Diversity in Policing: The Role of O�cer Race andGender in Police-Civilian Interactions in Chicago∗

Bocar Ba† Dean Knox‡ Jonathan Mummolo§ Roman Rivera¶

July 17, 2020

∗Data for this project were gathered in collaboration with Invisible Institute, with the help of Sam Stecklow andEmma Herman. We thank Rachel Ryley and the Lucy Parsons Lab.

†Assistant Professor of Economics, University of California Irvine, bba@uci.edu.‡Assistant Professor of Operations, Information and Decisions, Wharton School, University of Pennsylvania,

dcknox@upenn.edu.§Assistant Professor of Politics and Public A�airs, Princeton University, jmummolo@princeton.edu.¶PhD student, Dept. of Economics at Columbia University, r.g.rivera@columbia.edu.

Abstract

A history of nearly exclusively white male police forces in the U.S. has made diversify-ing personnel one of the oldest and most often proposed police reforms, but data challengeshave precluded micro-level evaluations of its impact. Using newly collected personnel dataand millions of ultra-�ne-grained records on o�cer deployment and behavior, we conduct adetailed quantitative case study of diversity in the Chicago Police Department. We show howo�cers from marginalized groups are consistently assigned to di�erent working conditionsthan white and male o�cers, meaning they typically encounter vastly di�erent circumstancesand civilian behaviors. As a result, coarse agency- or district-level analyses often fall short ofall-else-equal comparisons between o�cer groups, making it di�cult to disentangle o�cers’behavior from the environments in which they work—a crucial �rst step in evaluating thepromise of diversity reforms. To assess behavioral di�erences between o�cers of varyingracial, ethnic and gender pro�les, we leverage detailed records of daily patrol assignments toevaluate o�cers against their counterparts working in the same collections of city blocks, inthe same month and day of week, and at the same time of day. Compared with white o�-cers facing identical conditions, we show Black and Hispanic o�cers both make substantiallyfewer stops, arrests, and use force less often, especially against Black civilians. Much of thegaps in stops and arrests are due to a decreased focus on discretionary contact, such as stopsfor vaguely de�ned “suspicious behavior.” Hispanic and white o�cers exhibit highly similarbehavior toward Hispanic civilians, though Hispanic o�cers who speak Spanish appear tomake fewer arrests in general than those who do not speak Spanish. Within all racial/ethnicgroups, female o�cers are substantially less likely to use force relative to male o�cers. Takentogether, these results show the substantial impact of diversity on police treatment of minor-ity communities, and emphasize the need to consider multiple facets of police o�cers whencrafting personnel-driven reforms.

Racial disparities in police-civilian interactions and high-pro�le incidents of excessive forcecontinue to fuel allegations of abusive and discriminatory policing [2, 24]. Central to these cri-tiques are the fact that throughout the history of policing in the U.S.—from the formation of the�rst organized security patrols on slave plantations [34, 42], to the recent concentrated implemen-tation of aggressive tactics like “Stop, Question and Frisk” in communities of color [14]—manypolice forces in the U.S. have been nearly all white and male [11]. In turn, some of the mostfrequently proposed reforms aimed at reducing inequities and police brutality have centered onhiring more nonwhite [7] and female [10] o�cers. One agency that has undergone substan-tial diversi�cation in recent decades is the Chicago Police Department (CPD), transforming froma virtually homogeneous force to one in which half of sworn o�cers are minorities and overone-�fth are female. This history, combined with unprecedented data availability on the dailygeographic assignment and behavior of o�cers, allows for a thorough assessment of the practi-cal consequences of diversity in law enforcement. In this paper, we examine the Chicago case indepth to provide some of the most credible micro-level evidence to date on how o�cers of var-ied racial, ethnic, and gender pro�les engage in di�ering enforcement patterns when interactingwith civilians.

A central obstacle to rigorous evaluation of personnel reforms in policing has been the lack ofsu�ciently �ne-grained data on o�cer deployment and behavior. Most studies rely on inherentlylimited cross-jurisdiction analyses [e.g. 23, 8, 40], comparing aggregate outcomes like arrests anduses of force between agencies with di�ering levels of diversity. This approach often cannotdistinguish whether apparent di�erences in o�cer behavior arise due to selection (e.g. agenciesthat choose to diversify may di�er on unmeasured traits) or ecological fallacies (e.g. white andnonwhite o�cers within jurisdictions are assigned to work in di�erent circumstances—which wedemonstrate is in fact true). While strategies like agency-level pre-/post-reform comparisons havebeen employed to combat these sources of bias [27, 28, 12, 16], these aggregated analyses maskdetails of police-civilian encounters, precluding explorations of important sources of heterogene-ity. In cases where researchers have been able to obtain micro-level data on o�cers, analyses areoften limited in scope, e.g. tra�c accidents [46] or 911 calls [19, 45]; rely on di�cult-to-validatesurvey responses [33, 47]; or otherwise lack necessary data to make reliable inferences. Specif-ically, many studies of o�cer race and gender that rely on administrative data examine recordsof enforcement activities only, e.g. logs of pedestrian stops or arrests [3, 9, 20, 35, 44]. Theseapproaches omit data from shifts in which o�cers take no such actions—introducing potentiallysevere selection bias [17, 18, 21]. Using available (but often limited) data, prior work has shownsuggestive evidence that female o�cers are less likely to use force [e.g. 25, 32] though conclusionsvary [20]. Findings with regard to racial diversity have been decidedly mixed: in an exhaustivereview of the empirical literature on racial diversity in policing, one prominent legal scholar con-

2

cluded: “[t]he fairest summary of the evidence is probably that we simply do not know” [39,1225].

In this paper, we draw on a host of newly collected datasets that together allow us to over-come these longstanding limitations. Our data, which include personnel records covering roughly30,000 Chicago police o�cers, were assembled through years of open records requests and law-suits; they include o�cer demographics, language skills, shift assignments and career progres-sion. We use these records to paint a detailed picture of the diversi�cation of the departmentand how it has deployed o�cers from marginalized groups. We �nd substantial variation in de-ployment patterns across o�cers of di�erent races and genders.1 For example, Black o�cers aremore often deployed in high-crime areas, and female o�cers tend to work during di�erent timesof day. These di�erential deployment patterns highlight a source of confounding present in manyagency-level analyses of personnel reforms that �ner-grained data can address.

We link these personnel �les to timestamped, geolocated records of o�cers’ decisions to stop,arrest, and use force against civilians. After aggressively pruning these data to maximize the va-lidity of our analyses, we are left with a panel containing 2.9 million o�cer-shifts and 1.6 milliondetailed observations of enforcement events. Most importantly, we leverage �ne-grained infor-mation on o�cers’ daily patrol assignments—the speci�c times and beats (roughly, collectionsof city blocks) that they are assigned to patrol—to examine how o�cers of di�erent groups be-have when faced with identical circumstances and civilian behaviors. Crucially, this data shedslight not only on when o�cers take enforcement action against civilians, but also accounts fordecisions not to take action—thus avoiding a major source of selection bias.

We caution that our analysis does not directly estimate the future e�ects of hiring more di-verse o�cers, for several reasons. Chief among these are (i) the nature of police-civilian inter-actions is changing rapidly; (ii) racial/ethnic/gender di�erences in current o�cers’ behavior maynot map perfectly to those of future cohorts; (iii) deployment patterns will necessarily change asmore o�cers are hired from marginalized groups; and (iv) diversi�cation reforms may exert pow-erful spillover e�ects, e.g. through agency culture. Nevertheless, what we do show—enormousgaps in current o�cers’ enforcement patterns, in ways that are systematically associated withtheir personal backgrounds—is a critical step toward evaluating the promise of diversity reforms,particularly given the current dearth of detailed quantitative evidence on the issue. If o�cersof di�erent demographic pro�les do not perform di�erently when faced with identical circum-stances, then there is little hope that diversifying police agencies will yield tangible di�erencesin the treatment of marginalized civilians.2

1It is unclear whether Chicago consistently di�erentiates between sex and gender in administrative data. Weuse the term “gender” throughout in keeping with o�cer personnel �les we received in response to open recordsrequests.

2However, as we discuss below, diversity can potentially yield intangible gains like trust in government [13, 43].

3

It is important to note these behavioral gaps are not the “causal e�ect” of manipulating an o�-cer’s race: minority and female o�cers di�er from their counterparts in many unmeasured ways.Our focus is on estimating a far more policy-relevant quantity: the di�erence in performance thatcan be expected when deploying available o�cers of one group, relative to another group. Thisquantity is much more feasibly estimated because, as we explain below, it only requires adjustingfor external environmental factors, like the civilian behaviors faced by o�cers—not netting outinternal o�cer traits.

Our results highlight the need for much greater collection of policing data nationwide, par-ticularly with regard to deployment times and locations, if analysts are to rigorously evaluateo�cer behavior while holding environmental factors �xed. Our �ndings suggest deploying morewomen and o�cers of color, relative to white males, would reduce the amount of physical forceand enforcement activity endured by civilians, especially Black civilians. At a high level, we �nd:

(1) Minority o�cers are assigned to very di�erent districts, and even within districts, receivevastly di�erent geographic and temporal patrol assignments. Without accounting for these dif-ferences in working conditions, our work shows there is no way to meaningfully characterizethe di�erences in behavior between white/minority and male/female o�cers. These limitationsto traditional research designs suggest a need for greater transparency by police departments toenable careful evaluation of o�cer behavior.

(2) Compared to white o�cers working in the same places and times, Black o�cers make sig-ni�cantly fewer stops and arrests, and they use force less often. These di�erences are substantial,equivalent to 31%, 22%, and 35% of typical white-o�cer activity, respectively. Examining a widerange of o�cer activity, we show this is mostly driven by a decrease in discretionary activity (e.g.38% of the reduction in stops are due to a reduced focus on vaguely de�ned “suspicious behav-ior”), rather than lower enforcement of severe criminal behavior (decreased violent-crime arrestsaccount for only 12% of the overall Black-white gap in arrest volume). Moreover, higher levelsof enforcement by white o�cers fall primarily on marginalized groups: approximately 80% ofthe gap in o�cer stops, arrests, and uses of force is due to di�ering treatment of Black civilians.Hispanic o�cers display similar reductions in enforcement activity, relative to white o�cers, butto a lesser extent than their Black counterparts. In addition, we �nd suggestive evidence that His-panic o�cers who speak Spanish make fewer arrests than their non-Spanish-speaking counter-parts, underscoring an additional and previously under-explored, but potentially consequential,o�cer attribute.

(3) Relative to male o�cers facing identical circumstances, female o�cers make somewhatfewer arrests (di�erences equivalent to 7% of typical male arrest rates) and use force dramaticallyless (31%); moreover, over 80% of this gap is due to a reduced focus on Black civilians. This holdstrue even when comparing within racial/ethnic groups: across the board, Black, Hispanic, and

4

white policewomen use force less often than their co-racial/co-ethnic male counterparts, withreductions that are concentrated nearly entirely among interactions with nonwhite civilians.

Overall, our results suggest that deploying o�cers of various races and genders yields sub-stantial di�erences in police treatment of civilians. However, our �ndings also suggest the need toconsider multiple facets of police o�cers when crafting personnel-driven reforms. Like the civil-ians they police, o�cers are complex, not easily reducible to single demographic traits. Scholarsand reformers must take stock of the whole o�cer when considering the impact of diversityinitiatives on police-civilian interactions.

All data and interactive replication materials are publicly available at [LINK]. We encouragereaders to probe and extend our analyses.

1 The Case of the Chicago Police Department

Our analysis focuses exclusively on records from a single city, a feature that a�ords us unusuallydetailed data at the expense of geographic scope. Chicago represents a valuable opportunity toevaluate the e�ects of deploying diverse o�cers. It is a large and racially diverse metropolis, withmore than half the city’s population identifying as nonwhite. The police force has also diversi�edin recent decades, with roughly 22% of o�cers identifying as Black, 23% Hispanic, and 3% Asian.3

The agency’s o�cers are currently 22% female, a stark change from its 99% male composition in1970. The city is also heavily racially segregated,4 has a history of racial tensions between resi-dents and police, and has come under scrutiny in recent years for applying a range of controversialaggressive policing tactics such as “Stop and Frisk” on a wide scale. The agency received nationalattention for the 2014 killing of seventeen-year-old Laquan McDonald, an attempted cover-up,and ensuing social unrest. [6]. The department was condemned for its “code of silence,” [30] andthen-superintendent Garry McCarthy received widespread criticism for “encouraging the kind ofaggressive cop culture under which McDonald’s shooting took place," [26].

In some ways, this history makes Chicago unique, and to the extent these events contributedto disparities in the behavior of di�erent o�cer groups, it may be di�cult to extrapolate to othersettings in which racial tensions are less pronounced. But in other ways, it is the very existenceof these problems that makes Chicago such an important test case for proposed reforms: amongthe major departments, it is arguably the one in which reform has historically been most sorelyneeded. A single case study cannot hope to be the �nal word in an important, ongoing debate.But Chicago o�ers an invaluable opportunity to examine the role of diversity in policing using

3These �gures describe the racial distribution of the CPD in 2016 according to our personnel records; see Figure 1for additional details.

4See http://www.censusscope.org/us/print_rank_dissimilarity_white_black.html

5

unusually �ne-grained data, in a setting where concerns over racial inequity were pronounced,while o�ering a template for future scholars to conduct similar analyses elsewhere.

1.1 Data on Agency Personnel

To understand the history of how the CPD diversi�ed over time, we rely on both published ac-counts of department history and newly acquired quantitative data. Over a period of three years,we submitted open records requests to the CPD and the city’s Department of Human Resourcesseeking data on o�cer demographics and behavior. The resulting records include the name, race,gender, birth year, salary, language skill, unit assignments and appointment date of each o�cer[5, 36]. These newly acquired records allow us to reconstruct the history of CPD diversi�cationover a much longer period than previously possible. While the CPD intermittently publishesannual reports with aggregate demographics, these data cover only 1995–2010 and 2016–2017.5

Using our newly acquired records, we extend the time-series backward to 1970, allowing for acomprehensive descriptive portrait of the evolution of the demographic correspondence betweenCPD personnel and city residents.6

Figure 1 shows how the CPD, which currently employs about 12,000 sworn o�cers, slowlydiversi�ed over time. As [27] recounts, before a series of lawsuits, the CPD was slightly less thanthan 20% Black, in a city that was one-third Black in 1970. In the early 1970s, the Afro-AmericanPatrolmen’s League (AAPL) �led a discrimination suit against the CPD “on hiring, promotion,assignment, and discipline” [31], with the DOJ soon joining them [27]. In 1974, hiring quotas wereimposed, and Black hiring shares increased from 10% to 40% by 1975—though CPD compositionchanged more slowly due to low turnover [27]. These reforms also altered the gender compositionof the department within racial groups and across ranks; women made up a larger proportion ofBlack recruits, initially causing white women to trail Black women in new hiring and promotions[1, 92].

While the CPD has made strides in moving its demographic pro�le toward parity with thecity’s, there exists substantial divergence in working conditions. Figure 2 displays the averagecharacteristics of the districts—the 22 geographic regions delineated by the CPD—to which o�cergroups are assigned. While the �gure shows only minimal gendered variation in district assign-ments, the di�erences associated with o�cer race/ethnicity are stark. Black o�cers are assignedto districts with roughly 50% higher rates of violent crime, and perhaps most strikingly, Blacko�cers are assigned to districts with large co-racial resident populations—on average nearly 75%co-racial, far higher than the average 25–30% co-racial/co-ethnic districts where white and His-

5The CPD Annual Reports from 1966 to 1969 included counts of “police women” and “police matrons,” but re-porting on these statistics was discontinued after 1969.

6See SI Section A.2 for details on the coding on race/ethnicity.

6

Figure 1: Composition of CPD o�cers and city residents over time. Red, blue, and black ◦respectively depict the proportion of Black, Hispanic, and white active CPD o�cers in Decemberof each year, according to our personnel records. Dark and light gray regions respectively indicatethe proportion of female and male o�cers, using the same data. Data from CPD annual reports onthe demographics of sworn and exempt/command o�cers are available only for 1995–2010, 2016and 2017 (not shown); these are shown with ×. When available, these reports closely track ourpersonnel data and increase con�dence in our historical reconstructions. Lines indicating cityof Chicago decennial Census proportions for each racial/ethnic group, tabulated by the NationalHistorical Geographic Information System, are shown with ■ for reference.

0.00

0.25

0.50

0.75

1.00

1980 1990 2000 2010

Pro

port

ion

Group Chicago pop. CPD personnel records CPD reports

Race/Ethnicity Black Hispanic White

Gender Female Male

7

Figure 2: Characteristics of assigned districts. Average features of assigned geographic unitsare plotted for various o�cer groups, based on 1,089,707 monthly assignments from 2006–2016.Con�dence intervals cluster on district-month.

Violent crimesper 1,000 residents

per month

Property crimesper 1,000 residents

per month

Police officersper 1,000 residents

Proportion ofco-racial residents

Black

fem

ale

Black

male

Hispan

ic fe

male

Hispan

ic m

ale

Whi

te fe

male

Whi

te m

ale

Black

fem

ale

Black

male

Hispan

ic fe

male

Hispan

ic m

ale

Whi

te fe

male

Whi

te m

ale

Black

fem

ale

Black

male

Hispan

ic fe

male

Hispan

ic m

ale

Whi

te fe

male

Whi

te m

ale

Black

fem

ale

Black

male

Hispan

ic fe

male

Hispan

ic m

ale

Whi

te fe

male

Whi

te m

ale

0.00

0.25

0.50

0.75

0

2

4

6

0

2

4

6

0.0

0.5

1.0

1.5

2.0

Officer race and gender

Avera

ge c

hara

cteri

stic

s of

ass

igned

dis

tric

ts

panic o�cers serve. (SI Appendix B.1 contains a detailed discussion how districts are de�ned andanalysis of additional district characteristics. SI Figure 2 presents average district characteris-tics disaggregated by year. SI Figure B.2 shows that the demographic pro�le of assigned o�cerssomewhat tracks that of district residents, but o�cers tend to be disproportionately white.)

In supplemental analyses (SI Figures B.4–B.5) we also show considerable variation in assignedshift times. For example, 45% of standard shifts served by female o�cers are on third watch (4p.m. to midnight), compared to 55% for male o�cers. Similarly, 43% of standard shifts servedby Black o�cers are third watch, compared to 55% of white-o�cer shifts and 60% for Hispanico�cers. Moreover, Figures B.6–B.7 demonstrate that marginalized groups are tasked with pa-trolling di�erent sets of beats, compared to white or male o�cers serving in the same district.(All p-values < 0.001.)

These patterns underscore a central di�culty in testing whether o�cers with di�erent de-mographic pro�les perform their duties di�erently. Namely, white o�cers work in di�erent en-vironments from Black o�cers, on average (especially with regard to local racial composition),and men and women work during di�erent hours of the day. This means any inferred di�erencesin o�cer behavior that rely on data aggregated to large geographic units and time periods—suchas district-months or agency-years, the analytic strategy in much prior work—may simply re-�ect di�ering patrol environments, rather than di�erences in policing approaches. Fine-grained

8

assignment data overcomes this central obstacle to inference.

2 Data on O�cer Behavior

Our data also contain detailed data on o�cers’ stops, arrests and uses of force against civilians.We merge these with records of o�cers’ daily patrol assignments7 and U.S. Census data. Togetherthese data provide an unprecedentedly detailed view of the day-to-day behavior of o�cers in amajor U.S. law enforcement agency over an extended period. Table 1 contains aggregate infor-mation on stops, arrests, and uses of force from Jan. 2012 to Dec. 2015 (the period covered byour behavioral analysis) by o�cer group. Due to the small sample sizes associated with groupsincluding Asian Americans and Native Americans, our analysis is limited to Black, Hispanic, andwhite o�cers (together accounting for roughly 97% of o�cers for whom shift data is available).

Figure 3 illustrates the dataset’s various attributes by highlighting a small slice of its temporaland geographic coverage. The �gure maps activity during a four-month window in the CPD’sWentworth District (District 2), a highly segregated 7.5 square mile territory on Chicago’s SouthSide that is 96% Black and consistently ranks among the city’s most violent districts in per-capitacrime rates. The district is comprised of 15 patrol areas8 which are shaded according to their racialcomposition. Points distinguish between geolocated stops, arrests, and uses of force during thisperiod. The �gure also o�ers a detailed portrait of four anonymized CPD o�cers working inDistrict 2 in this time. For example, “O�cer A” is female, Hispanic, speaks both English andSpanish, and was born in 1965; “O�cer D” is a white male born in 1981 who does not speakSpanish. The �gure shows the speci�c beats to which o�cers were assigned over time and eacho�cer’s behavior while working in these beats.

7We are grateful to Lucy Parsons Labs for publicly releasing data on civilian stops and Rachel Ryley for generouslysharing data on beat assignments.

8Based on the CPD 2008 Annual Report.

9

Figu

re3:

Detailedview

oftheda

ta.T

herig

htpa

nel�

gure

map

spol

ice

activ

ityin

asin

gle

polic

edi

stric

t(W

entw

orth

,CPD

Dist

rict

2),w

ithgr

een

circ

les,

blue

squa

res,

and

red

cros

sesr

espe

ctiv

ely

indi

catin

gth

elo

catio

nsof

stop

s,ar

rest

s,an

dus

esof

forc

e.Po

lygo

nsre

pres

entg

eogr

aphi

cbe

ats

and

are

shad

edby

thei

rpr

opor

tion

ofm

inor

ityre

siden

ts.

Low

erle

ftpa

nels

char

tthe

beha

vior

offo

uran

onym

ized

o�ce

rsov

era

four

-mon

thpe

riod,

with

pane

lhea

ders

indi

catin

gag

e,ge

nder

,eth

nici

ty/r

ace,

and

lang

uage

abili

ty.E

n-ci

rcle

din

cide

ntsa

rede

scrib

edfu

rther

inth

ele

ftm

iddl

epa

nel,

whi

chre

port

civi

lian

and

inci

dent

spec

i�cs

.Fin

ally

,the

top

left

pane

lin

dica

tesh

owth

efou

rsel

ecte

do�

cers

area

ssig

ned

topa

trolb

eats

over

date

sand

times

,with

verti

calg

ray

bars

indi

catin

gw

eeke

nds.

10

Table 1: Summary of data on o�cer behavior (counts), 2012-2015. Summary statistics arereported after pruning o�cers, shifts, and event records aggressively to ensure common circum-stances in our behavioral analysis, as described in the main text.

Black Hispanic White Female Maleo�cers o�cers o�cers o�cers o�cers

Stops 253,609 356,541 729,078 264,552 1,074,676Arrests 47,406 65,587 132,285 43,630 201,648

Uses of force 1,355 2,081 4,514 1,125 6,825Shifts 830,062 689,239 1,413,977 740,240 2,193,038

O�cers 1,835 1,674 3,440 1,786 5,163

3 Research Design: Assessing the Impact of Deployment

These granular data permit precise comparisons between o�cers facing identical working con-ditions. We assemble a panel dataset in which each row represents an o�cer-shift—an assignedeight-hour patrol period—and contains detailed information on the o�cer’s actions and their con-text.9 In each of these 2.9 million patrol assignments, we compute the o�cer’s pro�le of stops,arrests, and uses of force. O�cers of di�erent demographic pro�les are then compared to peersworking in the same unique combination of month, day of week, shift time, and beat (roughly asmall collection of city blocks, averaging 0.82 square miles citywide)—a narrow slice of time andspace that we abbreviate “MDSBs.”

By making these precise comparisons, we ensure observed di�erences in o�cer behavior arenot due to disparities in external conditions. That is, we can safely assume all o�cers in anMDSB face the same set of average enforcement opportunities, civilian activity, neighborhoodattributes such as infrastructure and architecture, and time-varying conditions such as weatherand lighting. Put di�erently, this strategy ensures that the rates of various enforcement activitiesbeing compared across o�cers have the same denominators [15, 29]. We term this the “commoncircumstances” assumption. (For a detailed discussion of potential threats to this assumption, seeSI Appendix C.3). To further enhance the credibility of the common-circumstances assumption,we additionally subset to individuals for whom personnel data could be merged. Only those ofrank “police o�cer” are retained, eliminating sergeants and other higher-ranked o�cers thatmay not patrol as regularly. Non-standard shifts are dropped (retaining only �rst through thirdwatches), as are absences, shifts with assigned special duties (e.g. protest detail, station security,training), and unusual double- or triple-duty days in which a single o�cer serves multiple shifts.

9In analyzing o�cer behavior, we take patrol assignments—the level at which commanding o�cers typicallyexercise control—as the basic unit of analysis. In SI Section C.4, we examine the possibility that di�erences in clock-in/out times may be one mechanism behind observed di�erences in behavior between o�cer groups. While somestatistically signi�cant di�erences can be observed (roughly 0.1% disparities in patrol time), these gaps are two ordersof magnitude smaller than the di�erences observed in stops, arrests, and uses of force that make up our main results.

11

Events occurring while the o�cer was o� duty are also eliminated. SI Appendix A.1 describesthese datasets and our preprocessing procedures in detail.

For ease of interpretation, we present di�erences estimated using ordinary least squares withMDSB �xed e�ects, though our results are robust to several other estimators, including the ad-dition of �exible controls for experience (see SI Appendix C.2 for a discussion of estimation andSI Figures C.1-C.3 for these additional results). All statistical inferences are based on o�cer-levelblock bootstrap con�dence intervals that are robust to unobserved o�cer-speci�c peculiarities.Our behavioral analysis organizes patrol assignments into over 650,000 groups de�ned by MDSBs.Of these, roughly 300,000 allow for comparisons between o�cers of di�ering racial/ethnic groups(and thus contribute to our analyses); 230,000 contain both female and male o�cers; and 50,000contain Hispanic o�cers of di�ering language abilities. For details, see SI Appendix C.5.

Importantly, our assignment records allow tracking of o�cer-shifts whether or not they en-gaged in any enforcement activity. Here, our analysis di�ers from most prior studies of o�cerrace and gender using police administrative data [e.g. 3, 9, 20, 35, 44], which only rely on in-stances in which o�cers recorded an activity (e.g. a stop or arrest; see discussion in [21]). (Seealso [19], [45], and [46] for other strategies to combat the resulting selection bias.) Without thesepatrol assignment records, inaction is invisible to the analyst, and o�cers who do not recordany enforcement activity simply vanish from the data. In addition, as we show below, o�cers ofdi�erent races and ethnicities are assigned to work in systematically di�erent conditions. Patrolassignment data allows us to make apples-to-apples comparisons by adjusting for this fact.10

We caution that our approach is only partially informative about mechanisms. While we ob-serve that white and nonwhite o�cers behave di�erently under common circumstances, largelydriven by di�erential focus on low-level crimes, we cannot discern the psychological pathwaysor other channels through which these di�erences arise. But these results nonetheless help eval-uate the promise of proposed personnel reforms: they show what behavior can be expected whendeploying o�cers of a given demographic pro�le are deployed, on average, holding environmentalfactors �xed. If we cannot discern disparities in behavior across these o�cer groups, diversityreforms are unlikely to meaningfully alter the volume and character of policing. While investiga-tion of mechanisms are an important direction for future work, given the urgent need for evidenceon police reform e�cacy, careful estimation of overall e�ects is our paramount concern.

10Our data indicate that two o�cers were often involved in stopping or arresting a civilian. Throughout, suchinstances enter as separate rows in our o�cer-shift data, since each o�cer contributed to enforcement. In the stopdata, one is listed as “�rst” o�cer, potentially indicating a leading role, while in the arrest data no such labelingexists. We therefore conduct a robustness check using the stop data in SI Appendix C.8 that reanalyzes stops afterrestricting to �rst o�cer only. With minor exceptions, this yields nearly identical results.

12

4 Results

We now turn to whether deploying o�cers of di�erent demographic pro�les a�ects the volumeand nature of police-civilian interactions. Table 2 displays the average di�erences in the numberof stops, arrests and uses of force associated with Black and Hispanic o�cers (relative to whiteo�cers) and female o�cers (relative to male) working in the same MDSBs. We also note thatroughly 77% of Hispanic residents in the Chicago metropolitan area speak Spanish at home as of2015 [22]. However, our personnel data show less than half of Hispanic CPD o�cers can speakSpanish. Because of the potential for this language barrier to impede improved police-civilianinteractions within this ethnic group, we therefore also compare Hispanic o�cers who can andcannot speak Spanish.

Turning �rst to Black o�cers, we see that when faced with the same working conditions, thisgroup makes 1.52 fewer stops per 10 shifts, makes 1.94 fewer arrests per 100 shifts, and uses force1.02 fewer times per 1,000 shifts, on average, than white counterparts—that is, compared to whiteo�cers assigned to patrol the same beat, in the same month, on the same day of the week, andat the same shift time (all padj < 0.001 after Benjamini-Hochberg multiple-testing correction forcomparisons reported in Table 2). These gaps are large, representing reductions of 29%, 21%, and32% relative to typical stop, arrest, and use-of-force volume for white o�cers (see SI Table C.1 foraverage behavior by o�cer race).

Importantly, Table 2 also shows these disparities are not uniform across situations. Rather,they are driven by a reduced focus on engaging Black civilians (1.26 fewer stops per 10 shifts and1.46 fewer arrests per 100 shifts, 39% and 25% of typical white o�cer behavior) and a broad classof enforcement activities that can be thought of as relatively discretionary, including stops ofcivilians for “suspicious behavior” (-0.57 per 10 shifts, 31% less) or arrests for drug-related crimes(-0.31 per 100 shifts, 27% less). However, when it comes to enforcing violent crime, Black o�cerbehavior looks similar to that of white o�cers, with Black o�cers making only 0.22 fewer arrestsper 100 shifts for violent crimes than whites (only 11% less). In other words, when it comes topolicing the most serious o�enses, Black o�cers are roughly comparable to white o�cers. Blacko�cers also deploy force against Black civilians 0.85 fewer times per 1,000 shifts (38% less) thantheir white counterparts. (All adjusted p-values < 0.001.) In fact, reduced use of force against thiscivilian group accounts for 83% of the overall force disparity between white and Black o�cers. (SISection C.7 and SI Figures C.1–C.3 show that results are virtually identical using a wide range ofalternative estimators.) This pattern of results is remarkably in line with the hopes of proponentsof racial diversi�cation, who seek to reduce abusive policing and mass incarceration, especiallyin Black communities.

We also �nd substantial di�erences in the behavior of female o�cers relative to male o�cers.

13

Relative to male o�cers working on the same places and times, female o�cers make 0.61 fewertotal arrests per 100 shifts (-7% relative to typical male behavior) and 0.54 fewer arrests of Blackcivilians per 100 shifts (-9%, both padj. < 0.001). In fact, 89% of this disparity in arrest rate is due toreduced arrests of Black civilians. We also �nd that female o�cers use force 0.87 less times overall(-28%) and 0.71 fewer times per 1,000 shifts against Black civilians (-31%, both padj. < 0.001), withthe latter accounting for 82% of overall force reduction. (Appendix C.6 and SI Table C.4 shows thatthese gender di�erences are not speci�c to any racial/ethnic group: within each group, femaleo�cers use signi�cantly less force than their male counterparts.)

Results di�er in important ways for Hispanic o�cers. Like their Black colleagues, Hispanico�cers facing the same working conditions conduct fewer stops, arrests and uses of force thanwhite o�cers. However, these di�erences are far more modest. Strikingly, gaps are primarilydriven by less engagement with Black civilians, while Hispanic o�cers exhibit nearly the samevolume of enforcement activity against co-ethnic civilians as white o�cers, on average. Ourestimates indicate that Hispanic o�cers can be expected to make 0.28 fewer stops per 10 shifts(6% reduction relative to typical white o�cer behavior, padj. < 0.001); 0.44 fewer arrests per 100shifts (padj. = 0.012, 5% reduction); and 0.37 fewer uses of force per 1,000 shifts (padj. = 0.017,12% reduction) per shift. As the table shows, part of this heterogeneity is driven by Spanish-speaking Hispanic o�cers, who make 0.74 fewer arrests per 100 shifts on average than Hispanico�cers who do not speak the native language of many co-ethnic communities (6% less, padj. =0.021; however, SI Appendix C.2 shows that unlike all other results, this result is not robust tothe inclusion of �exible controls for o�cer experience). When it comes to stops and uses offorce, di�erences in the behavior of Spanish and non-Spanish-speaking Hispanic o�cers appearnegligible. However, we caution that null results may in part be due to the relatively small sizes ofthese language groups, which dramatically reduce overlap between Spanish- and non-Spanish-speaking Hispanic o�cers in MDSBs. An improved understanding of language and other culturalfactors remains an important direction for future work.

5 Discussion and Conclusion

Fatal encounters between white police o�cers and unarmed racial minorities continue to promptwidespread calls for law enforcement reforms. Personnel reforms are prominent among theseproposals, but evidence on their e�cacy has been limited by a severe scarcity of data. Usingunusually rich data on police activity in Chicago, we provide detailed micro-level evidence thatsheds light on the promise of diversity initiatives for reshaping police-civilian interactions.

Our results also underscore several previously unknown nuances and avenues for further ex-ploration. Faced with the identical working conditions, Black o�cers are less likely to stop, arrest,

14

Tabl

e2:O

�cers

ofdi�eringde

mog

raph

icpr

o�lesb

ehav

edi�eren

tlywhe

nfacing

common

circum

stan

ces.

Each

row

repo

rtsac

ompa

rison

betw

een

o�ce

rsof

vario

usde

mog

raph

icgr

oups

.Eac

hco

lum

nco

ntai

nsav

erag

eper

-shi

ftdi

�ere

nces

inap

olic

ebeh

avio

ram

ong

o�ce

rsas

signe

dto

the

sam

ebe

atan

dsh

ifttim

e,du

ring

the

sam

em

onth

and

day

ofw

eek.

Aste

risks

indi

cate

Benj

amin

i-H

ochb

ergp-

valu

esac

coun

ting

foro

�ce

r-le

velc

lust

erin

gan

dad

just

edfo

rmul

tiple

test

ing

of88

hypo

thes

es:∗

is<0.05

,∗∗

is<0.01

,an

d∗∗∗

is<0.001.

Alte

rnat

ive

estim

ator

s,in

clud

ing

adju

stin

gfo

ro�

cere

xper

ienc

e,ar

ere

porte

din

SIFi

gure

sC.1–

C.3

and

prod

uce

subs

tant

ivel

yid

entic

alre

sults

.

Stop

spe

r10

shifts

Civi

lian

race

Reas

onBl

ack

Hisp

anic

Whi

teO

ther

Loite

ring

Susp

icio

usD

rug

Tra�

cTo

tal

Blac

ko�

cers

(vsw

hite

)-1

.26∗∗∗

-0.13

∗∗∗

-0.13

∗∗∗

-0.25

∗∗∗

-0.03

∗∗∗

-0.57

∗∗∗

-0.17

∗∗∗

-0.50

∗∗∗

-1.52

∗∗∗

Hisp

anic

o�ce

rs(v

swhi

te)

-0.29

∗∗∗

0.05

-0.04

∗-0

.020.0

0-0

.17∗∗∗

-0.05

∗∗-0

.05-0

.28∗∗∗

Fem

ale

o�ce

rs(v

smal

e)-0

.03-0

.010.0

20.1

6∗∗∗

-0.01

-0.16

∗∗∗

-0.07

∗∗∗

0.07

-0.01

Span

ish-s

peak

ing

Hisp

anic

o�ce

rs0.0

30.0

1-0

.010.0

7-0

.02-0

.06-0

.030.0

80.0

4(v

snon

-Spa

nish

-spe

akin

g)

Arrests

per10

0sh

ifts

Civi

lian

race

:Re

ason

:Bl

ack

Hisp

anic

Whi

teO

ther

Dru

gTr

a�c

Prop

erty

Viol

ent

Tota

lBl

ack

o�ce

rs(v

swhi

te)

-1.46

∗∗∗

-0.29

∗∗∗

-0.18

∗∗∗

-0.93

∗∗∗

-0.31

∗∗∗

-0.18

∗∗∗

-0.29

∗∗∗

-0.23

∗∗∗

-1.94

∗∗∗

Hisp

anic

o�ce

rs(v

swhi

te)

-0.30

∗-0

.08-0

.08∗

-0.32

∗∗0.0

3-0

.08∗

-0.08

∗0.0

2-0

.44∗

Fem

ale

o�ce

rs(v

smal

e)-0

.54∗∗∗

-0.09

0.02

-0.28

∗∗∗

-0.17

∗∗∗

-0.02

-0.02

-0.12

∗∗-0

.61∗∗∗

Span

ish-s

peak

ing

Hisp

anic

o�ce

rs-0

.48∗

-0.20

-0.04

-0.17

-0.14

-0.13

-0.19

∗-0

.12-0

.74∗

(vsn

on-S

pani

sh-s

peak

ing)

Usesof

forcepe

r1,00

0sh

ifts

Civi

lian

race

Blac

kH

ispan

icW

hite

Tota

lBl

ack

o�ce

rs(v

swhi

te)

-0.85

∗∗∗

-0.10

∗-0

.05-1

.02∗∗∗

Hisp

anic

o�ce

rs(v

swhi

te)

-0.33

∗0.0

3-0

.06-0

.37∗

Fem

ale

o�ce

rs(v

smal

e)-0

.71∗∗∗

-0.13

∗∗-0

.00-0

.87∗∗∗

Span

ish-s

peak

ing

Hisp

anic

o�ce

rs-0

.020.2

00.2

10.3

8(v

snon

-Spa

nish

-spe

akin

g)

15

and use force against civilians, especially Black civilians, relative to white o�cers. These dispari-ties are driven by a reduced focus on the enforcement of discretionary stops and arrests for pettycrimes, including drug o�enses, which have long been thought to fuel mass incarceration [2]. Incontrast to these drastic di�erences in the policing of petty crime, Black o�cers’ enforcement ofviolent crime is only slightly lower than that of white o�cers.

But this pattern, which at �rst glance closely comports with the hopes of racial diversi�cationadvocates, is complicated by several additional results. For one, while Hispanic o�cers displaylower levels of enforcement activity than whites overall, their behavior toward Hispanic civiliansis broadly comparable to that of white o�cers. However, we �nd some evidence that after ac-counting for language ability, Spanish-speaking Hispanic o�cers make fewer arrests than theirnon-Spanish-speaking counterparts. We also �nd substantial di�erences in the behavior of fe-male o�cers—both relative to male o�cers generally and within racial and ethnic groups—withthe most substantial di�erences pertaining to use of force. The vast majority of these genderedreductions are driven by a reduced focus by female o�cers on arresting and using force againstBlack civilians. These results have important implications for the potential of personnel reforms.

This study has several limitations. First, we analyze data from a single city, allowing for an un-usually detailed analysis at some cost in generalizability. Unfortunately, the patchwork of roughly18,000 police agencies in the U.S., with its varying composition, rules and practices—combinedwith nonstandard data collection and sharing practices [15]—makes such precise evaluationsvirtually impossible in most jurisdictions. As additional similar data become available, patrol-assignment analyses o�er a useful template for other scholars to follow when testing whetherthese patterns hold in other places and times. A second limitation is that administrative dataalone can only partially identify the mechanisms connecting o�cer race and/or gender to behav-ior. While we demonstrate that much of these gaps in enforcement patterns arise from a reductionin discretionary policing and use of force, e.g. for drug o�enses, our data are unable to speak tothe psychological or other processes behind these observed disparities.

The e�ects we demonstrate strongly suggest that diversi�cation holds promise for reshapingpolice-civilian encounters. But retrospective estimates from one place and time can only o�era rough sense of the potential bene�ts of diversity initiatives nationwide. The extrapolation ofour results, even to future Chicago hiring, hinges on factors such as whether CPD hires froma comparable pool of potential employees. The nature of policing continues to evolve rapidly,and a complete understanding of the e�cacy of proposed reforms requires continued, in-depthresearch. As o�cers from marginalized communities join police forces in increasing numbers,their presence will necessarily lead to shifts in personnel deployment and department norms. Asagencies diversify, institutional culture may also begin to shift and exert its own e�ects on o�cerbehavior—an important long-run implication that is beyond the scope of our study. In addition,

16

while our analysis tests for tangible e�ects of diversi�cation, we emphasize that a large literatureon descriptive representation [4, 41, 13] suggests diversifying police agencies may yield intangiblebene�ts, namely, increased trust in government among historically underrepresented groups [43].If racial diversity a�ects levels of public trust, the cost-bene�t calculus when considering this classof reforms would be further complicated.

The e�ects of racial diversi�cation are likely neither simple nor monolithic. Like the civiliansthey serve, o�cers are multidimensional. Crafting e�ective personnel reforms requires thinkingbeyond the coarse demographic categories typically used in diversity initiatives, and considera-tion of how multiple o�cer attributes relate police to the civilians they serve.

References

[1] Adams, Megan Marie. 2012. The Patrolmen’s Revolt: Chicago Police and the Labor andUrban Crises of the Late Twentieth Century PhD thesis UC Berkeley.URL: https://escholarship.org/uc/item/9gp045vr

[2] Alexander, Michelle. 2010. The New Jim Crow: Mass Incarceration in the Age of Colorblind-ness. New York: The New Press.

[3] Antonovics, Kate and Brian G. Knight. 2009. “A new look at racial pro�ling: Evidence fromthe Boston Police Department.” The Review of Economics and Statistics 91(1).

[4] Arnold, R. Douglas. 1990. The Logic of Congressional Action. Yale University Press.

[5] Ba, Bocar A. 2019. “Going the Extra Mile: the Cost of Complaint Filing, Accountability, andLaw Enforcement Outcomes in Chicago.” Working paper .

[6] Byrne, John, Rick Pearson and Jeremy Gorner. 2018. “Mayor Rahm Emanuelrejects President Donald Trump’s call for Chicago police to use stop-and-frisktactics.” Working paper . https://www.chicagotribune.com/nation-world/

ct-trump-chicago-shootings-20181008-story.html.

[7] DOJ/EEOC. 2016. “Advancing Diversity in Law Enforcement.”. https://www.justice.

gov/crt/case-document/file/900761/download.

[8] Donohue III, John J. and Steven D. Levitt. 2001. “The impact of race on policing and arrests.”The Journal of Law and Economics 44(2):367–394.

[9] Fagan, Je�rey, Anthony A. Braga Rod K. Brunson and April Pattavina. 2016. “Stops andstares: Street stops, surveillance, and race in the new policing.” Fordham Urb. LJ 43:539–614.

17

[10] Fantz, Ashley. 2020. “Want to reform the police? Hire more women.” CNN . https:

//www.cnn.com/2020/06/23/us/protests-police-reform-women-policing-invs/

index.html.

[11] Forman Jr., James. 2017. Locking up our own: Crime and punishment in Black America. Farrar,Straus and Giroux.

[12] Garner, Maryah, Anna Harvey and Hunter Johnson. 2019. “Estimating E�ects of A�rmativeAction in Policing: A Replication and Extension.” International Review of Law and Economicsp. 105881.

[13] Gay, Claudine. 2002. “Spirals of trust? The e�ect of descriptive representation on the re-lationship between citizens and their government.” American Journal of Political Sciencepp. 717–732.

[14] Gelman, Andrew, Je�rey Fagan and Alex Kiss. 2007. “An Analysis of the New York City Po-lice Department’s "Stop-and-Frisk’" Policy in the Context of Claims of Racial Bias.” Journalof the American Statistical Association 102(429):813–823.

[15] Go�, Phillip Atiba and Kimberly Barsamian Kahn. 2012. “Racial bias in policing: Why weknow less than we should.” Social Issues and Policy Review 6(1):177–210.

[16] Harvey, Anna and Taylor Mattia. 2019. “Reducing Racial Disparities in Crime Victimization.”Working Paper . https://s18798.pcdn.co/annaharvey/wp-content/uploads/sites/

6417/2019/12/Victimization_Harvey_Mattia.pdf.

[17] Heckman, James J. 1977. “Sample selection bias as a speci�cation error (with an applicationto the estimation of labor supply functions).” NBER Working Paper (No. 172).

[18] Heckman, James J. 1979. “Sample Selection Bias as a Speci�cation Error.” Econometrica47(1):153–161.URL: http://www.jstor.org/stable/1912352

[19] Hoekstra, M. and C. Sloan. 2020. “Does Race Matter for Police Use of Force? Evidencefrom 911 Calls.” National Bureau of Economic Research, Working Paper (26774) . https:

//econrails.uzh.ch/api/agenda/event_attachments/269.

[20] Ho�man, Peter B. and Edward R. Hickey. 2005. “Use of force by female police o�cers.”Journal of Criminal Justice 33(2):145–151.

18

[21] Knox, Dean, Will Lowe and Jonathan Mummolo. 2020. “Administrative Records MaskRacially Biased Policing.” American Political Science Review . https://papers.ssrn.com/sol3/papers.cfm?abstract_id=33363387.

[22] Krogstad, Jens Manuel and Mark Hugo Lopez. 2017. “Use of Spanish declines among Latinosin major U.S. metros.” Pew Research Center . https://www.pewresearch.org/fact-tank/2017/10/31/use-of-spanish-declines-among-latinos-in-major-u-s-metros/.

[23] Legewie, Joscha and Je�rey Fagan. 2016. “Group threat, police o�cer diversity andthe deadly use of police force.” Columbia Public Law Research Paper . http://

raceandpolicing.issuelab.org/resources/25209/25209.pdf.

[24] Lerman, Amy and Vesla Weaver. 2014. Arresting Citizenship: The Democratic Consequencesof American Crime Control. University of Chicago Press.

[25] Lonsway, Kim, Michelle Wood Megan Fickling Alexandria De Leon M. Moore P. HarringtonE. Smeal and Katherine. Spillar. 2002. “Men, women, and police excessive force: A tale oftwo genders.”. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=33363387.

[26] Mcclell, Edward. 2018. “ChicagoâĂŹs Top Cop Was Fired After the Laquan McDonald Shoot-ing. Now He Wants to Be Mayor.”. Library Catalog: www.politico.com.URL: https://politi.co/2Io2Zy1

[27] McCrary, Justin. 2007. “The e�ect of court-ordered hiring quotas on the composition andquality of police.” American Economic Review 97(1):318–353.

[28] Miller, Amalia R and Carmit Segal. 2018. “Do Female O�cers Improve Law EnforcementQuality? E�ects on Crime Reporting and Domestic Violence.” The Review of Economic Studies86(5):2220–2247.

[29] Neil, Roland and Christopher Winship. 2019. “Methodological Challenges and Opportunitiesin Testing for Racial Discrimination in Policing.” Annual Review of Criminology 2(1):73–98.URL: https://doi.org/10.1146/annurev-criminol-011518-024731

[30] Police Accountability Task Force. 2016. Recommendations for Reform: Restoring Trustbetween the Chicago Police and the Communities they Serve. Technical report.

[31] Power, Politics, & Pride: Afro-American Patrolmen’s League. 2018. Library Catalog: interac-tive.wttw.com.URL: https://interactive.wttw.com/dusable-to-obama/afro-american-patrolmens-league

19

[32] Rabe-Hemp, Cara E. 2008. “Female o�cers and the ethic of care: Does o�cer gender impactpolice behaviors?” Journal of Criminal Justice 36(5):426–434.

[33] Raganella, Anthony J. and Michael D. White. 2004. “Race, gender, and motivation for becom-ing a police o�cer: Implications for building a representative police department.” Journal ofCriminal Justice 32(6):501–513.

[34] Reichel, Philip L. 1988. “Southern slave patrols as a transitional police type.” AmericanJournal of Police 7:51–77. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=

33363387.

[35] Ridgeway, Greg and John M. MacDonald. 2009. “Doubly Robust Internal Benchmarking andFalse Discovery Rates for Detecting Racial Bias in Police Stops.” Journal of the AmericanStatistical Association 104:661–668.

[36] Rim, Nayoung, Roman G. Rivera and Bocar A. Ba. 2019. “In-group Bias and the Police:Evidence from Award Nominations.” Working paper .

[37] Rubin, Donald B. 1974. “Estimating causal e�ects of treatments in randomized and non-randomized studies.” Journal of Educational Psychology 66:688–701.

[38] Rubin, Donald B. 1990. “Formal mode of statistical inference for causal e�ects.” Journal ofstatistical planning and inference 25(3):279–292.

[39] Sklansky, David Alan. 2005. “Not your father’s police department: Making sense of the newdemographics of law enforcement.” J. Crim. L. & Criminology 96:1209–1244.

[40] Smith, Brad. 2003. “The Impact of Police O�cer Diversity on Police-Caused Homicides.”Policy Studies 31(2):147–162.

[41] Tate, Katherine. 2001. “The political representation of blacks in Congress: Does race mat-ter?” Legislative Studies Quarterly pp. 623–638.

[42] Turner, K. B., David Giacopassi and Margaret Vandiver. 2006. “Ignoring the past: Coverageof slavery and slave patrols in criminal justice texts.” Journal of Criminal Justice Education17(1):181–195.

[43] Tyler, T.R. and J. Fagan. 2008. “Legitimacy and cooperation: Why do people help the police�ght crime in their communities?” Ohio State Journal of Criminal Law 6(1):231–275.

[44] Weisburst, EK. 2019. “Police use of force as an extension of arrests: Examining disparitiesacross civilian and o�cer race.” American Economic Asssociation: Papers and Proceedings 109.

20

[45] Weisburst, Emily. 2020. “‘Whose help is on the way?’ The importance of individual po-lice o�cers in law enforcement outcomes.” Working Paper . https://drive.google.com/file/d/1z04Hrq1Muv9z0oTuFRWddtp8Pbg7m4aa/view.

[46] West, Jeremy. 2018. “Racial Bias in Police Investigations.”. Working Paper https://people.ucsc.edu/~jwest1/articles/West_RacialBiasPolice.pdf.

[47] White, Michael D., Jonathon A. Cooper Jessica Saunders and Anthony J. Raganella. 2010.“Motivations for becoming a police o�cer: Re-assessing o�cer attitudes and job satisfactionafter six years on the street.” Journal of Criminal Justice 38(4):520–530.

21

Supplemental Information

Table of ContentsA Detailed Description of Data 2

A.1 CPD Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

A.2 Coding Race and Ethnicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

A.3 U.S. Census Merge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

A.4 Preprocessing of Patrol Assignments . . . . . . . . . . . . . . . . . . . . . . . . 3

A.5 Preprocessing of Police Behavior Data . . . . . . . . . . . . . . . . . . . . . . . . 4

B Descriptive Analysis of Assignment Patterns 4

B.1 District Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

B.2 O�cer Demographics and Patrol Assignments . . . . . . . . . . . . . . . . . . . 8

C O�cer Behavior 13

C.1 Unadjusted Average Behavior by Various O�cer Groups . . . . . . . . . . . . . 13

C.2 Estimand . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

C.3 Potential Threats to Validity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

C.4 Shift Duration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

C.5 Variation in Explanatory Variable . . . . . . . . . . . . . . . . . . . . . . . . . . 20

C.6 Gender Heterogeneity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

C.7 Alternative Estimators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

C.8 Robustness Checks: Multiple Stopping O�cers . . . . . . . . . . . . . . . . . . . 27

1

A Detailed Description of Data

A.1 CPD Data

The administrative data from the CPD used in this study span multiple datasets collected in col-laboration with the Invisible Institute, Sam Stecklow, and Emma Herman over the course of threeyears (2016-2019). We obtained these records from the Chicago Police Department or Chicago De-partment of Human Resources via Freedom of Information Act (FOIA) or through court orderedreleases stemming from requests made by Invisible Institute and Jaime Kalven. CPD provided thefollowing data: rosters of all available current and past o�cers up to 2018, unit history data forindividual o�cers from the 1930s to 2016, Tactical Response Reports from 2004 to 2018 (i.e. use offorce reports), and arrest data with arresting o�cers and arrestee demographic information from2001 to 2017. The Chicago Department of Human Resources provided data on o�cers’ languageskills up to 2019 and o�cers’ home address in 2004, 2005 and in early 2019. We supplement ourcore data with data on “Stop, Question and Frisk” (SQF) activity between 2012-2015, which wasshared by the Lucy Parson’s Lab. Finally, the Automated Daily Attendance and Assignment sheetdata for each police district between 2012 and 2015 was obtained via a FOIA request to the CPDand shared by Rachel Ryley.

These data and others have been used to construct rich pro�les of Chicago Police O�cers.While no �le contains a unique identi�er (star numbers change over time, names are common,etc.), we constructed unique o�cer pro�les through a successive merge process described here.Each �le contains some identifying information such as of demographic data (birth year, race,gender) or other characteristics (name, start/badge number, appointed date, resignation date, cur-rent unit). We used these identifying characteristics to �rst de-duplicate o�cers within a �le andto then merge to pre-existing o�cer data with inter-�le unique identi�ers. The merging processitself is an iterative-pairwise matching method, where the o�cers in each dataset are repeatedlymerged on identifying characteristics and any successful 1-to-1 match in a round removes thematched o�cers from the next round of merging.

The resulting data contains records on 33,000 police o�cers appointed between March of 1936to February of 2018. The number of years and o�cers varies across analyses in our paper due tomissing data (for example, assignment data only exists for the years 2012–2015).

A.2 Coding Race and Ethnicity

We determine race/ethnicity of CPD o�cers based on demographic data obtained from the CPDthrough FOIA. The CPD usually classi�es race/ethnicity in at most 7 mutually exclusive groups:white/Caucasian, white Hispanic, Black/African American, Black Hispanic, Asian/Paci�c Islander,

2

Native American/Native Alaskan, and unknown/missing. However, there are inconsistencies inhow races and ethnicities are coded across �les. For example, some �les do not include “BlackHispanic” as a racial category, (very few o�cers are ever classi�ed as Black Hispanic), and some�les contain outdated racial categories which we update to the best of our ability. For consistency,we classify “Hispanic” and “White Hispanic” as “Hispanic”; “Black” and “Black Hispanic” (rarecases) as “Black.” “White” in our analysis refers to non-Hispanic white. If an o�cer has multipleraces associated with them across di�erent datasets, we aggregate by most common non-missingraces.

For Census and American Community Survey data, we construct corresponding race cate-gories as follows: any Hispanic individual is coded Hispanic; white and Black are comprised ofindividuals who are coded as not Hispanic and white (Black) alone.

A.3 U.S. Census Merge

District and beat demographic data was constructed using the 2010 US Census data and the CPD’spre-2012 beat map. The centroid of each census tract was identi�ed, then the demographic in-formation of all the centroids inside a beat were aggregated to determine the beat’s populationand demographic composition. District demographics were determined by aggregating across allbeats within that district. Post-2012 district and beat demographics were constructed based onthe pre-2012 beat data discussed previously and using a crosswalk that maps pre-2012 beats tocurrent (2018) beats and their respective districts.

A.4 Preprocessing of Patrol Assignments

We restrict analysis to patrol assignments in which Black, Hispanic, or white o�cers serve.Asian/Paci�c Islander and Native American/Alaskan Native o�cers are not examined due tosmall sample sizes. Within this subset, we further drop non-standard assignments (notably in-cluding “protest detail,” “station supervisor,” and “station security” assignments, as well as spe-cial assignments for training, compensatory time, and excused sick leave). Patrol assignments inwhich o�cers are indicated as non-present are also dropped. These steps are intended to ensurethat o�cers nominally patrolling a beat are in fact actively circulating in the assigned geographicarea, improving the plausibility of the common-circumstances assumption. For the same reason,we drop double shifts (patrol assignment slots in which the assigned o�cer served for more thanone shift on the same day) to address the possibility that o�cers behave di�erently due to fatiguein these circumstances. We also eliminate o�cers assigned to non-standard watches (i.e., otherthan �rst through third watches). Finally, we drop o�cers at ranks other than “police o�cer.”This step eliminates police sergeants, who serve in 8% of beat assignments but make very few

3

stops and arrests, as well as legal o�cers, helicopter pilots, explosives technicians, and caninehandlers.

A.5 Preprocessing of Police Behavior Data

Events are merged to the remaining patrol assignments based on o�cer ID and date. This stepdiscards a large number of events, including those involving o�cers of higher ranks and incidentsoccurring on rest days. For stops, arrests, and uses of force, we drop all events that occur out-side of the reported patrol start/stop times, eliminating o�-duty activity. Non-standard shifts aredropped (retaining only �rst–third watches), as are absences, shifts with assigned special duties(e.g. protest detail, station security, training), and double- or triple-duty days in which a singleo�cer serves multiple shifts.

Stops for “dispersal” and “gang and narcotics-related loitering” are coded as loitering stops;those that are “gang / narcotics related” are coded as drug stops; “investigatory stops” and stopsof “suspicious persons” are coded as suspicious behavior; and stops under the “Repeat O�enderGeographic Urban Enforcement Strategy (ROGUES)” program are combined with the “other”category. For stops, if a single o�cer is reported as both primary and secondary stopping o�cer,only one event is retained.

Arrests for municipal code violations and outstanding warrants are categorized as “other.”

B Descriptive Analysis of Assignment Patterns

B.1 District Characteristics

The CPD currently subdivides Chicago into 22 policing districts which correspond to CPD units,in which the majority of police o�cers work. A typical district covers roughly ten square miles.There were 25 districts (numbered 1 - 25) until 2012, at which time 3 smaller districts (ranking18th, 21st, and 25th in land area1) were eliminated and merged with other districts. Districts23 and 21 and District 13 were eliminated and absorbed into neighboring districts in March andDecember of 2012, respectively. While District 23 was mostly absorbed by District 19 and most ofDistrict 13 was absorbed by District 12, signi�cant parts of District 21 were absorbed by Districts1, 2, and 9.

Figure B.1 illustrates the types of districts to which o�cers of each demographic group areassigned. This analysis takes each unique combination of racial/ethnic and gender, identi�es allo�cers in that group, and then compute their assigned districts’ average characteristics. Four

1See Chicago Police Department 2008 Annual Report, page 37, https://home.chicagopolice.org/wp-content/uploads/2014/12/2008-Annual-Report.pdf

4

Figure B.1: Average characteristics of assigned geographic districts for various o�cer groups, byyear, based on 1,089,707 monthly assignments to geographic units from 2006–2016. Con�denceintervals cluster on district-month.

Black officers Hispanic officers White officers

Vio

lent crim

es

per 1

,00

0 re

sidents

per m

onth

Pro

perty

crimes

per 1

,00

0 re

sidents

per m

onth

Police

office

rsper 1

,00

0 re

sidents

Pro

portio

n o

fco

-racia

l resid

ents

2006 2008 2010 2012 2014 2016 2006 2008 2010 2012 2014 2016 2006 2008 2010 2012 2014 2016

0.0

0.5

1.0

1.5

2.0

0

2

4

6

8

0

2

4

6

8

0.00

0.25

0.50

0.75

Year

Avera

ge c

hara

cteri

stic

s of

ass

igned

dis

tric

ts

Officer race Black Hispanic White Gender Female Male

dimensions are examined: violent crime rates, property crime rates, police o�cer density, andproportion of co-racial residents. (These results are analogous to those presented in Figure 2, butbreak out each year separately rather than pooling.) Results are highly comparable, indicatingthat reported patterns are not an artifact of pooling across years.

We now turn to two district-level analyses. Figure B.2 plots the relationship between a po-lice district’s resident demographic pro�le (e.g. the proportion of residents that are Black) ando�cer demographic pro�le (the proportion of o�cers assigned to that district that are Black).White-dominated districts have virtually no minority o�cers assigned, and districts with size-able minority populations tend to have more o�cers of the corresponding race. However, o�cersare disproportionately white compared to district residents: a number of districts dominated byBlack residents nonetheless have sizeable contingents of white o�cers. For example, Wentworth

5

Figure B.2: Racial and ethnic composition of o�cers’ assigned districts. In each panel,each point represents a police district. The horizontal axis indicates the proportion of civilians ofa given racial/ethnic group residing in 2010 Census data, and the vertical axis depicts the shareof o�cers assigned to that district in January 2010 from the same racial/ethnic group.

Black Hispanic White

0.0 0.5 1.00.0 0.5 1.00.0 0.5 1.0

0.00

0.25

0.50

0.75

1.00

Proportion of district civilians

Pro

port

ion o

f off

icers

ass

igned t

o d

istr

ict

(CPD District 2, depicted in Figure 3) is 98% Black, but 24% of o�cers assigned there are white.The disparity is even starker in Austin (CPD District 15), where a 96% Black resident population ispoliced by a unit that is 55% white. (See SI Section A.3 for details on the computation of residentdemographics.)

Figure B.3 displays signi�cant over-time changes in the racial composition of o�cers assignedto a district. In this �gure, the vertical slice at 2010 corresponds to the results plotted in Figure B.2.The proportion of Black o�cers assigned to some districts (e.g. districts 3, 5, 6, 7) while holdingsteady in others. Temporal discontinuities are due to changes in district boundaries or eliminationof police districts.

6

Figure B.3: Racial composition of police districts. Each panel depicts a geographic policedistrict. Points represent the racial composition of district residents. Lines represent monthlyproportions of o�cers assigned to a district that belong to each racial group. Districts 21, 23, and13 were eliminated during the observation period.

District 21: Prairie District 22: Morgan Park District 23: Town Hall District 24: Rogers Park District 25: Grand Central

District 16: Jefferson Park District 17: Albany Park District 18: Near North District 19: Belmont District 20: Lincoln

District 11: Harrison District 12: Monroe District 13: Wood District 14: Shakespeare District 15: Austin

District 6: Gresham District 7: Englewood District 8: Chicago Lawn District 9: Deering District 10: Ogden

District 1: Central District 2: Wentworth District 3: Grand Crossing District 4: South Chicago District 5: Calumet

2005 2010 2015 2005 2010 2015 2005 2010 2015 2005 2010 2015 2005 2010 2015

0.00

0.25

0.50

0.75

1.00

0.00

0.25

0.50

0.75

1.00

0.00

0.25

0.50

0.75

1.00

0.00

0.25

0.50

0.75

1.00

0.00

0.25

0.50

0.75

1.00

Prop

ortio

n

Race Black Hispanic White

7

B.2 O�cer Demographics and Patrol Assignments

Among o�cers assigned to a particular police district, considerable variation exists in the exactpatrol assignments that o�cers receive. We examine each unit individually, tabulating o�cerrace and shift time assignments (�rst, second and third watch, respectively corresponding to thenominal duty periods of midnight to 8 a.m., 8 a.m. to 4 p.m., and 4 p.m. to midnight). Figures B.4–B.5 depict the frequency of each shift period, respectively showing that the pattern of assignmentsdi�ers dramatically by o�cer race and gender. For example, white o�cers in Wentworth (District2) almost exclusively serve from 4 p.m. to midnight, whereas Black o�cers are more likely to beassigned to mid-day shifts.

Figures B.6–B.7 examine the pattern of patrol beat assignments by race/ethnicity and gender,respectively. They show that, for example, relative to white o�cers, Black o�cers are far morefrequently deployed to assigned beat 202—which roughly corresponds to a patrol area in thedistrict’s southwest corner (depicted in Figure 3) that has extremely high police activity and a highconcentration of Black residents. These results undermine analyses in a wide array of previousstudies that aggregate at high levels of geography (for example, controlling for district or unitassignment) and which assume that o�cers face homogeneous conditions within these crudegroupings.

8

Figu

reB.

4:Assigne

dsh

ifttimeby

o�cerrace.

Each

pane

ldep

icts

o�ce

rsw

ithin

age

ogra

phic

polic

edi

stric

t.W

ithin

adi

stric

t,ea

chro

wsh

owst

hepr

opor

tion

ofsh

iftas

signm

ents

forB

lack

,Hisp

anic

,orw

hite

o�ce

rsto

�rst

wat

ch(m

idni

ghtt

o8

a.m.),

seco

ndw

atch

(8a.m

.to

4p.

m.),

and

third

wat

ch(4

p.m

.to

mid

nigh

t).D

arke

rcel

lsin

dica

teah

ighe

rpro

porti

onof

assig

nmen

tsto

ashi

fttim

e,an

den

tries

ina

row

sum

toun

ity.T

he�g

ure

dem

onst

rate

stha

twith

inan

ypa

rticu

lard

istric

t,Bl

ack

and

Hisp

anic

o�ce

rsar

eca

lled

tose

rve

atve

rydi

�ere

nttim

esof

day

(p<0.001

fora

llw

ithin

-dist

ricts

tatis

tical

test

sofi

ndep

ende

nce)

.

Dis

trict

21:

Pra

irie

Dis

trict

22:

Mor

gan

Park

Dis

trict

23:

Tow

n H

all

Dis

trict

24:

Rog

ers

Park

Dis

trict

25:

Gra

nd C

entra

l

Dis

trict

16:

Jef

fers

on P

ark

Dis

trict

17:

Alb

any

Park

Dis

trict

18:

Nea

r Nor

thD

istri

ct 1

9: B

elm

ont

Dis

trict

20:

Lin

coln

Dis

trict

11:

Har

rison

Dis

trict

12:

Mon

roe

Dis

trict

13:

Woo

dD

istri

ct 1

4: S

hake

spea

reD

istri

ct 1

5: A

ustin

Dis

trict

6: G

resh

amD

istri

ct 7

: Eng

lew

ood

Dis

trict

8: C

hica

go L

awn

Dis

trict

9: D

eerin

gD

istri

ct 1

0: O

gden

Dis

trict

1: C

entra

lD

istri

ct 2

: Wen

twor

thD

istri

ct 3

: Gra

nd C

ross

ing

Dis

trict

4: S

outh

Chi

cago

Dis

trict

5: C

alum

et

1st w

atch

2nd

wat

ch3r

d w

atch

1st w

atch

2nd

wat

ch3r

d w

atch

1st w

atch

2nd

wat

ch3r

d w

atch

1st w

atch

2nd

wat

ch3r

d w

atch

1st w

atch

2nd

wat

ch3r

d w

atch

1st w

atch

2nd

wat

ch3r

d w

atch

1st w

atch

2nd

wat

ch3r

d w

atch

1st w

atch

2nd

wat

ch3r

d w

atch

1st w

atch

2nd

wat

ch3r

d w

atch

1st w

atch

2nd

wat

ch3r

d w

atch

1st w

atch

2nd

wat

ch3r

d w

atch

1st w

atch

2nd

wat

ch3r

d w

atch

1st w

atch

2nd

wat

ch3r

d w

atch

1st w

atch

2nd

wat

ch3r

d w

atch

1st w

atch

2nd

wat

ch3r

d w

atch

1st w

atch

2nd

wat

ch3r

d w

atch

1st w

atch

2nd

wat

ch3r

d w

atch

1st w

atch

2nd

wat

ch3r

d w

atch

1st w

atch

2nd

wat

ch3r

d w

atch

1st w

atch

2nd

wat

ch3r

d w

atch

1st w

atch

2nd

wat

ch3r

d w

atch

1st w

atch

2nd

wat

ch3r

d w

atch

1st w

atch

2nd

wat

ch3r

d w

atch

1st w

atch

2nd

wat

ch3r

d w

atch

1st w

atch

2nd

wat

ch3r

d w

atch

Blac

k

His

pani

c

Whi

te

Blac

k

His

pani

c

Whi

te

Blac

k

His

pani

c

Whi

te

Blac

k

His

pani

c

Whi

te

Blac

k

His

pani

c

Whi

te

Blac

k

His

pani

c

Whi

te

Blac

k

His

pani

c

Whi

te

Blac

k

His

pani

c

Whi

te

Blac

k

His

pani

c

Whi

te

Blac

k

His

pani

c

Whi

te

Blac

k

His

pani

c

Whi

te

Blac

k

His

pani

c

Whi

te

Blac

k

His

pani

c

Whi

te

Blac

k

His

pani

c

Whi

te

Blac

k

His

pani

c

Whi

te

Blac

k

His

pani

c

Whi

te

Blac

k

His

pani

c

Whi

te

Blac

k

His

pani

c

Whi

te

Blac

k

His

pani

c

Whi

te

Blac

k

His

pani

c

Whi

te

Blac

k

His

pani

c

Whi

te

Blac

k

His

pani

c

Whi

te

Blac

k

His

pani

c

Whi

te

Blac

k

His

pani

c

Whi

te

Blac

k

His

pani

c

Whi

te

Assi

gned

shi

ft

0.00

0.25

0.50

0.75

Prop

ortio

n of

offi

cer-s

hifts

ass

igne

d to

shi

ft

9

Figu

reB.

5:Assigne

dsh

ifttim

eby

o�cerge

nder.E

ach

pane

ldep

icts

o�ce

rsw

ithin

age

ogra

phic

polic

edi

stric

t.W

ithin

adi

stric

t,ea

chro

wsh

owst

hepr

opor

tion

ofsh

iftas

signm

ents

forf

emal

eorm

aleo

�ce

rsto

�rst

wat

ch(m

idni

ghtt

o8a

.m.),

seco

ndw

atch

(8a.m

.to

4p.

m.),

and

third

wat

ch(4

p.m

.to

mid

nigh

t).D

arke

rcel

lsin

dica

tea

high

erpr

opor

tion

ofas

signm

ents

toa

shift

time,

and

entri

esin

aro

wsu

mto

unity

.The

�gur

ede

mon

stra

test

hatw

ithin

any

parti

cula

rdist

rict,

fem

ale

and

mal

eo�

cers

are

calle

dto

serv

eat

very

di�e

rent

times

ofda

y(p<0.001

fora

llw

ithin

-dist

ricts

tatis

tical

test

sofi

ndep

ende

nce)

.

Dis

tric

t 2

1:

Pra

irie

Dis

tric

t 2

2:

Morg

an P

ark

Dis

tric

t 2

3:

Tow

n H

all

Dis

tric

t 2

4:

Rogers

Park

Dis

tric

t 2

5:

Gra

nd C

entr

al

Dis

tric

t 1

6:

Jeff

ers

on P

ark

Dis

tric

t 1

7:

Alb

any P

ark

Dis

tric

t 1

8:

Near

Nort

hD

istr

ict

19

: B

elm

ont

Dis

tric

t 2

0:

Linco

ln

Dis

tric

t 1

1:

Harr

ison

Dis

tric

t 1

2:

Monro

eD

istr

ict

13

: W

ood

Dis

tric

t 1

4:

Shake

speare

Dis

tric

t 1

5:

Aust

in

Dis

tric

t 6

: G

resh

am

Dis

tric

t 7

: Engle

wood

Dis

tric

t 8

: C

hic

ago L

aw

nD

istr

ict

9:

Deeri

ng

Dis

tric

t 1

0:

Ogden

Dis

tric

t 1

: C

entr

al

Dis

tric

t 2

: W

entw

ort

hD

istr

ict

3:

Gra

nd C

ross

ing

Dis

tric

t 4

: South

Chic

ago

Dis

tric

t 5

: C

alu

met

1st

watc

h2

nd w

atc

h3

rd w

atc

h1

st w

atc

h2

nd w

atc

h3

rd w

atc

h1

st w

atc

h2

nd w

atc

h3

rd w

atc

h1

st w

atc

h2

nd w

atc

h3

rd w

atc

h1

st w

atc

h2

nd w

atc

h3

rd w

atc

h

1st

watc

h2

nd w

atc

h3

rd w

atc

h1

st w

atc

h2

nd w

atc

h3

rd w

atc

h1

st w

atc

h2

nd w

atc

h3

rd w

atc

h1

st w

atc

h2

nd w

atc

h3

rd w

atc

h1

st w

atc

h2

nd w

atc

h3

rd w

atc

h

1st

watc

h2

nd w

atc

h3

rd w

atc

h1

st w

atc

h2

nd w

atc

h3

rd w

atc

h1

st w

atc

h2

nd w

atc

h3

rd w

atc

h1

st w

atc

h2

nd w

atc

h3

rd w

atc

h1

st w

atc

h2

nd w

atc

h3

rd w

atc

h

1st

watc

h2

nd w

atc

h3

rd w

atc

h1

st w

atc

h2

nd w

atc

h3

rd w

atc

h1

st w

atc

h2

nd w

atc

h3

rd w

atc

h1

st w

atc

h2

nd w

atc

h3

rd w

atc

h1

st w

atc

h2

nd w

atc

h3

rd w

atc

h

1st

watc

h2

nd w

atc

h3

rd w

atc

h1

st w

atc

h2

nd w

atc

h3

rd w

atc

h1

st w

atc

h2

nd w

atc

h3

rd w

atc

h1

st w

atc

h2

nd w

atc

h3

rd w

atc

h1

st w

atc

h2

nd w

atc

h3

rd w

atc

h

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Ass

igned

shift

0.2

0.4

0.6

Pro

port

ion o

f off

icer-

shifts

ass

igned

to s

hift

10

Figu

reB.

6:Assigne

dbe

atby

o�cerrace.

Each

pane

ldep

icts

o�ce

rsw

ithin

age

ogra

phic

polic

edi

stric

t.W

ithin

adi

stric

t,ea

chro

wsh

owst

hepr

opor

tion

ofsh

iftas

signm

ents

forB

lack

,Hisp

anic

,orw

hite

o�ce

rsto

beat

s,or

geog

raph

icpa

trola

reas

.Dar

kerc

ells

indi

cate

ahi

gher

prop

ortio

nof

assig

nmen

tsto

abe

at,a

nden

tries

ina

row

sum

toun

ity.

The

�gur

ede

mon

stra

tes

that

with

inan

ypa

rticu

lard

istric

t,Bl

ack

and

Hisp

anic

o�ce

rsar

eca

lled

tose

rve

atve

rydi

�ere

ntlo

catio

ns(p<0.001

fora

llw

ithin

-dist

ricts

tatis

tical

test

sofi

ndep

ende

nce)

.

Dis

trict

25:

Gra

nd C

entra

l

Dis

trict

22:

Mor

gan

Park

Dis

trict

23:

Tow

n H

all

Dis

trict

24:

Rog

ers

Park

Dis

trict

19:

Bel

mon

tD

istri

ct 2

0: L

inco

lnD

istri

ct 2

1: P

rairi

e

Dis

trict

16:

Jef

fers

on P

ark

Dis

trict

17:

Alb

any

Park

Dis

trict

18:

Nea

r Nor

th

Dis

trict

13:

Woo

dD

istri

ct 1

4: S

hake

spea

reD

istri

ct 1

5: A

ustin

Dis

trict

10:

Ogd

enD

istri

ct 1

1: H

arris

onD

istri

ct 1

2: M

onro

e

Dis

trict

7: E

ngle

woo

dD

istri

ct 8

: Chi

cago

Law

nD

istri

ct 9

: Dee

ring

Dis

trict

4: S

outh

Chi

cago

Dis

trict

5: C

alum

etD

istri

ct 6

: Gre

sham

Dis

trict

1: C

entra

lD

istri

ct 2

: Wen

twor

thD

istri

ct 3

: Gra

nd C

ross

ing

250025042549255925742579255625532547259225992594257825762543258525062545258425412570258325902572255125402501258225752580258625542548250325072593257725302552250525502546251025442534252025132542253125232525252225152521251225352532253325242514251125732502

2274

2278

2253

2276

2277

2242

2203

2251

2275

2280

2250

2285

2206

2252

2241

2240

2254

2294

2243

2290

2293

2289

2270

2200

2299

2207

2284

2282

2222

2279

2283

2230

2281

2201

2212

2232

2220

2233

2210

2223

2205

2234

2213

2221

2272

2211

2202

2382

2384

2388

2343

2395

2390

2381

2342

2333

2306

2330

2350

2300

2307

2372

2386

2394

2305

2352

2340

2351

2393

2324

2341

2301

2320

2323

2332

2322

2399

2310

2312

2331

2311

2313

2302

2408247024032458245924062448245724042456244924712446244024902400249524942454245524922453245124522493244224822407247624472499248324202410244324412450243024122481248624872485244524842424242224442401241124332405241324312423243224022472

19911997195819771989198719921996190319571943195219931982199519421947197419491904194619701948195419081955195319561941194019901950198819451981199419511984198319441980198519301900190719061999198619331905192619211911191519721910192519201935191419311932192319121934191319221971192419011902

2008

2046

2070

2080

2081

2087

2054

2095

2043

2000

2092

2053

2090

2044

2099

2040

2084

2052

2050

2042

2041

2051

2006

2086

2094

2072

2001

2030

2085

2093

2010

2013

2031

2005

2023

2020

2024

2032

2022

2011

2012

2033

2071

2002

2100

2112

2122

2124

2141

2142

2170

2172

2174

2181

2182

2183

2199

2132

2195

2101

2133

2102

2110

2130

2131

2123

2105

2113

2120

2111

1683

1604

1686

1688

1606

1654

1699

1674

1645

1696

1607

1608

1682

1656

1653

1651

1690

1643

1601

1640

1695

1681

1642

1670

1652

1650

1600

1641

1630

1610

1631

1605

1620

1613

1671

1633

1611

1612

1623

1632

1624

1614

1622

1634

1621

1602

1703

1784

1754

1791

1785

1796

1743

1789

1782

1744

1741

1708

1742

1745

1774

1752

1700

1750

1740

1797

1755

1790

1770

1788

1787

1795

1753

1705

1707

1730

1710

1732

1701

1786

1723

1720

1751

1712

1799

1731

1711

1722

1733

1713

1772

1724

1702

1851

1876

1843

1854

1850

1882

1877

1886

1846

1856

1853

1840

1852

1889

1845

1812

1844

1890

1871

1842

1807

1883

1805

1870

1822

1800

1824

1841

1814

1873

1821

1806

1810

1872

1888

1834

1833

1830

1899

1831

1801

1820

1832

1813

1811

1823

1802

1304

1307

1308

1342

1390

1352

1341

1395

1351

1381

1353

1322

1365

1324

1370

1301

1306

1350

1313

1300

1372

1305

1311

1333

1310

1330

1332

1320

1331

1323

1312

1302

1408

1483

1488

1487

1474

1451

1465

1489

1456

1495

1444

1486

1455

1407

1494

1446

1472

1450

1443

1400

1454

1445

1405

1442

1403

1453

1421

1452

1490

1424

1430

1499

1482

1440

1484

1412

1432

1411

1420

1441

1434

1471

1410

1401

1433

1413

1414

1423

1422

1431

1402

1500

1507

1570

1506

1580

1540

1508

1583

1544

1555

1584

1505

1590

1581

1504

1582

1586

1543

1530

1595

1542

1593

1556

1553

1545

1554

1551

1520

1571

1585

1541

1550

1511

1501

1510

1599

1513

1552

1524

1533

1523

1532

1522

1512

1531

1572

1502

1000100710701076107910811091105510031056108610741053109210061045105410441077105110651075108410421096100810721043108810041090104110461031105010521032103410821001109510201023108910831022109310051010103310401099103010121013101110241021101410711002

1100110711081147115711741146117911581177119211731145117511531178115511931176118911441154119011651185110611561194115111701150118411821103118611421143115211401181119911721187119511831159114111051132113411151110113011011120112511141188112111131131112211351124112311121111113311711102

12071273127712881248129512471274127912561265128012811246124412061291129412541285127012501208125112921201124012551290129912411283125212091253121412431200124512931215122112421234123012351210123312111271122012051212122312251272121312321224122212311202

703707708775776

7775781785788795786757758770789754784755792706750794783746753787756741700772751790752774730793743745705740742799720701735714744724722782702723732734710733731725726715771712713711

806880886889879876878874877875870883850890803853857895893856894858840852844848882851896854807884847846842800899845805822843831885892849830835810859855832821823841801820873825833871824834814815811812813872802

983974986991903947981984994993959958965908995970950987945941942953978956951944946943985971990901940989955900907954952906999910930920932921923905988924934913935922912972931933915973914925911902

400404408483484485487488489496457477475481478482452454405442458450444440443465441407456486446403494451490406470495479447414401499445412472431453410421422411423471420430455432424413433402434

500503504507546555581582583584585586598549556565576577597578591547575579564572554596544550595501545553590506505540531508570543542552551512532523599502530541533511513510520522524571

600607608657661670675676677678682691695689686690656683662668688699693663681694665640655605653647646644684633687643622672642650645602654606641614632630634631624651652685623601612620621610611671613

173

141

104

134

142

140

190

106

199

193

170

110

123

114

108

124

130

131

180

100

192

122

120

112

133

171

101

107

132

113

121

105

111

102

172

200203207208251253255256270274275276277278279285286289297284254282204252247250296281245205272246283240214294206234222223244280224290202233212243235213211287232299242241230295201215288220210225271231221

300307308356365370373374375376377378379381387388354395391382346303352304350355351347306386353389384390368344385369380333372340305345371343314341330342301311399321332320331334383323324313310322312302

Blac

kH

ispa

nic

Whi

te

Blac

kH

ispa

nic

Whi

te

Blac

kH

ispa

nic

Whi

te

Blac

kH

ispa

nic

Whi

te

Blac

kH

ispa

nic

Whi

te

Blac

kH

ispa

nic

Whi

te

Blac

kH

ispa

nic

Whi

te

Blac

kH

ispa

nic

Whi

te

Blac

kH

ispa

nic

Whi

te

Blac

kH

ispa

nic

Whi

te

Blac

kH

ispa

nic

Whi

te

Blac

kH

ispa

nic

Whi

te

Blac

kH

ispa

nic

Whi

te

Blac

kH

ispa

nic

Whi

te

Blac

kH

ispa

nic

Whi

te

Blac

kH

ispa

nic

Whi

te

Blac

kH

ispa

nic

Whi

te

Blac

kH

ispa

nic

Whi

te

Blac

kH

ispa

nic

Whi

te

Blac

kH

ispa

nic

Whi

te

Blac

kH

ispa

nic

Whi

te

Blac

kH

ispa

nic

Whi

te

Blac

kH

ispa

nic

Whi

te

Blac

kH

ispa

nic

Whi

te

Blac

kH

ispa

nic

Whi

te

Assi

gned

bea

t

0.0

0.1

0.2

0.3

0.4

0.5

Prop

ortio

n of

offi

cer-s

hifts

ass

igne

d to

bea

t

11

Figu

reB.

7:Assigne

dbe

atby

o�cerge

nder.E

ach

pane

ldep

icts

o�ce

rsw

ithin

age

ogra

phic

polic

edi

stric

t.W

ithin

adi

stric

t,ea

chro

wsh

owst

hepr

opor

tion

ofsh

iftas

signm

ents

forf

emal

eor

mal

eo�

cers

tobe

ats,

orge

ogra

phic

patro

lare

as.D

arke

rcel

lsin

dica

tea

high

erpr

opor

tion

ofas

signm

ents

toa

beat

,and

entri

esin

aro

wsu

mto

unity

.The

�gur

ede

mon

stra

test

hatw

ithin

any

parti

cula

rdi

stric

t,fe

mal

ean

dm

ale

o�ce

rsar

eca

lled

tose

rve

atve

rydi

�ere

ntlo

catio

ns(p

<0.001

for

allw

ithin

-dist

ricts

tatis

tical

test

sof

inde

pend

ence

).

Dis

tric

t 2

5:

Gra

nd C

entr

al

Dis

tric

t 2

2:

Morg

an P

ark

Dis

tric

t 2

3:

Tow

n H

all

Dis

tric

t 2

4:

Rogers

Park

Dis

tric

t 1

9:

Belm

ont

Dis

tric

t 2

0:

Linco

lnD

istr

ict

21

: Pra

irie

Dis

tric

t 1

6:

Jeff

ers

on P

ark

Dis

tric

t 1

7:

Alb

any P

ark

Dis

tric

t 1

8:

Near

Nort

h

Dis

tric

t 1

3:

Wood

Dis

tric

t 1

4:

Shake

speare

Dis

tric

t 1

5:

Aust

in

Dis

tric

t 1

0:

Ogden

Dis

tric

t 1

1:

Harr

ison

Dis

tric

t 1

2:

Monro

e

Dis

tric

t 7

: Engle

wood

Dis

tric

t 8

: C

hic

ago L

aw

nD

istr

ict

9:

Deeri

ng

Dis

tric

t 4

: South

Chic

ago

Dis

tric

t 5

: C

alu

met

Dis

tric

t 6

: G

resh

am

Dis

tric

t 1

: C

entr

al

Dis

tric

t 2

: W

entw

ort

hD

istr

ict

3:

Gra

nd C

ross

ing

2579

2549

2574

2504

2547

2553

2592

2599

2594

2543

2559

2585

2590

2570

2545

2583

2541

2540

2584

2501

2506

2572

2556

2582

2551

2507

2500

2578

2580

2575

2503

2586

2576

2593

2548

2530

2577

2554

2510

2552

2505

2550

2520

2544

2546

2542

2515

2513

2514

2522

2534

2533

2525

2532

2511

2531

2512

2524

2521

2523

2535

2573

2502

2203

2285

2253

2277

2279

2280

2250

2251

2276

2275

2274

2206

2241

2289

2242

2252

2240

2278

2254

2207

2270

2243

2294

2290

2293

2200

2299

2283

2284

2282

2281

2230

2201

2210

2232

2223

2220

2233

2205

2213

2222

2212

2221

2202

2234

2211

2272

2300

2380

2388

2394

2343

2390

2381

2306

2384

2382

2342

2350

2330

2386

2340

2352

2351

2393

2301

2305

2395

2324

2332

2322

2341

2307

2333

2399

2310

2320

2312

2323

2372

2331

2311

2313

2302

2408

2458

2403

2459

2404

2406

2448

2456

2470

2457

2490

2449

2446

2440

2495

2471

2400

2493

2452

2487

2482

2442

2451

2454

2453

2455

2492

2407

2450

2430

2476

2494

2443

2447

2499

2441

2405

2410

2420

2401

2445

2486

2481

2485

2484

2433

2483

2444

2422

2412

2432

2411

2431

2402

2424

2413

2423

2472

19741977199119581997190319921943195719871952190819891904198219471954194919941942195319951955194619501941199319561990194519401948197019961951194419831988198019811907190619841900199919301986190519261985193419201914193519321910190119311922192419331972191219251911192119151923191319711902

2008

2080

2054

2081

2087

2070

2095

2043

2090

2046

2092

2099

2053

2044

2086

2040

2084

2052

2000

2050

2042

2006

2041

2051

2094

2001

2093

2072

2005

2030

2010

2085

2013

2012

2023

2031

2020

2011

2032

2024

2022

2033

2071

2002

2100

2170

2195

2199

2101

2105

2130

2181

2183

2141

2182

2174

2113

2131

2123

2112

2110

2120

2122

2132

2102

2133

2124

2111

2172

1606

1608

1604

1686

1672

1688

1654

1699

1674

1645

1696

1682

1651

1607

1656

1640

1653

1601

1681

1690

1670

1643

1695

1642

1652

1650

1641

1600

1630

1612

1610

1620

1605

1632

1623

1613

1633

1631

1671

1621

1622

1634

1611

1624

1614

1602

1703

1754

1784

1791

1796

1785

1787

1743

1789

1744

1707

1708

1797

1742

1752

1782

1750

1774

1740

1770

1700

1755

1790

1705

1741

1753

1745

1795

1788

1730

1710

1786

1701

1720

1799

1712

1751

1722

1733

1732

1723

1713

1731

1711

1724

1772

1702

1807

1856

1889

1876

1851

1840

1843

1882

1886

1877

1854

1852

1853

1883

1890

1844

1870

1805

1850

1842

1846

1830

1800

1810

1841

1806

1845

1801

1820

1899

1888

1871

1814

1833

1873

1832

1834

1813

1822

1811

1831

1821

1812

1824

1823

1802

1872

1342

1352

1308

1304

1381

1341

1390

1353

1365

1351

1301

1370

1350

1300

1306

1395

1307

1310

1330

1320

1305

1331

1333

1332

1313

1311

1312

1372

1322

1323

1302

1324

1465

1486

1450

1456

1451

1490

1489

1487

1499

1452

1445

1494

1446

1408

1472

1455

1443

1444

1403

1453

1440

1474

1405

1430

1400

1454

1482

1495

1407

1442

1488

1484

1483

1401

1441

1420

1413

1410

1422

1423

1412

1421

1434

1424

1432

1411

1402

1414

1433

1431

1471

1508

1546

1570

1506

1580

1540

1500

1590

1581

1504

1555

1507

1544

1593

1582

1530

1556

1583

1584

1543

1595

1542

1553

1554

1545

1586

1541

1585

1551

1599

1550

1520

1510

1505

1524

1512

1501

1552

1523

1511

1522

1513

1532

1533

1572

1531

1571

1502

1075

1081

1082

1074

1091

1084

1076

1077

1065

1079

1008

1000

1096

1044

1090

1006

1070

1092

1042

1045

1088

1043

1046

1001

1086

1089

1083

1007

1050

1095

1054

1040

1051

1055

1041

1010

1004

1052

1056

1099

1053

1030

1072

1093

1005

1020

1031

1012

1023

1034

1022

1014

1032

1021

1024

1011

1013

1033

1071

1002

115811571146117911741103117311771145119311751153115511901189116511501108117011061144119411991156117811921100115411761185118611511184114011521107114311721142119511871188115911821130111011201183110111151132110511411181111311241122113111341111113511211125111411231133111211711102

1209

1258

1248

1208

1273

1247

1294

1265

1206

1256

1244

1201

1288

1246

1291

1200

1277

1280

1285

1283

1270

1279

1250

1253

1292

1290

1274

1299

1252

1254

1240

1295

1207

1281

1251

1255

1243

1245

1241

1293

1242

1230

1225

1220

1210

1214

1205

1221

1234

1202

1215

1231

1235

1223

1212

1211

1232

1233

1224

1213

1272

1222

1271

7775759703757788789758785795784792706775708786770750776790730700781794799746755756754707774741753752740783701793705745751720782744711743725742710787723702733722731735713715734732724712714726772771

884878879874875876880877806870857850889890807895853893840894854852882896851886800883899842803858843856885847892810848830846801845844855820841805849859811831873821815832871824822812833825814834835813823872802

987974995948903991994947984993978983958942965959970908941945953901950940951930990988943944956955986900907952910946999981954971989906985931920905912922935934923925911933921913972914932915924973902

408483489448449457481487496485403404450458407465452488454442486400444475440482490484443456446494470406451495499477453410401447445479478420405430472455441423431412411424422421413433414402432434471

503

564

598

577

508

565

591

504

549

597

556

547

546

572

586

570

550

506

507

555

585

596

554

540

501

545

590

553

500

581

582

583

544

584

595

599

530

578

576

579

575

552

543

551

542

520

510

541

502

505

531

512

511

522

533

532

513

523

524

571

691657661668662695656663665690688608640650655693646699605607689686683644647670600676678684685675687645654630601642677606694672643681651641653652682632610620602622612631633623621634624614611613671

104

173

141

192

134

140

142

190

106

107

199

193

170

108

110

100

180

123

112

111

121

130

122

114

113

124

132

131

105

101

120

133

102

171

172

203208285204250256296253297255282252251270247280289287290294245207246281240284274299254283276277206279244230286201200275278288295210211243220272225205223242213214215235212241231232224222233234221202271

308373387381365370374391388304303389356375355395347352379346354350351376369380386353368390307340377344399300382306385378384372330345383320310301343305342341332331314313321324323334311333322312302371

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Fem

ale

Male

Ass

igned

beat

0.0

0.1

0.2

0.3

0.4

0.5

Pro

port

ion o

f off

icer-

shifts

ass

igned

to b

eat

12

C O�cer Behavior

C.1 Unadjusted Average Behavior by Various O�cer Groups

Table C.1: Average events per shift, by o�cer racial/ethnic group. Mean number of stops,arrests, and uses of force without adjustment for time or location. Typical behavior is reportedfor Black, Hispanic, and white o�cers individually, as well as the average pooling three o�-cer races. Records associated with Native American/Alaskan and Asian/Paci�c Islander o�cersare excluded due to small sample sizes. O�cer behavior toward Native American/Alaskan andAsian/Paci�c Islander civilians is not included for the purposes of computing total and reason-speci�c events. Values are scaled for ease of interpretation.

Mean Mean Mean MeanBehavior (Pooled) (Black o�.) (Hisp. o�.) (White o�.)

Stops per 10 shifts:Civilian race: Black 3.03 2.61 3.09 3.25Civilian race: Hispanic 0.92 0.21 1.43 1.10Civilian race: White 0.57 0.21 0.61 0.75Reason: Drug 0.48 0.18 0.63 0.58Reason: Loitering 0.06 0.03 0.09 0.07Reason: Other 1.31 1.16 1.35 1.39Reason: Suspicious 1.54 0.89 1.73 1.84Reason: Tra�c 1.18 0.81 1.39 1.30Total: 4.58 3.06 5.19 5.17

Arrests per 100 shifts:Civilian race: Black 5.64 4.95 5.93 5.91Civilian race: Hispanic 1.76 0.40 2.52 2.19Civilian race: White 0.89 0.32 0.98 1.17Reason: Tra�c 0.70 0.35 0.78 0.87Reason: Drug 0.94 0.36 1.24 1.13Reason: Other 3.10 2.04 3.46 3.54Reason: Property 1.52 1.15 1.66 1.68Reason: Violent 2.11 1.82 2.38 2.16Total: 8.37 5.71 9.52 9.36

Force per 1,000 shifts:Civilian race: Black 1.99 1.45 2.10 2.26Civilian race: Hispanic 0.41 0.09 0.58 0.50Civilian race: White 0.26 0.07 0.28 0.36Total: 2.71 1.63 3.02 3.19

13

Table C.2: Average events per shift, by o�cer gender. Mean number of stops, arrests, anduses of force without adjustment for time or location. Typical behavior is reported for female andmale o�cers separately, as well as the pooled average. Records associated with Native Ameri-can/Alaskan and Asian/Paci�c Islander o�cers are excluded for consistency with racial/ethnicanalyses. O�cer behavior toward Native American/Alaskan and Asian/Paci�c Islander civiliansis not included for the purposes of computing total and reason-speci�c events. Values are scaledfor ease of interpretation.

Mean Mean MeanBehavior (Pooled) (Female o�.) (Male o�.)

Stops per 10 shifts:Civilian race: Black 3.03 2.39 3.25Civilian race: Hispanic 0.92 0.65 1.02Civilian race: White 0.57 0.49 0.59Reason: Drug 0.48 0.27 0.55Reason: Loitering 0.06 0.04 0.07Reason: Other 1.31 1.23 1.34Reason: Suspicious 1.54 1.06 1.71Reason: Tra�c 1.18 0.99 1.25Total: 4.58 3.59 4.92

Arrests per 100 shifts:Civilian race: Black 5.64 4.03 6.18Civilian race: Hispanic 1.76 1.11 1.98Civilian race: White 0.89 0.69 0.95Reason: Tra�c 0.70 0.45 0.79Reason: Drug 0.94 0.47 1.09Reason: Other 3.10 2.09 3.44Reason: Property 1.52 1.21 1.63Reason: Violent 2.11 1.68 2.26Total: 8.37 5.90 9.20

Force per 1,000 shifts:Civilian race: Black 1.99 1.10 2.29Civilian race: Hispanic 0.41 0.20 0.48Civilian race: White 0.26 0.20 0.28Total: 2.71 1.52 3.11

14

Table C.3: Average events per shift forHispanic o�cers, by language ability. Mean numberof stops, arrests, and uses of force without adjustment for time or location. Typical behavior isreported for Spanish-speaking and non-Spanish-speaking Hispanic o�cers separately. O�cerbehavior toward Native American/Alaskan and Asian/Paci�c Islander civilians is not includedfor the purposes of computing total and reason-speci�c events. Values are scaled for ease ofinterpretation.

Mean MeanBehavior (Spanish Hisp. o�.) (Non-Spanish Hisp. o�.)

Stops per 10 shifts:Civilian race: Black 2.15 4.03Civilian race: Hispanic 1.50 1.37Civilian race: White 0.65 0.59Reason: Drug 0.53 0.72Reason: Loitering 0.06 0.11Reason: Other 1.25 1.47Reason: Suspicious 1.47 2.01Reason: Tra�c 1.05 1.74Total: 4.36 6.05

Arrests per 100 shifts:Civilian race: Black 3.99 7.84Civilian race: Hispanic 2.26 2.81Civilian race: White 0.98 0.99Reason: Tra�c 0.46 1.10Reason: Drug 0.72 1.76Reason: Other 2.54 4.38Reason: Property 1.55 1.80Reason: Violent 2.06 2.69Total: 7.33 11.73

Force per 1,000 shifts:Civilian race: Black 1.35 2.80Civilian race: Hispanic 0.52 0.66Civilian race: White 0.30 0.28Total: 2.21 3.79

15

C.2 Estimand

At a high level, the goal of our analysis is to evaluate the policy e�ect of a personnel reform thatincreases the representation of minorities in the CPD by assigning them to positions that wouldotherwise be �lled by white individuals. The analysis is conducted at the level of the patrolassignment slot. Commanding o�cers are assumed to have a �xed set of patrol assignments thatmust be �lled, where each slot is associated with a beat (geographic patrol area) and shift time(temporal window). Multiple slots may be available for a particular beat and shift time, but eachslot can be �lled by only one o�cer.

We organize beat assignments into groups, indexed by i, based on unique combinations ofmonth (Mi), day of week (Di), shift time (�rst/second/third watch, Si), and beat (Bi), or uniqueMDSBs. Patrol assignment slots within a MDSB are indexed by j. For each slot, the realizedpattern of o�cer behavior is denoted Yi,j(Ri,j), where Ri,j is the demographic pro�le (race/ethnicityand/or gender) of the o�cer assigned to a particular slot. Our notation implicitly makes the stableunit treatment value assumption [SUTVA, 38], which requires that (1) there do not exist �nergradations of o�cer identity (i.e., within the broad racial/ethnic and gender categories used) thatwould result in di�ering potential o�cer behavior, and (2) that potential outcomes do not verydepending on the racial/ethnic and gender identities of o�cers assigned to other slots.2

The slot-level policy e�ect is the di�erence in potential outcomes [37] Yi,j(r) − Yi,j(r ′), thechange in behavior that would have realized if an o�cer of demographic pro�le r had been as-signed to the patrol assignment slot, rather than another o�cer of pro�le r ′. These slot-levelcounterfactual di�erences are fundamentally unobservable. Instead, we target the average policye�ect within the subset of F MDSBs for which policy e�ects can be feasibly estimated (i.e., forwhich variation in o�cer demographic pro�les exists). This quantity is

� =1FAi

F∑i=1

Ai∑j=1Yi,j(r) − Yi,j(r ′),

where Ai is the number of patrol assignment slots available within MDSB i and Ai is the averageslot count across MDSBs. This can be rewritten as the weighted average of MDSB-speci�c e�ects,

2We explore the validity of this second assumption to the extent possible in SI Appendix C.8, in which stops madeby two o�cers are reanalyzed. In this section, we re-compute our estimates of di�erential stopping behavior afterexcluding the second reporting o�cer from our analysis; the resulting estimates are highly similar.

16

�i , with weights given by Ai .

� =F∑i=1

Ai

∑Fi′=1 Ai′ (

1Ai

Ai∑j=1Yi,j(r) − Yi,j(r ′))

=F∑i=1

Ai

∑Fi′=1 Ai′

�i .

As we discuss in Section 3, a key identifying assumption is that

Yi,j(r), Yi,j(r ′) ⟂⟂ Ri |Mi = m,Di = d, Si = s, Bi = b.

Informally, this requires that minority o�cers are not selectively assigned to slots within MDSBs,at least in ways that matter for potential o�cer behavior. (Hypothetically speaking, this inde-pendence condition could be achieved even without adjusting for MDSB if white and nonwhiteo�cers were randomly assigned locations and times to patrol.)

Our primary results estimate this quantity with an ordinary least squares (OLS) regression ofthe form Yi,j = �i +∑r �r1(Ri,j = r) + "i, j, where �i represents a �xed e�ect for MDSB i. Unbiased-ness of this estimator requires the additional assumption that MDSB-speci�c policy e�ects arehomogeneous, or that �i = � for all i. It is well known that when this assumption is violated, OLSrecovers the weighted average of �is with weights corresponding to the variance of o�cer de-mographic pro�les within strata. To allow for the possibility of non-homogeneous policy e�ectsand other departures from our modeling assumptions, we therefore apply a number of alterna-tive estimators, which are described in detail in SI Section C.7. As we show in SI Figures C.1–C.3,these alternative results are virtually identical to our primary results.

C.3 Potential Threats to Validity

For full transparency, we highlight a number of possible threats to the validity of our analysisgiven our analytic goal. Confounding factors in this scenario include all variables that correlatewith o�cer race and/or gender (depending on the analysis) in ways that violate the common-circumstances assumption. An example would be if Black and white o�cers were assigned to thesame beat and shift, but Black o�cers were ordered to stay in their patrol cars the entire timewhile white o�cers were allowed to freely roam the beat, meaning Black and white o�cers facedsystematically di�erent working conditions for reasons beyond their control. Another examplewould be if there are unobserved di�erences within a MDSB (e.g., a beat is more dangerous onone particular Tuesday evening in a month, perhaps due to a scheduled protest) and o�cers ofone group are preferentially assigned to patrol according to those di�erences. We assess thatconfounding of this type is extremely rare, because MDSB are de�ned in such a �ne-grained way

17

that comparisons are made within groups of roughly four beat-shifts (e.g., all Tuesday eveningsin January 2012 for beat 251).

However, because we are not seeking to identify the e�ect of race per se, other correlatesof o�cer race which do not violate the common-circumstances assumption do not obstruct ourability to evaluate this counterfactual. Examples of these innocuous correlates include: (1) Blackand white o�cers possessing di�erent levels of education which in turn lead to di�erential en-forcement; or (2) male and female o�cers choosing to focus on di�erent corners of their beatsonce assigned in ways that in�uence policing outcomes. In the latter case, o�cers still were as-signed to face common circumstances (our key identifying assumption) but chose to turn a blindeye to certain subsets of civilian behavior. These facets represent di�erent mechanisms throughwhich the policy intervention of interest a�ects police-civilian interactions, but would not biasestimates relating to o�cer deployment.

18

C.4 Shift Duration

We consider the possibility that stops, arrests, and uses of force are driven by di�erent amountsof time spent patrolling. Even among o�cers assigned to a particular shift time (a nominal eight-hour patrol period), minor variation exists in the precise start and end of the o�cer’s duty time.Of the o�cer-shifts analyzed, 85.5% are 9 hours in duration, with 8.5- and 8-hour shifts makingup an additional 7.8% and 5.1%, respectively (percentages are based on rounding shift durationto the nearest 6 minutes). In �xed-e�ect regression analyses that compare o�cers within uniqueMDSB combinations, we estimate that shifts of Black o�cers are 0.0094 hours shorter (roughly0.1% shorter) than their white counterparts assigned to the same MDSB, and Hispanic o�cer shiftdurations are statistically indistinguishable from those of white o�cers. Because these di�erencesare two orders of magnitude smaller than reported di�erences in behavior, patrol time disparitiesare unlikely to be a mechanism driving observed racial gaps in stops, arrests, and force.

19

C.5 Variation in Explanatory Variable

In analyzing how policing behavior varies with o�cer demographic characteristics, we comparethe recorded decisions of di�erent o�cers facing the same set of circumstances. To do so, weexamine 653,527 unique combinations of month, day of week, shift number, and beat (MDSBs).Of these, 572,067 MDSBs have more than one assigned o�cer, a requirement to make any within-MDSB comparison. Single-o�cer MDSBs can arise if, for example, a beat requires only one o�certo patrol and o�cer schedules are stable (e.g., if one individual consistently serves all �rst watchson Mondays for the month). To make cross-group comparisons, we further require that di�erento�cer groups have served in the same MDSB.

There are 294,963 MDSBs that contain overlap between multiple assigned o�cer racial/ethnicgroups (e.g., one Black o�cer and one white o�cer); 229,143 MDSBs contain overlap betweenboth female and male o�cers; and 49,721 MDSBs contain overlap between Spanish-speaking andnon-Spanish-speaking Hispanic o�cers. Due to the smaller number of Hispanic o�cers and theresulting low overlap rates, power for detecting di�erences between Spanish- and non-Spanish-speaking Hispanic o�cer behavior is relatively low.

20

C.6 Gender Heterogeneity

Here, we demonstrate that di�erences in behavior between male and female o�cers remain evenwhen comparing within members of a single racial/ethnic group. Female o�cers consistently useless force overall, and less force toward Black civilians in particular, than their co-racial/co-ethnicmale counterparts (all padj. ≤ 0.023). In addition, we �nd that Black female o�cers make slightlyfewer drug stops than Black male o�cers, though this gap is di�cult to interpret given the largernumber of stops for miscellaneous other reasons (both padj. < 0.001).

21

Tabl

eC.

4:W

ithineach

racial/ethnicgrou

p,o�

cers

ofdi�eringge

ndersbe

have

di�eren

tlywhe

nfacing

common

circum

-stan

ces.

Each

row

repo

rtsa

com

paris

onbe

twee

nfe

mal

ean

dm

ale

o�ce

rsof

apa

rticu

larr

acia

l/eth

nicg

roup

.Eac

hco

lum

nco

ntai

nsav

erag

eper

-shi

ftdi

�ere

nces

inap

olic

ebeh

avio

ram

ong

o�ce

rsas

signe

dto

thes

ameb

eata

ndsh

ifttim

e,du

ring

thes

amed

ayof

wee

kan

dm

onth

.Ast

erisk

sind

icat

eBe

njam

ini-H

ochb

ergp-

valu

esac

coun

ting

foro

�ce

r-le

velc

lust

erin

gan

dad

just

edfo

rmul

tiple

test

ing

of66

hypo

thes

es:∗

is<0.05

,∗∗

is<0.01

,and

∗∗∗

is<0.001.

Stop

spe

r10

shifts

(gen

derdi�eren

ceswithinracial/ethnicgrou

p)Ci

vilia

nra

ceRe

ason

Blac

kH

ispan

icW

hite

Oth

erLo

iterin

gSu

spic

ious

Dru

gTr

a�c

Tota

lBl

ack

fem

ale

o�ce

rs(v

smal

e)0.1

90.0

20.0

20.2

2∗∗∗

-0.00

-0.06

-0.04

∗∗∗

0.12

0.23

Hisp

anic

fem

ale

o�ce

rs(v

smal

e)-0

.09-0

.010.0

40.1

1-0

.02-0

.16-0

.030.0

5-0

.03W

hite

fem

ale

o�ce

rs(v

smal

e)-0

.03-0

.02-0

.000.1

1-0

.00-0

.16∗

-0.05

0.07

-0.04

Arrests

per10

0sh

ifts

(gen

derdi�eren

ceswithinracial/ethnicgrou

p)Ci

vilia

nra

ce:

Reas

on:

Blac

kH

ispan

icW

hite

Oth

erD

rug

Tra�

cPr

oper

tyVi

olen

tTo

tal

Blac

kfe

mal

eo�

cers

(vsm

ale)

-0.33

-0.01

0.00

-0.18

-0.05

0.03

-0.01

-0.13

-0.34

Hisp

anic

fem

ale

o�ce

rs(v

smal

e)-0

.62∗

-0.05

-0.01

-0.15

-0.20

-0.15

-0.10

-0.09

-0.69

Whi

tefe

mal

eo�

cers

(vsm

ale)

-0.40

-0.09

0.07

-0.24

-0.14

-0.01

0.04

-0.09

-0.44

Usesof

forcepe

r1,00

0sh

ifts

(gen

derdi�eren

ceswithinracial/ethnicgrou

p)Ci

vilia

nra

ceBl

ack

Hisp

anic

Whi

teTo

tal

Blac

kfe

mal

eo�

cers

(vsm

ale)

-0.52

∗-0

.01-0

.02-0

.54∗

Hisp

anic

fem

ale

o�ce

rs(v

smal

e)-1

.13∗∗∗

-0.37

-0.02

-1.52

∗∗

Whi

tefe

mal

eo�

cers

(vsm

ale)

-0.61

∗-0

.19-0

.00-0

.87∗

22

C.7 Alternative Estimators

Our primary analysis of o�cer behavior uses OLS regression with MDSB �xed e�ects, of theform Yi,j = �i + ∑r �r1(Ri,j = r) + "i,j , where �i represents a �xed e�ect for MDSB i. As wediscuss in SI Section C.2, this estimator will deviate from the desired average policy e�ect (i.e.,the average e�ect of replacing white o�cers assigned to a particular patrol assignment slot with aminority o�cer on resulting stop, arrest, and use-of-force volume) if MDSB-speci�c policy e�ectsare highly variable in a way that is associated with the proportion of minority/female o�cers thatare assigned to MDSBs (in this case, it is well known that OLS recovers the weighted average ofMDSB-speci�c policy e�ects, where weights are determined by variance of o�cer race withinthe MDSB).

To gauge robustness of our results to the violation of this assumption, we present alternativeestimates in SI Figures C.1–C.3 below. The �rst alternative estimator takes the within-MDSBdi�erence in behavior between average patrol assignments between given o�cer demographicpro�les, then aggregates these according to the number of patrol assignment slots in each MDSB.Following the notation de�ned in SI Section C.2, this estimator can be written as

F∑i=1

Ai

∑Fi′=1 Ai′

Ai∑j=1(

∑Aij=1 Yi,j 1(Di,j = d)∑Ai

j=1 1(Di,j = d)−∑Ai

j=1 Yi,j 1(Di,j = d ′)∑Ai

j=1 1(Di,j = d ′) ).

To assess the extent to which results are driven by large MDSBs, we further compute theunweighted average of MDSB-speci�c estimated e�ects:

1F

F∑i=1

Ai∑j=1(

∑Aij=1 Yi,j 1(Di,j = d)∑Ai

j=1 1(Di,j = d)−∑Ai

j=1 Yi,j 1(Di,j = d ′)∑Ai

j=1 1(Di,j = d ′) ).

Finally, we consider the possibility that observed demographic di�erences in o�cer behaviorare driven by di�erences in experience between o�cer groups. If this were the case, it wouldundermine the applicability of our results to the e�ect of a hiring reform that brought in additionalminority rookie o�cers. To examine whether these di�erences impact our results, we extend theregression speci�cation by adding additional linear and quadratic terms for each o�cer’s lengthof service. Speci�cally, we estimate Yi,j = �i +∑r �r1(Ri,j = r) + 1Si,j + 2S2i,j + "i,j , where Si,j is thelength of service of o�cer j as of the month corresponding to MDSB i occurred. This robustnesstest generally does not alter our substantive conclusions; the sole exception is that di�erencesin the number of arrests made by Spanish-speaking and non-Spanish-speaking Hispanic o�cersvanish when adjusting for length of service.

23

Figu

reC.

1:Stop

s.Po

ints

(err

orba

rs)d

epic

test

imat

eddi

�ere

nces

(95%

bloc

k-bo

otst

rap

con�

denc

eint

erva

ls)in

then

umbe

rofc

ivili

anst

ops

cond

ucte

dby

Blac

kan

dH

ispan

ico�

cers

pers

hift,

rela

tive

tow

hite

o�ce

rspa

trolli

ngin

the

sam

em

onth

,day

grou

p(w

eek-

day/

wee

kend

),sh

ifttim

e(�

rst/s

econ

d/th

irdw

atch

),an

dbe

at.R

esul

tspr

esen

ted

usin

gfo

urdi

�ere

ntes

timat

ors.

His

panic

off

icers

(vs

whit

e)

Sp

anis

h-s

peaki

ng

His

panic

off

icers

(vs

non-s

peaki

ng

)

Bla

ck o

ffic

ers

(vs

whit

e)

Fem

ale

off

icers

(vs

male

)

Civ. r

ace:

Bla

ck

Civ. r

ace:

Hisp

anic

Civ. r

ace:

Whi

te

Reaso

n: O

ther

Reaso

n: L

oite

ring

Reaso

n: S

uspi

ciou

sRea

son:

Dru

g

Reaso

n: T

raffi

c

Tota

lCiv

. rac

e: B

lack

Civ. r

ace:

Hisp

anic

Civ. r

ace:

Whi

te

Reaso

n: O

ther

Reaso

n: L

oite

ring

Reaso

n: S

uspi

ciou

sRea

son:

Dru

g

Reaso

n: T

raffi

c

Tota

l

-0.2

-0.1

0.0

0.1

-0.2

-0.1

0.0

0.1

Differences in stops per shift byofficers of each racial/ethnic group,

versus white officers facing CC

Count-

weig

hte

d m

ean o

f w

ithin

-MD

SB

diffe

rence

sO

LS w

ith M

DS

B F

EO

LS w

ith M

DS

B F

E a

nd

exp

eri

ence

contr

ols

Sim

ple

mean o

f w

ithin

-MD

SB

diffe

rence

s

24

Figu

reC.

2:Arrests.

Poin

ts(e

rror

bars

)dep

icte

stim

ated

di�e

renc

es(9

5%bl

ock-

boot

stra

pco

n�de

nce

inte

rval

s)in

the

num

ber

ofar

rest

scon

duct

edby

Blac

kan

dH

ispan

ico�

cers

pers

hift,

rela

tive

tow

hite

o�ce

rspa

trolli

ngin

the

sam

em

onth

,day

grou

p(w

eek-

day/

wee

kend

),sh

ifttim

e(�

rst/s

econ

d/th

irdw

atch

),an

dbe

at.R

esul

tspr

esen

ted

usin

gfo

urdi

�ere

ntes

timat

ors.

His

panic

off

icers

(vs

whit

e)

Sp

anis

h-s

peaki

ng

His

panic

off

icers

(vs

non-s

peaki

ng

)

Bla

ck o

ffic

ers

(vs

whit

e)

Fem

ale

off

icers

(vs

male

)

Civ. r

ace:

Bla

ck

Civ. r

ace:

Hisp

anic

Civ. r

ace:

Whi

te

Reaso

n: O

ther

Reaso

n: D

rug

Reaso

n: T

raffi

c

Reaso

n: P

rope

rty

Reaso

n: V

iole

nt

Tota

lCiv

. rac

e: B

lack

Civ. r

ace:

Hisp

anic

Civ. r

ace:

Whi

te

Reaso

n: O

ther

Reaso

n: D

rug

Reaso

n: T

raffi

c

Reaso

n: P

rope

rty

Reaso

n: V

iole

nt

Tota

l

-0.2

-0.1

0.0

0.1

-0.2

-0.1

0.0

0.1

Differences in arrests per 10 shifts byofficers of each racial/ethnic group,

versus white officers facing CC

Count-

weig

hte

d m

ean o

f w

ithin

-MD

SB

diffe

rence

sO

LS w

ith M

DS

B F

EO

LS w

ith M

DS

B F

E a

nd

exp

eri

ence

contr

ols

Sim

ple

mean o

f w

ithin

-MD

SB

diffe

rence

s

25

Figu

reC.

3:Usesof

Force.

Poin

ts(e

rror

bars

)dep

icte

stim

ated

di�e

renc

es(9

5%bl

ock-

boot

stra

pco

n�de

nce

inte

rval

s)in

the

num

ber

ofus

esof

forc

eby

Blac

kan

dH

ispan

ico�

cers

per

shift

,rel

ativ

eto

whi

teo�

cers

patro

lling

inth

esa

me

mon

th,d

aygr

oup

(wee

k-da

y/w

eeke

nd),

shift

time

(�rs

t/sec

ond/

third

wat

ch),

and

beat

.Res

ults

pres

ente

dus

ing

four

di�e

rent

estim

ator

s.

His

panic

off

icers

(vs

whit

e)

Sp

anis

h-s

peaki

ng

His

panic

off

icers

(vs

non-s

peaki

ng

)

Bla

ck o

ffic

ers

(vs

whit

e)

Fem

ale

off

icers

(vs

male

)

Civ. r

ace:

Bla

ckCiv

. rac

e: H

ispan

icCiv

. rac

e: W

hite

Tota

l

Civ. r

ace:

Bla

ckCiv

. rac

e: H

ispan

icCiv

. rac

e: W

hite

Tota

l

-0.1

0.0

0.1

0.2

-0.1

0.0

0.1

0.2

Differences in uses of force per 100 shiftsby officers of each racial/ethnic group,

versus white officers facing CC

Count-

weig

hte

d m

ean o

f w

ithin

-MD

SB

diffe

rence

sO

LS w

ith M

DS

B F

EO

LS w

ith M

DS

B F

E a

nd

exp

eri

ence

contr

ols

Sim

ple

mean o

f w

ithin

-MD

SB

diffe

rence

s

26

C.8 Robustness Checks: Multiple Stopping O�cers

Data on stops of civilians indicate that the vast majority of enforcement is jointly conducted bytwo o�cers, but one is listed as the primary o�cer in police records. (Arrest records also oftencontain more than one o�cer, but contain have no indication of the lead o�cer.) In our mainanalysis, we treat a stop by two o�cers as two incidents in the data, as both o�cers contributeto the decision to engage a civilian. To gauge the extent to which this decision drives our re-sults, we present an alternative analysis of stops in which we use only data on �rst reportingo�cers, respectively. Results are substantively unchanged. Note that the reduction in female-o�cer drug stops (versus male o�cers) loses signi�cance, and a handful of other comparisonsappear to become marginally signi�cant. However, general patterns remain substantively iden-tical, with smaller coe�cients re�ecting the fact that roughly half of all stop events have beendiscarded. Given the lack of information on arresting o�cer roles, we do not conduct a similarrobustness test for the arrest analysis. Our only option would be to drop an o�cer from eacharrest at random, which would in expectation merely produce identical patterns with attenuatedcoe�cients.

27

Tabl

eC.

5:Rob

ustnessof

demog

raph

icdi�eren

cesto

analyz

ingon

estop

ping

o�cerpe

ren

coun

ter..

Follo

win

gTa

ble

2,ea

chro

wre

ports

aco

mpa

rison

betw

een

o�ce

rsof

vario

usde

mog

raph

icgr

oups

.Eac

hco

lum

nco

ntai

nsav

erag

epe

r-sh

iftdi

�ere

nces

ina

polic

ebe

havi

oram

ong

o�ce

rsas

signe

dto

the

sam

ebe

atan

dsh

ifttim

e,du

ring

the

sam

em

onth

and

day

ofw

eek.

Aste

risks

indi

cate

Benj

amin

i-Hoc

hber

gp-

valu

esac

coun

ting

foro

�ce

r-le

velc

lust

erin

gan

dad

just

edfo

rmul

tiple

test

ing:

∗is<0.05

,∗∗

is<0.01

,and

∗∗∗

is<0.001.

Resu

ltsar

esu

bsta

ntiv

ely

iden

tical

toTa

ble

2.

Stop

spe

r10

shifts

Civi

lian

race

Reas

onBl

ack

Hisp

anic

Whi

teO

ther

Loite

ring

Susp

icio

usD

rug

Tra�

cTo

tal

Blac

ko�

cers

(vsw

hite

)-0

.88∗∗∗

-0.11

∗∗∗

-0.10

∗∗∗

-0.21

∗∗∗

-0.02

∗∗∗

-0.42

∗∗∗

-0.12

∗∗∗

-0.32

∗∗∗

-1.09

∗∗∗

Hisp

anic

o�ce

rs(v

swhi

te)

-0.34

∗∗∗

-0.01

-0.05

∗∗∗

-0.05

0.01

-0.20

∗∗∗

-0.05

∗∗-0

.10∗

-0.40

∗∗∗

Fem

ale

o�ce

rs(v

smal

e)0.1

20.0

40.0

30.1

9∗∗∗

-0.00

-0.10

∗∗-0

.010.1

2∗∗∗

0.20∗

Span

ish-s

peak

ing

Hisp

anic

o�ce

rs-0

.01-0

.080.0

10.0

6-0

.04-0

.04-0

.13∗

0.06

-0.08

(vsn

on-S

pani

sh-s

peak

ing)

28