007 (0.009)

Observations 143,457 143,457 143,307 143,457 143,457

Panel B: Sentencing Outcomes Pretrial Release (0/1)

(6)

Sentence Length (IHS)

(7)

Sentence (IHS) (control for BOL)

(8)

Below-Guidelines Sentence (0/1)

(9) Interaction 0.023**

(0.009) -0.122***

(0.043)

-0.054 (0.034)

0.007 (0.009)

Male Def (0/1)

-0.134*** (0.010)

0.861*** (0.042)

0.640*** (0.033)

-0.089*** (0.011)

Male Pros (fraction)

-0.007 (0.009)

0.082* (0.047)

0.031 (0.033)

0.003 (0.009)

Observations 141,046 143,457 143,457 143,456 OLS regressions. ***: p<0.01; **: p<0.05; *: p<0.10. Standard errors are clustered at the courthouse level. All regressions include demographic controls listed in Table 3, month and year fixed effects, and courthouse-section fixed effects. All regressions restricted to courthouse-sections with at least 25 observations.

74  

Table A.8. Mixed-Gender Cases

All Cases Solo Defendants

Multiple Ds Different Gender

Multiple Ds Same Gender

(1) (2) (3) (4) Dependent Variable: Base Offense Level Interaction -0.542***

(0.149)

-0.641*** (0.175)

-0.448** (0.226)

0.366 (0.570)

Male Defendant 1.938*** (0.161)

2.068*** (0.142)

1.751*** (0.222)

1.540*** (0.571)

Male Prosecutor 0.421** (0.185)

0.517*** (0.169)

0.284 (0.276)

-0.459 (0.594)

Observations 143,457 87,046 26,099 24,474

OLS regressions of the the defendants base offense level. ***: p<0.01; **: p<0.05; *: p<0.10. Standard errors are clustered at the courthouse level. All regressions include demographic controls listed in Table 3, month and year fixed effects, and courthouse-section type fixed effects. All regressions restricted to courthouse-sections with at least 25 observations. Columns (3)-(4) also control for the number of defendants in each case.

75  

Table A.9. Courthouse-Sections in which Prosecutor Gender Uncorrelated with Five Characteristics Full Sample Full Sample p>0.01 p>0.01 p>0.05 p>0.05 p>0.10 p>0.10 (1) (2) (3) (4) (5) (6) (7) (8) Dependent Variable: Base Offense Level Interaction -0.542***

(0.149)

- -0.600*** (0.162)

- -0.585*** (0.191)

- -0.546*** (0.200)

-

Male Defendant 1.938*** (0.161)

1.554*** (0.095)

1.951*** (0.132)

1.530*** (0.088)

2.026*** (0.178)

1.610*** (0.118)

1.948*** (0.196)

1.551*** (0.140)

Male Prosecutor 0.421** (0.185)

-0.027 (0.099)

0.515*** (0.187)

0.020 (0.103)

0.480** (0.214)

-0.008 (0.118)

0.352* (0.200)

-0.101 (0.120)

Observations 143,457 143,457 84,614 84,614 54,823 54,823 39,229 39,229 OLS regressions of the defendant’s base offense level. ***: p<0.01; **: p<0.05; *: p<0.10. Standard errors are clustered at the courthouse level. All regressions include demographic controls listed in Table 3, month and year fixed effects, and courthouse-section fixed effects. All regressions restricted to courthouse-sections with at least 25 observations.

  

76

Table A.10. Individual Prosecutor Fixed Effects (1) (2) (3) Dependent Variable: Base Offense Level Interaction -0.596***

(0.167)

-0.458*** (0.153)

-0.513*** (0.176)

Male Defendant 1.940*** (0.182)

1.766*** (0.161)

1.681*** (0.184)

Male Prosecutor 0.449** (0.215)

- -

Prosecutor FEs X Prosecutor-Offense FEs X Observations 127,296 127,296 102,717 OLS regressions of the the defendants base offense level. ***: p<0.01; **: p<0.05; *: p<0.10. Standard errors are clustered at the courthouse level. All regressions include demographic controls listed in Table 3, month and year fixed effects, and courthouse-section fixed effects. All regressions restricted to courthouse-section, prosecutor, or prosecutor-section cells with at least 25 observations.