the impact of racial and ethnic diversity in policing · the impact of racial and ethnic diversity...
TRANSCRIPT
The Impact of Racial and Ethnic Diversity inPolicing∗
Bocar Ba† Dean Knox‡ Jonathan Mummolo§ Roman Rivera¶
February 15, 2020
Abstract
Racial and ethnic diversi�cation is among the oldest proposed police reforms, butsevere data constraints have forced scholars to rely almost exclusively on problematiccross-agency comparisons to evaluate its e�ects. We assess the impact of diversi�-cation using millions of observations of Chicago police o�cers’ behavior, mergedwith data on their socio-demographic traits. Compared with white o�cers facingthe same working conditions, we show that black o�cers make substantially fewerdiscretionary stops (e.g. for “suspicious behavior”), fewer arrests for petty crimes,and use force less often. These reductions stem primarily from a reduced focus onblack civilians, a pattern concordant with e�orts to end abusive policing and massincarceration in African American communities. However, we also �nd evidencethat coarse identity groupings mask large variation in police behavior. Most notably,we �nd Hispanic o�cers who cannot speak Spanish—the native language of manyco-ethnic communities—stop Hispanic residents at the highest rates. Taken together,these results show the substantial impact of diversity on police treatment of minoritycommunities, and emphasize the need to consider multiple facets of police o�cerswhen crafting personnel-driven reforms.
∗Data for this project were gathered in collaboration with Invisible Institute, with the help of Sam Steck-low and Emma Herman. We thank Rachel Ryley, the Lucy Parsons Lab, and L2, Inc. for generously sharingadditional data, and Ted Enamorado for research assistance.
†Postdoctoral associate, University of Pennsylvania, Law School, [email protected].‡Assistant Professor of Politics, Princeton University, [email protected].§Assistant Professor of Politics and Public A�airs, Princeton University, [email protected].¶PhD student, Dept. of Economics at Columbia University, [email protected].
Racial disparities in the volume and nature of police-civilian interactions (Gelman,Fagan and Kiss, 2007; Knox, Lowe and Mummolo, Forthcoming; Lerman and Weaver, 2014)have for decades fueled persistent allegations of racial bias in policing. Central to thesecritiques are the fact that throughout most of U.S. history—from the formation of the�rst organized security patrols on slave plantations (Reichel, 1988; Turner, Giacopassiand Vandiver, 2006), to the concentrated implementation of “Broken Windows” policingin communities of color in recent decades (Wilson and Kelling, 1982)—many police forceshave been comprised of nearly all white o�cers (Forman Jr., 2017). In turn, one of themost often proposed police reforms aimed at reducing racial inequities in policing hasbeen to increase racial diversity within law enforcement agencies (DOJ/EEOC, 2016).
But despite a large body of research in economics (Antonovics and Knight, 2009), crim-inology (Cao and Cullen, 1996; Shusta and Wong, 1995), law (Forman Jr., 2017), and po-litical science (Garner, Harvey and Johnson, 2019), the e�cacy of this reform strategyremains contested. In some cases, scholars have concluded diversity has no detectableimpact on police violence (Smith, 2003). Other prominent studies conclude the reformchanges the racial composition of arrested civilians (Levitt, 2012; McCrary, 2007). In anexhaustive review of the empirical literature on racial diversity in policing, one legalscholar concluded: “[t]he fairest summary of the evidence is probably that we simplydo not know” (Sklansky, 2005, 1225).
The central obstacle to evaluating this reform has been a lack of su�ciently �ne-grained data on police o�cer deployment and behavior. A comprehensive assessmentrequires analysts to test whether minority police o�cers behave di�erently from whiteo�cers when faced with otherwise identical circumstances. Without the ability to makesuch careful comparisons, virtually all prominent work in this area has relied on cross-agency comparisons (e.g. Legewie and Fagan, 2016; Donohue III and Levitt, 2001; Smith,2003). This approach cannot distinguish whether apparent di�erences in o�cer behaviorarise due to selection (e.g., diverse agencies di�er on unmeasured traits) or ecological falla-cies (Johnson et al., 2019) (e.g., white and nonwhite o�cers in a jurisdictions are assignedto work in di�erent circumstances, which we demonstrate is in fact true). In addition,even the handful of studies that have leveraged plausibly exogenous variation in agencyracial composition (McCrary, 2007; Garner, Harvey and Johnson, 2019; Harvey and Mat-tia, 2019)1 mask important details of police-civilian interactions and the contexts in which
1McCrary (2007),Garner, Harvey and Johnson (2019) and Harvey and Mattia (2019) employ di�erence-in-di�erences designs and leverage the timing of court-ordered a�rmative action mandates on law enforce-
1
they unfold. Most prior work also lacks basic information on police and civilians, much ofwhich is relevant to evaluating this policy. For example, skeptics of diversi�cation arguethat minorities choosing to apply to police academies are a self-selected group that maynot mirror the socio-demographic backgrounds and political views of the broader minor-ity population. This critique is in line with much prior work on selection in bureaucraticsta�ng (Clinton and Lewis, 2008; Forman Jr., 2017; Lewis, 2010; Linos and Riesch, 2019;Wilson, 1989). In other words, regardless of an o�cer’s race and ethnicity, it may be thecase that “blue is blue.” If so, diversity-based hiring reforms may not deliver meaningfulchange.
In this paper, we draw on a host of newly collected data sets that together allow us toovercome these longstanding limitations. Our data, which track more than 33,000 Chicagopolice o�cers over time, was assembled through years of open records requests and law-suits and allows for the most credible and �ne-grained assessment of the impact of o�cerdiversity on police behavior to date. It includes timestamped and geolocated records ofo�cers’ decisions to stop, arrest, and use force, as well as civilian complaints that arisefrom o�cer behavior. After aggressively pruning these data to maximize the validity ofour analyses, we construct a panel of 2.9 million detailed observations of events linkedto individual o�cers and speci�c patrol assignments. Our data further includes infor-mation on o�cers’ demographics, language skills, places of residence, political prefer-ences/participation, and career progressions. Most crucial for the validity of our policyevaluation, we leverage data on o�cers’ patrol assignments—i.e., the speci�c days andshifts that o�cers were assigned to patrol speci�c beats, or narrow sets of city blocks.This feature allows us to compare o�cers of di�erent racial and ethnic groups who areworking under the same conditions even when they do not record any enforcement activi-ties. This means we can e�ectively control for environmental factors including di�erentialexposure to civilians across o�cers, overcoming a fundamental limitation in prior work(Knox, Lowe and Mummolo, Forthcoming; Knox and Mummolo, 2020). In short, our dataallow us to credibly approximate, for the �rst time, whether and how white and nonwhiteo�cers perform their jobs di�erently.
We present �ve broad conclusions, several of which undermine prior studies that reliedon coarsely aggregated data at the city or agency level:
(1) Minority o�cers are assigned to very di�erent districts, and even within districts,
ment agencies to estimate agency-level e�ects of o�cer diversity on various outcomes including crime andcrime victimization.
2
receive vastly di�erent geographic and temporal patrol assignments. Without accountingfor these di�erences in working conditions, there is no way to meaningfully characterizethe di�erences between minority and white o�cer behavior.
(2) Like Chicago city residents, white, black and Hispanic o�cers tend to live in areasof the city with high shares of co-racial and co-ethnic civilians, a pattern which suggestslow levels of inter-group contact when o�-duty. Black o�cers also have similar politicalpreferences to black civilians, while Hispanic and white o�cers are substantially morepolitically conservative than their civilian counterparts.
(3) Black o�cers make roughly 31% fewer stops and 22% fewer arrests, and they useforce 35% less often, relative to white o�cers working in the same places and times. Ex-amining a wide range of o�cer activity, we show this is mostly driven by a decreasein discretionary activity (e.g. a 53% decrease in loitering stops and 30% fewer drug ar-rests), rather than lower enforcement of severe criminal behavior (only a 10% decrease inviolent-crime arrests). Moreover, relatively higher levels of enforcement by white o�cersfall primarily on black civilians. Black o�cers stop and use force against black civiliansat a rate that is roughly 40% lower than white o�cers.
(4) Hispanic o�cers behave comparably to white o�cers in the aggregate. But thereis massive variation in behavior among Hispanic o�cers. Those who can speak Spanishmake signi�cantly fewer stops and arrests than o�cers who cannot speak Spanish, and aretherefore less likely to have cultural ties to immigrant communities. In fact non-Spanish-speaking o�cers stop Hispanic civilians at even higher rates than white o�cers.
(5) Despite clear evidence that black o�cers make fewer stops and arrests and use forceless often than white o�cers, we �nd civilians lodge complaints against black and whiteo�cers working in the same places and times in roughly equal number overall. However,black o�cers (relative to white o�cers) and Spanish-speaking Hispanic o�cers (relativeto non-Spanish-speaking Hispanic o�cers) receive substantially more complaints for mi-nor infractions, a category that includes failure to �le paperwork, while appearing com-parable in numerous other categories. These patterns suggest the possibility of bias on thepart of civilians when lodging complaints against some groups of minority o�cers. How-ever, due to data limitations on civilian complaints which we outline below, we presentthese results as suggestive, requiring further investigation.
Overall, our results suggest that diversi�cation has a substantial impact on the treat-ment of minority civilians. However, our �ndings also suggest the need to consider mul-tiple facets of police o�cers when crafting personnel-driven reforms. Like the civilians
3
they police, o�cers are complex, not easily reducible to racial or ethnic categories. Schol-ars and reformers must take stock of other factors that may a�ect how o�cers relate toresidents, including political preferences and language skills, to develop e�ective policy.
All data and interactive replication materials are publicly available at [LINK]. We en-courage readers to probe and extend our analyses.
1 Data
Over a period of three years, we submitted open records requests to the Chicago PoliceDepartment (CPD) and the city’s Department of Human Resources. The resulting recordsinclude the name, race, gender, birth year, salary, language skill, unit assignments andappointment date of each o�cer, as well as behavioral outcomes including arrests, usesof force, and associated complaints (Ba, 2019; Rim, Rivera and Ba, 2019). We merged thesedata with records on stops of civilians and o�cer beat assignments, proprietary recordson voter turnout and party registration,2 and U.S. census data. SI A.1 describes these datasets and our pre-processing procedures in detail.
We begin with 4.2 million stops in the period of interest, including stops by o�cers onspecial duty, stops that occurred outside their assigned shift times, or other idiosyncraticcircumstances that make principled comparisons di�cult. After aggressive pruning tomaximize the validity of our comparisons, we retain 1.3 million stop records linked to in-dividual o�cers and speci�c patrol assignments. Similarly, after merging and pruning, weretain records of nearly 250,000 arrests and 8,000 uses of force. In this period, we matchroughly 6,000 records of civilian complaints (these complaint records are often incom-plete, a feature we discuss further below). Table 1 contains aggregate information on fourbehaviors tabulated by o�cer race: stops, arrests, uses of force, and civilian complaintsfrom Jan. 2012 to Dec. 2015, the period that informs our behavioral analysis. Due to thesmall sample sizes associated with o�cer groups including Asian Americans and NativeAmericans, our analysis is limited to three racial and ethnic groups: white, black and His-panic o�cers, (together accounting for roughly 97% of o�cers for whom shift assignmentdata is available).
Figure 1 illustrates the dataset’s various attributes by highlighting a small slice of its2We are grateful to Lucy Parsons Labs for publicly releasing data on civilian stops, Rachel Ryley for
generously sharing data on beat assignments, and L2, Inc. for allowing access to its national voter �ledatabase.
4
Table 1: Summary of data on o�cer behavior (counts), 2012-2015
Black Hispanic Whiteo�cers o�cers o�cers
Stops 253,609 356,541 729,078Arrests 47,406 65,587 132,285
Uses of force 1,355 2,081 4,514Civilian complaints 1,661 1,345 2,922
temporal and geographic coverage. The �gure maps activity during a four-month windowin the CPD’s Wentworth District (District 2), a highly segregated 7.5 square mile territoryon Chicago’s South Side that is 96% black and consistently ranks among the city’s mostviolent districts in per-capita crime rates. The district is comprised of 15 beats (small col-lections of city blocks assigned for patrols),3 which are shaded according to their racialcomposition. Points distinguish between geolocated stops, arrests, uses of force and civil-ian complaints during this period.
The �gure also describes the behavior of four CPD o�cers working in District 2 duringthis time, renamed to maintain anonymity. For example, one o�cer is female, Hispanic,speaks both English and Spanish, has no political party a�liation, and was born in 1965;another is a white male Republican who does not speak Spanish, born in 1981. The �gureshows the speci�c beats to which o�cers were assigned over time, each o�cer’s behav-ior while working in these beats, the political party a�liation of each o�cer (Democrat,Republican, or una�liated), and whether they turned out to vote in the 2012 Novembergeneral election. The �gure also highlights particular events associated with these o�-cers. For example, one summary details an incident that occurred on Oct. 8, 2012 at 7:13p.m. in which an o�cer used force against a 20-year old unarmed black male. The eventoccurred outdoors, during clear weather in a vacant lot, and the civilian was not injuredduring the encounter. The o�cer, who was not in uniform during the event, was also notinjured. The precise location of the event is denoted on the map in an encircled red “X.”
With data this granular, we are not only able to describe the makeup of the police forcerelative to the city population on a host of dimensions, providing a rich context absent inprior work. We can also make the most valid comparisons to date between the behaviorof white and nonwhite o�cers facing common working conditions. To accomplish this,in our behavioral analysis, we restrict our data to the years 2012-2015, the years covered
3Based on Chicago Police Department 2008 Annual Report.
5
by our patrol assignment data. We further restrict analysis to the 87% of o�cers employedat the rank “police o�cer” (pay grade D|1) to isolate o�cers who work primarily on citystreets, rather than behind desks or in atypical environments (e.g. excluding o�cers atrank “helicopter pilot” or “superintendent’s chief of sta�”).4
In the next section, we further detail our analytic strategy.
4This is one of the few o�cer level covariates we condition on in our analysis. We do this because itrepresents a potential violation of our key assumption: that white and nonwhite o�cers common externalcircumstances within month-day-shift-beat units. We detail this rationale in the following section.
6
Figure 1: Detailed view of the data. The right panel �gure maps police activity in a single police district (Wentworth, CPDDistrict 2), with green circles, blue squares, and red crosses respectively indicating the locations of stops, arrests, and uses offorce. Polygons represent geographic beats and are shaded by their proportion of minority residents. Lower left panels chartthe behavior of four anonymized o�cers over a four-month period, with panel headers indicating age, gender, ethnicity/race,language ability, and political orientation. In addition to mapped events, these panels indicate general election turnout andcivilian-�led complaints with purple triangles and orange asterisks, respectively. Encircled incidents are described further inthe left middle panel, which report civilian and incident speci�cs. Finally, the top left panel indicates how the four selectedo�cers are assigned to patrol beats over dates and times, with vertical gray bars indicating weekends.
7
2 Empirical Strategy
Our main goal is to assess the impact of o�cer diversity on the treatment of civilians.Speci�cally, we seek to test whether o�cers of various racial and ethnic groups per-form their jobs di�erently, on average, when faced with the same circumstances. How-ever, previous policing studies have lacked information on the context in which police-civilian interactions occur, as well as insight into how o�cers relate to civilians on socio-demographic dimensions.
Our analysis therefore proceeds in two main parts. The �rst is a descriptive analy-sis of the concordance between several features of o�cers and the civilians they police.Beyond shared racial and ethnic identity, these include place of residence, political pref-erences and political participation. Many reformers have asserted the importance of ad-ditional factors that may condition the impact of racial diversity, such as the frequencywith which o�cers interact with co-racial and ethnic civilians, whether o�cers reside inthe communities they patrol (Goodyear, 2014) and whether o�cers share common goalswith respect to police reform, among other attributes (Lockwood and Prohaska, 2015).Indeed, there are strong reasons to suspect that racial diversi�cation alone will not bringabout meaningful change depending on the prevalence and distribution of these otherfeatures. Hiring conservative black o�cers who believe policing needs no substantial re-form (Eckhouse, 2019; Federico and Holmes, 2005; Weaver, 2007; Weitzer and Tuch, 2004),or Hispanic o�cers who cannot speak the native language of many immigrant commu-nities, for example, may do little to transform the nature of police-civilian interactions.Our analysis thus begins with a rich description of how o�cers relate to the civilians theypolice on a host of dimensions.
Following this descriptive assessment, we proceed to our main analysis, presentingcomparisons of the behavior of o�cers belonging to various racial and ethnic groupswho are working under virtually identical conditions. Unless white and nonwhite o�cersbehave di�erently on the job, it is unlikely that diversi�cation will yield tangible bene�tsfor policed populations. The ideal experiment to evaluate this question would be to ran-domize white and nonwhite o�cer into patrol assignments. Given this randomization,white and nonwhite o�cers would encounter the same on-average conditions, and ob-served behavior across o�cer groups could inform a valid causal assessment of the policyreform of interest. (For a formal discussion of our quantity of interest, see SI Section B.)
While we cannot directly implement this experiment, our data allow us to recover a
8
close approximation. We assemble a vast panel data set, where each row contains infor-mation on a particular o�cer’s activities during a particular shift, including contextualinformation on those activities.5 In each of these 2.9 million patrol assignments, we com-pute the volume of stops, arrests, uses of force, and civilian complaints. We then compareo�cers of di�erent racial and ethnic groups working in the same unique month, day group(weekend vs. weekday), shift number (one of three nominal eight-hour watches eachday), and beat (a small collection of city blocks, averaging 0.82 square miles citywide)—extremely narrow slices of time and space that we label MDSBs for short. Because we arecomparing o�cers working in these narrow units, we can credibly assume that observeddi�erences in o�cer behavior are not due to disparities in external conditions. That is, weassume that white and nonwhite o�cers within MDSBs face the same set of enforcementopportunities, and are faced with: the same types of civilian activity; common neigh-borhood attributes such as infrastructure and architecture; and common time-varyingconditions such as seasons and lighting. We term this the “common circumstances” (CC)assumption. For ease of interpretation, we present di�erences estimated using ordinaryleast squares regression with MDSB �xed e�ects, though our results are robust to severalother estimators (see SI Figures A6-A9 for these additional results). All statistical infer-ences are based on o�cer-level block bootstrap con�dence intervals that are robust tounobserved o�cer-speci�c peculiarities. Our behavioral analysis organizes patrol assign-ments into nearly 207,000 groups de�ned by MDSBs, of which roughly 121,000 containassigned o�cers of di�erent ethno-racial groups (i.e., vary on the explanatory variable ofinterest and thus contribute to our estimates).
Importantly, because we have data on micro-level assignments, we are able to retaino�cers in our panel analysis in every time period whether or not they engaged in anyenforcement activity. Here, our analysis di�ers from most prior studies of police admin-istrative data, which only rely on instances in which o�cers recorded an activity (e.g.a stop or arrest; see discussion in Knox, Lowe and Mummolo (Forthcoming)). Withoutthese patrol assignment records, inaction is invisible to the analyst, and o�cers who donot record any enforcement activity simply vanish from the data. To our knowledge, no
5In analyzing o�cer behavior and associated complaints, we take patrol assignments—the level at whichcommanding o�cers typically exercise control—as the basic unit of analysis. In SI Section C.5, we examinethe possibility that di�erences in clock-in/out times may be one mechanism behind observed di�erencesin behavior between o�cer racial groups. While some statistically signi�cant di�erences can be observed(roughly 0.1% disparities in patrol time), these gaps are two orders of magnitude smaller than the di�erencesobserved in stops, arrests, uses of force, and civilian complaints.
9
previous study has been able to leverage patrol assignments in this fashion.6
We stress that our goal is to estimate the �rst-order impact of a policy intervention—the counterfactual substitution of an o�cer of di�ering race into a patrol assignment.(This conceptual manipulation di�ers substantially from the imagined scenario of chang-ing an o�cer’s race or ethnicity while holding their other attributes �xed, an arguablyimpossible counterfactual (Greiner and Rubin, 2011; Holland, 1986).) The policy inter-vention we examine may have second-order e�ects as well. For example, the presenceof additional nonwhite o�cers could a�ect department culture in ways that eventuallyalter the nature of policing. Such e�ects are far more di�cult to evaluate rigorously andbeyond the scope of our analysis.
For full transparency, we highlight a number of possible threats to the validity ofour analysis given our analytic goal. Confounding factors in this scenario include allvariables that correlate with o�cer race in ways that violate the CC assumption. Anexample would be if black and white o�cers were assigned to the same beat and shift, butblack o�cers were ordered to stay in their patrol cars the entire time while white o�cerswere allowed to freely roam the beat. Another violation, given that we control for daygroup but not day of week, would occur if white and nonwhite o�cers were systematicallyassigned to work particular days of the week in ways that correlated with criminal activity.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 CC assumption do not obstruct our ability toevaluate 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�erentialenforcement; or (2) black and white o�cers choosing to focus on di�erent corners of theirbeats once assigned in ways that in�uence policing outcomes. These facets of o�cer racewould represent di�erent mechanisms through which the policy intervention of interesta�ects police-civilian interactions, but would not bias causal estimates relating to minorityo�cer deployment. (For a related discussion of conceptualizing race as a causal variable,see Knox, Lowe and Mummolo (Forthcoming) and Sen and Wasow (2016)).
This analysis, while a substantial advance over prior work, has important limitations.First, we measure relative di�erences in o�cer behavior, which do not necessarily implyracial bias. For example, if we were to �nd that white and nonwhite o�cers use force at
6In some cases, our data indicate that two o�cers were involved in stopping a civilian. Throughout,such instances enter as separate rows in our o�cer-shift data, since in such cases each o�cer made a stop.However, to gauge robustness, Si Section C.10 presents a reanalysis of stops after restricting the data to the�rst reporting o�cer, which yields essentially identical results.
10
equal rates against nonwhite civilians under identical circumstances, that may indicateno racial bias. However, it may also indicate a high but equal level of bias across o�cers.A second limitation is that we cannot identify speci�c mechanisms connecting o�cerrace to behavior. The behavioral di�erences we observe could be due to di�erences inhow o�cers engage civilians, di�erences in how civilians behave toward certain o�cergroups, or some combination of the two. The e�ects we demonstrate still inform thee�cacy of the policy of racial diversi�cation, but the mechanisms behind these e�ectsdeserve further study, a point to which we return in the conclusion.
In addition, while our analysis tests for tangible e�ects of diversi�cation, we empha-size this policy reform may yield additional intangible bene�ts, namely, increased trustin government among minorities. A long literature has argued that trust in police is acrucial element of public safety, compelling residents to more readily cooperate with lawenforcement (e.g. Tyler and Fagan, 2008) and thereby potentially increasing the chancesthat o�cers solve crimes. If racial diversity a�ects levels of public trust, the cost-bene�tcalculus when considering this policy would be further complicated.
Finally, our analysis is con�ned to a single city, a feature that a�ords us unusuallydetailed data at the expense of geographic scope. Given the substantial progress our dataallow us to make in quantifying the impact of this policy, we view this as a worthwhiletrade. These data also emanate from a setting that is ideal for investigating the role ofo�cer diversity in shaping police-civilian interactions: both Chicago and its police de-partment are racially diverse, and the city has come under �re for a range of aggressivepolice tactics and allegations of misconduct, making it an important case for the studyof a widely proposed police reform (Rivera and Ba, 2019). Our analysis also provides atemplate for future work to be conducted in other jurisdictions.
3 Racial Representation in Chicago
The CPD, which currently employs about 12,000 sworn o�cers, exhibits substantial levelsof racial and ethnic diversity, making it a useful test case for the e�cacy of this policy.However, this has not always been the case: the department has slowly and steadily di-versi�ed over time. Figure 2 visualizes this process. The �gure’s left panel plots the shareof CPD employees belonging to each racial and ethnic group since 1970, as well as thecorresponding shares of the city population belonging to those groups (see SI SectionA.2 for details on the coding of race and ethnicity). The right panel compares the share of
11
new CPD recruits belonging to racial and ethnic groups to the city’s population over time.White o�cers have historically made up a disproportionate share of the CPD, but theirrepresentation has been steadily declining, dropping from 78% (61%) to 54% (36%) of theagency (city) between 1970 and 2010. These declines in white representation have beeno�set by gains by minority groups, with sharp rates of increase for Hispanics in particularin both the CPD and among city residents. At present, the CPD is roughly 24% black and21% Hispanic. The �gure also shows that blacks have been declining as a share of cityresidents since 1980, which has likely contributed to stagnant growth in the CPD’s shareof black o�cers. In recent years, only about 15% of new CPD recruits have been black,down from a high during this period of nearly 35% in the late 1970s. While the agencydoes not perfectly mirror the racial composition of the city, it does contain substantialshares of nonwhite o�cers—a necessary feature for our evaluation.
4 Race and geographic assignment
Given Chicago’s high levels of racial segregation7 among residents, it is unclear the extentto which o�cers interact with co-racial and co-ethnic civilians while on duty. If suchinteractions were infrequent, the impact of diversi�cation could be limited. To assess this,Figure 3 plots the the composition of police districts (25 sectors of the city, each covering9.25 square miles, on average) against the composition of o�cers assigned to work inthose districts. In general, white-dominated districts have virtually no minority o�cersassigned, and districts with sizeable minority populations tend to have more o�cers ofthe corresponding race. However, a number of districts dominated by black residentsnonetheless have sizeable contingents of white o�cers. For example, Wentworth (CPDDistrict 2, depicted in Figure 1) is 95% black, but 24% of o�cers assigned there are white.The disparity is even starker in Austin (CPD District 15), where a 93% black residentpopulation is policed by a unit that is 55% white. (See SI Section A.3 for details, and SIFigure A1 for additional analyses.)
Further, among o�cers assigned to a particular district, considerable variation exists inthe exact patrol assignments that o�cers receive. To show this, we examine each unit indi-vidually, tabulating o�cer race and shift time assignments (�rst, second and third watch,respectively corresponding to the nominal duty periods of midnight to 8 a.m., 8 a.m. to4 p.m., and 4 p.m. to midnight). SI Figure A2 depicts the frequency of each shift period,
7See http://www.censusscope.org/us/print_rank_dissimilarity_white_black.html.
12
Figure 2: Composition of CPD o�cers and city residents over time. The left paneldepicts the proportion of active CPD o�cers belonging to a racial/ethnic group in Jan-uary of each year. The right panel shows the corresponding proportion for incomingrecruits, aggregated by �ve-year windows (e.g., 1970–1974, 1975–1980) to reduce noisefrom relatively small recruitment cohorts. In both plots, decennial Census proportionsfor each racial/ethnic group for the city of Chicago, tabulated by the National HistoricalGeographic Information System, are provided for reference.
Composition of full department,by year
Composition of recruits,by 5-year window
1970 1980 1990 2000 2010 1970 1980 1990 2000 20100.00
0.25
0.50
0.75
Pro
port
ion
Race/Ethnicity Black Hispanic White
Group Chicago city population Chicago Police Department
13
Figure 3: Racial and ethnic composition of o�cers’ assigned districts. In each panel,points corresponding to police districts indicate the proportion of civilians of a givenracial/ethnic group residing in 2010 Census data, together with the corresponding 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
showing that the pattern of assignments di�ers dramatically by o�cer race. For example,white o�cers in Wentworth (District 2) almost exclusively serve from 4 p.m. to midnight,whereas black o�cers are more likely to be assigned to mid-day shifts. Similarly, SI Fig-ure A3 shows that relative to white o�cers, black o�cers are far more frequently deployedto assigned beat 202—roughly corresponding to a patrol area in the district’s southwestcorner (depicted in Figure 1) that has extremely high police activity and a high concen-tration of black residents. These results undermine analyses in a wide array of previousstudies that aggregate at high levels of geography (for example, controlling for districtor unit assignment) and which assume that o�cers face homogeneous conditions withinthese crude groupings.
These patterns not only establish that o�cers tend to police co-racial residents, butalso underscore a central di�culty in testing whether white and nonwhite o�cers per-form their duties di�erently. Namely, white and nonwhite o�cers work in di�erent envi-ronments on average, especially with regard to local racial composition. This means anyinferred di�erences in o�cer behavior that rely on data aggregated to large geographicunits, such as districts or agencies—the analytic strategy in most prior work—may simplybe artifacts of the divergent environments they police. Our data allow us to overcome thisobstacle to inference.
14
5 Do O�cers Live Where They Work?
Our data contain the home addresses of CPD employees for the year 2004, which allowsus to assess whether o�cers reside in or near the areas of the city where they work.8
Home location may serve as a rough proxy for all manner of attributes, such that o�cersliving in their assigned district may have more in common with the residents they police.While CPD o�cers are required to live within the city limits, Chicago’s extreme racial andeconomic segregation means that living in a di�erent part of town often implies a whollydi�erent lived experience. To the extent that o�cers reside in the same areas they patrol,some reformers have asserted they may have greater familiarity with residents, inspiregreater levels of trust, and exhibit greater levels of empathy toward the civilians theyserve (Goodyear, 2014; Gramlich, 2017). And by assessing the correspondence betweenthe racial composition of areas where o�cers live and work, we can gain insight intowhether o�cers are likely coming into contact with various civilian groups when o�duty, i.e. in a less adversarial context.
Figure 4 shows that much like city residents, Chicago Police o�cers are racially seg-regated. The left panel of the �gure depicts the residences of 13,635 o�cers for whomgeocoded 2004 addresses were available, perturbed randomly by 500 meters to maintainprivacy. The center panel shows only those o�cers assigned to work in their home dis-trict in 2004. The right panel shows the racial composition of each district for reference.As the �gure makes clear, o�cers tend to live among co-racial and co-ethnic civilians, apattern echoed in Table 2.
Table 2: Racial composition of o�cers’ home districts, averaged over o�cers.Civilian demographics of the districts in which o�cers live, averaged over o�cers in eachracial group. Boxes indicate co-ethnic/co-racial proportions.
Proportion Black Residents Proportion Hispanic Residents Proportion White ResidentsO�cer White 0.219 0.207 0.525O�cer Hispanic 0.207 0.324 0.420O�cer Black 0.630 0.153 0.199
8In addition, we analyze local demographics here at the district level rather than the beat level, becausethe geographic boundaries of beats change over time, sometimes in unobservable ways. However, this doesnot compromise our behavioral analysis. Within a given month, beat boundaries remain �xed, so we canbe sure o�cers compared within beats in a given month are patrolling to same geographic areas.
15
Figure 4: Police residences by o�cer race. The left panel shows the rough locations of all CPD o�cer residences in 2004,perturbed to maintain privacy. The center panel shows those o�cers whose homes are located in their assigned district. Theright panel shows the local racial composition of each district.
Officer residence locations
BlackHispanicWhite
Officers residing in assigned district District civilian racial composition
16
6 O�cers’ Political Preferences
Another contextual factor that may contribute to the e�ectiveness of diversi�cation isconcordance between the political preferences of o�cers and residents within racial andethnic groups. Democrats and Republicans hold markedly di�erent views on criminaljustice matters (Eckhouse, 2019; Federico and Holmes, 2005; Weaver, 2007; Weitzer andTuch, 2004). If nonwhite o�cers skew conservative, there may be little reason to expectthat racial diversi�cation would change the nature of policing. To investigate this, wemerge our o�cer data with a commercial voter database, following Enamorado, Fi�eldand Imai (2017), to obtain information on o�cers’ political activity and partisan leanings;details of the procedure are given in SI Section A.4.9
Figure 5 displays the share of each racial and ethnic group a�liating with the Demo-cratic and Republican parties—as well as the shares a�liating with neither party—by birthyear among both city residents and CPD o�cers. This tabulation, based on contempora-neous voter turnout data, allows us to assess the degree to which CPD o�cers’ currentpolitical preferences resemble those of Chicago residents, as well as the degree to whichthis correspondence varies across cohorts. Turning �rst to black o�cers, we see thatcompared to city residents, o�cers are slightly more likely to be Democrats, slightly lesslikely to be nonpartisan, and slightly more likely to be Republican. Though there is somevariation in the magnitude of these di�erences across cohorts, in the aggregate, 7% ofblack o�cers a�liate with the Republican party, compared to 1% of black city residents.But Hispanic and white o�cers diverge sharply from minority populations: while theshare of Democratic o�cers is comparable to the share of Democratic civilians, o�cersare substantially less (more) likely to be nonpartisan (Republican). In the aggregate, 22%of Hispanic o�cers and 40% of white o�cers a�liate with the Republican party, comparedwith 8% and 28% among co-racial and co-ethnic residents, respectively. In general, blacko�cers appear fairly similar to co-racial residents in terms of political preferences, whileHispanic and white o�cers appear markedly more conservative. To the extent politicalpreferences predict styles of policing, this pattern suggests that black civilians may expe-rience improved treatment by black o�cers relative to other groups of o�cers. In fact,we �nd precisely this pattern in our behavioral analysis below.
9Voters in Chicago are not required to register with a party to vote in party primaries, but can onlyparticipate in one party’s primary in each election. L2 codes party a�liation based on a proprietary process,and we use their coding here. As a robustness check, we also code party ID based on which party primaryvoters participated in during the 2016 cycle and obtain highly similar results; see SI Figure A5.
17
Figure 5: Police more likely to be Republican and to vote than co-Racial and co-ethnic civilians. Cohort political participation and composition of CPD o�cers andChicago residents based on current political orientation. O�cers are aggregated to �ve-year windows on birth year (1950–1954, 1955–1960, etc.). For turnout in previous electionsfrom 2000–2016, see SI Section C.3 and SI Figure A4.
Voted in 2016 Democratic Non-Partisan Republican
Bla
ckH
ispanic
White
1950
1960
1970
1980
1990
1950
1960
1970
1980
1990
1950
1960
1970
1980
1990
1950
1960
1970
1980
1990
0.00.20.40.60.8
0.00.20.40.60.8
0.00
0.25
0.50
0.75
Birth year
Pro
port
ion
Race/Ethnicity Black Hispanic White
Group Chicago city population Chicago Police Department
18
In addition, Figure 5 shows police across nearly all cohorts are far more likely to votethan city residents. On average, white, Hispanic, and black citizens’ (o�cers’) votingrates in 2016 were 74% (84%), 61% (68%) and 60% (77%), respectively. Given longstandingcorrelations between turnout and strength of partisanship (Campbell et al., 1960), thisresult suggests that not only do many o�cers a�liate with the Republican party at higherrates, they are also likely to be stronger partisans across the board.
7 Racial Diversity and Police Behavior
We now turn to assessing whether racial diversity among police o�cers a�ects the natureand volume of police-civilian interactions. Figures 6-8 display the average di�erences inthe number of stops, arrests and uses of force associated with black and Hispanic o�cersrelative to white o�cers working in the same MDSBs. Turning �rst to black o�cers, wesee that when faced with the same conditions, black o�cers make roughly 0.16 fewerstops, make 0.020 fewer arrests, and use force 0.0011 fewer times per shift, on average,than white o�cers facing the same circumstances—that is, counterparts assigned to thesame month, the same part of week, the same shift time, and the same beat. These gapsrespectively represent reductions of 31%, 22%, and 41% over typical stop, arrest, and use-of-force volume for white o�cers (see SI Table A1 for detailed tables of average behaviorby o�cer race).
Importantly, Figure 7 also shows that these disparities are not uniform across situa-tions. Rather, they are driven by a reduced focus on engaging black civilians (41% fewerstops, 26% fewer arrests) and a broad class of enforcement activities that can be thoughtof as relatively discretionary, including stops of civilians for “suspicious behavior” (33%less) or arrests for drug-related crimes (30% less). However, when it comes to enforcingviolent crime, black o�cer behavior looks similar to that of white o�cers, with black o�-cers making only 10% fewer arrests for violent crimes than whites. In other words, whenit comes to policing the most serious o�enses, black o�cers are comparable to white of-�cers. Black o�cers also deploy force 35% less often than their white counterparts. Thisdi�erence is driven largely by reduced uses of force against black civilians: decreased useof force against this group accounts for 83% of the overall disparity between white andblack o�cers. (SI Section C.6 and SI Tables A2–A6 report detailed analyses for the resultspresented here; SI Section C.7 and SI Figures A6–8 show that results are virtually identicalusing a wide range of alternative estimators.) This pattern of results is remarkably in line
19
with the hopes of proponents of racial diversi�cation, who have asserted the policy couldreduce abusive policing and mass incarceration in black communities. Our results implythat if the CPD were to hire additional similarly situated black o�cers from the same poolof recruits from which current o�cers are drawn, and deployed them in the same manneras current o�cers, then black civilians would be subject to signi�cantly lower levels ofdetainment for petty crime and police violence.
Results di�er in important ways for Hispanic o�cers. As with black o�cers, Hispanico�cers facing the same working conditions exhibit fewer stops, arrests and uses of forcethan white o�cers, but the magnitudes of these di�erences are far more modest. Strik-ingly, these negative di�erences are driven by less engagement with black civilians, whileHispanic o�cers exhibit nearly the same volume of enforcement activity against Hispaniccivilians as white o�cers on average. These estimates imply that Hispanic o�cers can beexpected to produce 0.032 fewer stops (6% reduction relative to white o�cers) and 0.00034fewer uses of force (11% reduction) per shift.
In addition to stops, arrests and uses of force, Figure 9 displays average di�erences incomplaints �led by civilians against di�erent groups of o�cers. These results di�er some-what from our analyses of the previous three behaviors. Despite engaging civilians at alower rate, black o�cers receive roughly the same number of complaints as white o�cersworking in the same places and under the same conditions, with two notable exceptions:black o�cers are signi�cantly less likely to receive a complaint related to improper searchor arrest of a civilian (0.22 fewer complaints per 1,000 shifts, a 36% reduction), and theyare estimated to receive a large but statistically insigni�cant number of additional minorprocedural complaints (e.g. relating to improper �ling of paperwork or failure to provideservice). Hispanic o�cers, on the other hand, receive 0.16 fewer complaints per 1,000shifts relative to comparable white o�cers. This gap is primarily driven by a decrease incomplaints (1) from black civilians (0.10 fewer complaints per 1,000 shifts, a 19% reduc-tion) and (2) relating to improper use of force (0.09 fewer complaints per 1,000 shifts, areduction of 21%).
We caution that complaint records are incomplete, in part due to extensive missingnessand procedural hurdles that impede their �ling (Ba, 2019). For instance, of complaints�led between January 2011 and July 2014, investigations were closed for 38% because theinvestigator was not able to link the allegation to a speci�c accused o�cer. In addition,complaints may incorporate elements of civilian perception and bias (e.g., �ling of petty orretaliatory complaints). We note that our analysis of complaints necessarily di�ers from
20
Figure 6: Stops. Points (error bars) depict estimated di�erences (95% block-bootstrap con�dence intervals) in the number ofcivilian stops conducted by black and Hispanic o�cers per shift, relative to white o�cers patrolling in the same month, daygroup (weekday/weekend), shift time (�rst/second/third watch), and beat. Corresponding regression tables are reported in SISection C.6, and results from alternative estimators are shown in SI Section C.7.
Black officers (vs white) Hispanic officers (vs white)
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
-0.15
-0.10
-0.05
0.00
Diffe
rence
s in
sto
ps
per
shift
by
off
icers
of
each
raci
al/eth
nic
gro
up,
vers
us
whit
e o
ffic
ers
faci
ng C
C
21
Figure 7: Arrests. Points (error bars) depict estimated di�erences (95% block-bootstrap con�dence intervals) in the number ofarrests conducted by black and Hispanic o�cers per shift, relative to white o�cers patrolling in the same month, day group(weekday/weekend), shift time (�rst/second/third watch), and beat. Corresponding regression tables are reported in SI SectionC.6, and results from alternative estimators are shown in SI Section C.7. Results are qualitatively similar when holding o�cergender constant; see SI Figures A13–A14 for details.
Black officers (vs white) Hispanic officers (vs white)
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
-0.25
-0.20
-0.15
-0.10
-0.05
0.00
Diffe
rence
s in
arr
est
s per
10
shifts
by
off
icers
of
each
raci
al/eth
nic
gro
up,
vers
us
whit
e o
ffic
ers
faci
ng C
C
22
Figure 8: Uses of Force. Points (error bars) depict estimated di�erences (95% block-bootstrap con�dence intervals) in thenumber of uses of force by black and Hispanic o�cers per shift, relative to white o�cers patrolling in the same month, daygroup (weekday/weekend), shift time (�rst/second/third watch), and beat. Corresponding regression tables are reported in SISection C.6, and results from alternative estimators are shown in SI Section C.7.
Black officers (vs white) Hispanic officers (vs white)
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
-0.15
-0.10
-0.05
0.00
Diffe
rence
s in
use
s of
forc
e p
er
10
0 s
hifts
by o
ffic
ers
of
each
raci
al/eth
nic
gro
up,
vers
us
whit
e o
ffic
ers
faci
ng C
C
23
Figure 9: Complaints. Points (error bars) depict estimated di�erences (95% block-bootstrap con�dence intervals) in the numberof civilian complaints registered against black and Hispanic o�cers per shift, relative to white o�cers patrolling in the samemonth, day group (weekday/weekend), shift time (�rst/second/third watch), and beat. Corresponding regression tables arereported in SI Section C.6, and results from alternative estimators are shown in SI Section C.7.
Black officers (vs white) Hispanic officers (vs white)
Civ. r
ace:
Bla
ck
Civ. r
ace:
Hisp
anic
Civ. r
ace:
Whi
teRea
son:
Oth
er
Reaso
n: A
rrest
/sea
rch
Reaso
n: F
orce
Reaso
n: U
nkno
wn
Tota
lCiv
. rac
e: B
lack
Civ. r
ace:
Hisp
anic
Civ. r
ace:
Whi
teRea
son:
Oth
er
Reaso
n: A
rrest
/sea
rch
Reaso
n: F
orce
Reaso
n: U
nkno
wn
Tota
l
-0.4
-0.2
0.0
0.2
Diffe
rence
s in
com
pla
ints
per
1k
shifts
by o
ffic
ers
of
each
raci
al/eth
nic
gro
up,
vers
us
whit
e o
ffic
ers
faci
ng C
C
24
our analysis of the previous three behaviors due to these data constraints. In particular, theexact time of the incident which generated a complaint is often unknown. By necessity,if an incident is reported on the date of the accused o�cer’s shift, we assume the incidentoccurred while the o�cer was on duty. (If complaint-generating incidents occur whileo�cers are o�-duty, after departing their assigned beat, the CC assumption is likely to beviolated.) Given these limitations, we advise readers to treat our conclusions regardingcivilian complaints as tentative. However, we include these data in our analysis becausethey o�er a unique civilian perspective on the appropriateness of police behavior thatis di�cult to obtain from o�cer-generated reports, and suggest several patterns whichfuture research should further examine.
8 O�cers’ Spanish Language Pro�ciency
Our initial analysis indicated that white and Hispanic o�cers behaved comparably towardHispanic civilians, suggesting hiring and deploying additional similarly situated Hispanico�cers would not alter the treatment of Hispanic civilians by police, on average. However,this analysis does not account for the role of language in police-civilian interactions. Ac-cording to a Pew Research Center study, roughly 77% of Hispanic residents in the Chicagometropolitan area speak Spanish at home as of 2015 (Krogstad and Lopez, 2017). However,personnel data we obtained through open records requests show less than half of HispanicCPD o�cers can speak Spanish. This could preclude e�ective communication with im-migrant populations. Moreover, Hispanic o�cers who cannot speak Spanish may divergefrom many Hispanic residents in their cultural backgrounds and lived experiences. If His-panic o�cers cannot speak the native language of the co-ethnic civilians they serve, thee�ects of co-ethnic policing may not materialize.
To investigate this possibility, we use the same analytic approach as in the previoussection to compare the behavior of Hispanic o�cers who can and cannot speak Spanish ontheir volume of stops, arrests, uses of force and civilian complaints. These comparisons,displayed in Figure 10, reveal important heterogeneity in o�cer behavior within this var-ied ethnic group.10 Relative to Hispanic o�cers who speak Spanish and are working in thesame places and times, Hispanic o�cers who cannot speak Spanish exhibit substantiallyhigher numbers of stops, arrests and uses of force. For example, non-Spanish-speakingHispanic o�cers make roughly 9% more stops and 13% more arrests of Hispanic civilians,
10SI Figure A10 shows similar di�erences between Spanish-speaking Hispanic o�cers and white o�cers.
25
relative to Spanish-speaking Hispanic o�cers facing a comparable pool of civilian ac-tivity. Perhaps counterintuitively, non-Spanish-speaking o�cers stop signi�cantly moreHispanic civilians than even white o�cers. However, like black o�cers, Spanish-speakingHispanic o�cers receive more minor complaints (those falling into the “other” category).
26
Figure 10: Spanish Language Ability. Points (error bars) depict estimated di�erences (95% block-bootstrap con�dence inter-vals) in the number of civilian stops, arrests, uses of force, and civilian complaints generated by non-Spanish-speaking Hispanic(NSSH) o�cers per shift, relative to Spanish-speaking Hispanic (SSH) facing common circumstances (CC) o�cers patrolling inthe same month, day group (weekday/weekend), shift time (�rst/second/third watch), and beat.
Force uses per 100 shifts Complaints per 1,000 shifts
Stops per shift Arrests per 10 shifts
Civ. r
ace:
Bla
ckCiv
. rac
e: H
ispan
icCiv
. rac
e: W
hite
Tota
l
Civ. r
ace:
Bla
ck
Civ. r
ace:
Hisp
anic
Civ. r
ace:
Whi
te
Reaso
n: O
ther
Reaso
n: A
rrest
/sea
rch
Reaso
n: F
orce
Reaso
n: U
nkno
wnTo
tal
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
s
Reaso
n: D
rug
Reaso
n: T
raffi
cTo
tal
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
ntTo
tal
0.00
0.05
0.10
0.15
0.20
-0.4
-0.2
0.0
0.2
0.00
0.03
0.06
0.09
0.00
0.02
0.04
0.06
0.08
Diffe
rence
s in
behavio
rs f
or
NS
SH
off
icers
, vers
us
SS
H o
ffic
ers
faci
ng
CC
27
9 Discussion and Conclusion
A string of fatal encounters between white police o�cers and unarmed civilians of colorhas led to renewed calls for law enforcement reforms. Prominent among these propos-als is the racial diversi�cation of police agencies, but the e�ects of this widely studiedreform have remained uncertain due to severe data limitations. Using an unusually richindividual-level data set on police activity in Chicago, we provide the most credible as-sessment to date of the impact of this reform on police-civilian interactions.
Our results reveal a much more nuanced story than previous accounts—which arelargely based on agency-level summary statistics (e.g. McCluskey and McCluskey, 2004;Smith, 2003)—have been able to show. Using demographic and political data on o�cers,we �rst show that o�cers are geographically assigned in ways that increase interactionswith co-racial and co-ethnic residents. Though the majority of o�cers do not live in thepolice districts where they are assigned to work, they do tend to live among co-racialand co-ethnic residents, suggesting lower levels of inter-group contact when o� duty.In addition, while the political preferences of black o�cers are fairly comparable to co-racial city residents, white and Hispanic o�cers are markedly more conservative thantheir civilian counterparts—a noteworthy pattern given the correlation between politicalconservatism and support for punitive criminal justice policies.
After establishing these descriptive portraits of police o�cers, we report a range ofanalyses characterizing police behavior and how racial diversity translates into meaning-ful but sometimes unexpected di�erences in the lives of the policed. Faced with the sameworking conditions, black o�cers are less likely to stop, arrest, and use force against civil-ians, especially black civilians, relative to white o�cers. These disparities are driven by areduced focus on the enforcement of discretionary stops and arrests for petty crimes, in-cluding drug o�enses, which have long been thought to fuel mass incarceration (Alexan-der, 2010). In contrast to these drastic di�erences in the policing of petty crime, blacko�cers’ enforcement of violent crime is only slightly lower than that of white o�cers.But this pattern, which closely comports with the hopes of advocates of racial diversi�ca-tion, is complicated by several additional results. For one, while Hispanic o�cers displaylower levels of enforcement activity than whites overall, their behavior toward Hispaniccivilians is broadly comparable to that of white o�cers. And after accounting for lan-guage ability, we �nd non-Spanish speaking Hispanic o�cers engage Hispanic civiliansfar more often than their Spanish-speaking counterparts.
28
Our analysis of civilian-generated complaints provides suggestive evidence of the roleof civilian behavior. While black o�cers engage in lower levels of enforcement activitythan white o�cers, they receive a roughly equal number of civilian complaints. Thispattern could be explained by several mechanisms. On the one hand, it may indicate thatblack o�cers engage in misconduct at higher rates, though we �nd this unlikely giventhat most of the complaint disparity stems from allegations of minor infractions. Anotherexplanation is that civilians judge the behavior of black o�cers more harshly than thatof white o�cers, or that civilians believe complaints against white o�cers are less likelyto be taken seriously and therefore opt not to �le them in a variety of situations. Ourability to parse these mechanisms with current data is limited, and we encourage futureresearch to investigate these patterns further. Future work should also seek to discernwhether the patterns we reveal stem from di�erential levels of proactive behavior acrossracial and ethnic groups of o�cers, as opposed to di�erential reactions by civilians towardo�cers of various groups. Doing so would likely require close, unobtrusive observation ofpolice-civilian encounters, perhaps via the analysis of dash or body-worn camera footage.
Finally, while our analysis focuses on evaluating policies related to race and ethnic-ity, other o�cer attributes—most notably, o�cer gender (Lockwood and Prohaska, 2015;Raganella and White, 2004)—that deserve renewed and extended analysis. Our data andresearch design o�er an immediate path for researchers to explore the impact of this andother personnel related reforms. Taken together, our results show that the e�ects of racialdiversi�cation are neither simple nor monolithic. Like the civilians they serve, o�cersare multidimensional. Crafting e�ective personnel reforms requires thinking beyond thecoarse racial and ethnic categories typically used in diversity initiatives, and considerationof how multiple o�cer attributes relate police to the civilians they serve.
29
References
Alexander, Michelle. 2010. The New Jim Crow: Mass Incarceration in the Age of Colorblind-ness. New York: The New Press.
Antonovics, Kate and Brian G. Knight. 2009. “A new look at racial pro�ling: Evidencefrom the Boston Police Department.” The Review of Economics and Statistics 91(1).
Ba, Bocar A. 2019. “Going the Extra Mile: the Cost of Complaint Filing, Accountability,and Law Enforcement Outcomes in Chicago.” Working paper .
Campbell, Angus, Philip E. Converse, Warren E. Miller and Donald E. Stokes. 1960. TheAmerican Voter. University of Chicago Press.
Cao, Liqun, James Frank and Francis T. Cullen. 1996. “Race, community context andcon�dence in the police.” American Journal of Police 15(1):3–22.
Clinton, Joshua D. and David E. Lewis. 2008. “Expert opinion, agency characteristics, andagency preferences.” Political Analysis 16(1):3–20.
DOJ/EEOC. 2016. “Advancing Diversity in Law Enforcement.”. https://www.justice.
gov/crt/case-document/file/900761/download.
Donohue III, John J. and Steven D. Levitt. 2001. “The impact of race on policing andarrests.” The Journal of Law and Economics 44(2):367–394.
Eckhouse, Laurel. 2019. “Race, Party, and Representation in Criminal Justice Politics.” TheJournal of Politics 81(3).
Enamorado, Ted, Benjamin Fi�eld and Kosuke Imai. 2017. “Using a Probabilistic Model toAssist Merging of Large-scale Administrative Records.” Working Paper .
Federico, Christopher M. and Justin W. Holmes. 2005. “Education and the interface be-tween racial perceptions and criminal justice attitudes.” Political Psychology 26:47–75.
Forman Jr., James. 2017. Locking up our own: Crime and punishment in Black America.Farrar, Straus and Giroux.
Garner, Maryah, Anna Harvey and Hunter Johnson. 2019. “Estimating E�ects of A�rma-tive Action in Policing: A Replication and Extension.” International Review of Law andEconomics p. 105881.
30
Gelman, Andrew, Je�rey Fagan and Alex Kiss. 2007. “An Analysis of the New York CityPolice Department’s "Stop-and-Frisk’" Policy in the Context of Claims of Racial Bias.”Journal of the American Statistical Association 102(429):813–823.
Goodyear, Sarah. 2014. “Should Cops Have to Live Where TheyWork?” CityLab . https://www.citylab.com/equity/2014/08/
should-cops-have-to-live-where-they-work/378858/.
Gramlich, John. 2017. “Black and white o�cers see many key aspects of polic-ing di�erently.”. https://www.pewresearch.org/fact-tank/2017/01/12/
black-and-white-officers-see-many-key-aspects-of-policing-differently/2017.
Greiner, James D. and Donald B. Rubin. 2011. “Causal e�ects of perceived immutablecharacteristics.” Review of Economics and Statistics 93:775–785.
Harvey, Anna and Taylor Mattia. 2019. “Reducing Racial Disparities in Crime Vic-timization.” Working Paper . https://s18798.pcdn.co/annaharvey/wp-content/
uploads/sites/6417/2019/12/Victimization_Harvey_Mattia.pdf.
Holland, Paul W. 1986. “Statistics and causal inference.” Journal of the American StatisticalAssociation 81(396):945–60.
Johnson, David J., Trevor Tress, Nicole Burkel, Carley Taylor and Joseph Cesario. 2019.“O�cer characteristics and racial disparities in fatal o�cer-involved shootings.” Pro-ceedings of the National Academy of Sciences . Published online ahead of print July 22,2019. https://www.pnas.org/content/early/2019/07/16/1903856116.
Knox, Dean and Jonathan Mummolo. 2020. “Making Inferences About Racial Disparitiesin Police Violence.” Proceedings of the National Academy of Sciences 117:1261–1262.
Knox, Dean, Will Lowe and Jonathan Mummolo. Forthcoming. “Administrative RecordsMask Racially Biased Policing.” American Political Science Review . https://papers.
ssrn.com/sol3/papers.cfm?abstract_id=33363387.
Krogstad, Jens Manuel and Mark Hugo Lopez. 2017. “Use of Span-ish declines among Latinos in major U.S. metros.” Pew ResearchCenter . https://www.pewresearch.org/fact-tank/2017/10/31/
use-of-spanish-declines-among-latinos-in-major-u-s-metros/.
31
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.
Lerman, Amy and Vesla Weaver. 2014. Arresting Citizenship: The Democratic Consequencesof American Crime Control. University of Chicago Press.
Levitt, Steven D. 2012. “Understanding why crime fell in the 1990s: Four factors thatexplain the decline and six that do not.” The Journal of Economic Perspectives 18(1):163–190.
Lewis, David E. 2010. “Measurement and public service motivation: New insights, oldquestions.” International Public Management Journal 13(1):46–55.
Linos, Elizabeth and Nefara Riesch. 2019. “Thick Red Tape and the Thin Blue Line: A FieldStudy on Reducing Administrative Burden in Police Recruitment.”.
Lockwood, Daniel and Ariane Prohaska. 2015. “Police O�cer Gender and Attitudes to-ward Intimate Partner Violence: How policy can eliminate stereotypes.” InternationalJournal of Criminal Justice Studies 10(1):77–90.
McCluskey, Cynthia Perez and John D. McCluskey. 2004. “Diversity in polic-ing: Latino representation in law enforcement.” Journal of Ethnicity in Crimi-nal Justice 2:67–81. https://www.pewresearch.org/fact-tank/2017/01/12/
black-and-white-officers-see-many-key-aspects-of-policing-differently/2017.
McCrary, Justin. 2007. “The e�ect of court-ordered hiring quotas on the composition andquality of police.” American Economic Review 97(1):318–353.
Raganella, Anthony J. and Michael D. White. 2004. “Race, gender, and motivation forbecoming a police o�cer: Implications for building a representative police department.”Journal of criminal justice 32(6):501–513.
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.
Rim, Nayoung, Roman G. Rivera and Bocar A. Ba. 2019. “In-group Bias and the Police:Evidence from Award Nominations.” Working paper .
32
Rivera, Roman G. and Bocar A. Ba. 2019. “The E�ect of Police Oversight on Crime andAllegations of Misconduct: Evidence from Chicago.” Working paper .
Rubin, Donald B. 1974. “Estimating causal e�ects of treatments in randomized and non-randomized studies.” Journal of Educational Psychology 66:688–701.
Rubin, Donald B. 1990. “Formal mode of statistical inference for causal e�ects.” Journal ofstatistical planning and inference 25(3):279–292.
Sen, Maya and Omar Wasow. 2016. “Race as a bundle of sticks: Designs that estimate ef-fects of seemingly immutable characteristics.” Annual Review of Political Science 19:499–522.
Shusta, Robert M., Deena R. Levine Phillip R. Harris and Herbert Z. Wong. 1995. Mul-ticultural law enforcement: Strategies for Peacekeeping in a Diverse Society. 9 ed. YaleUniversity Press.
Sklansky, David Alan. 2005. “Not your father’s police department: Making sense of thenew demographics of law enforcement.” J. Crim. L. & Criminology 96:1209–1244.
Smith, Brad. 2003. “The Impact of Police O�cer Diversity on Police-Caused Homicides.”Policy Studies 31(2):147–162.
Turner, K. B., David Giacopassi and Margaret Vandiver. 2006. “Ignoring the past: Cov-erage of slavery and slave patrols in criminal justice texts.” Journal of Criminal JusticeEducation 17(1):181–195.
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.
Weaver, Vesla M. 2007. “Frontlash: Race and the development of punitive crime policy.”Studies in American political development 21:230–265.
Weitzer, Ronald and Steven A. Tuch. 2004. “Reforming the police: Racial di�erences inpublic support for change.” Criminology 42(2):391–416.
Wilson, James Q., ed. 1989. Bureaucracy: What government agencies do and why they do it.
Wilson, James Q. and George L. Kelling. 1982. “Broken windows: the police and neigh-borhood safety.” Atlantic Monthly 249(3):29–38.
33
Supplemental Materials
Table of ContentsA Detailed Description of Data 2
A.1 CPD Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
A.2 Coding race . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
A.3 U.S. Census Merge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
A.4 Voter File Merge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
A.5 Preprocessing of patrol assignments . . . . . . . . . . . . . . . . . . . . 4
A.6 Preprocessing of police behavior . . . . . . . . . . . . . . . . . . . . . . 5
B Causal Estimand in Behavioral Analysis 5
C Additional Results 7
C.1 Race and District Assignment . . . . . . . . . . . . . . . . . . . . . . . 7
C.2 O�cer Race and Patrol Assignments . . . . . . . . . . . . . . . . . . . . 9
C.3 O�cers’ Political Preferences . . . . . . . . . . . . . . . . . . . . . . . . 12
C.4 O�cer Behavior: Overall Means . . . . . . . . . . . . . . . . . . . . . . 15
C.5 O�cer Behavior: Shift Duration . . . . . . . . . . . . . . . . . . . . . . 17
C.6 O�cer Behavior: Regression Tables . . . . . . . . . . . . . . . . . . . . 18
C.7 O�cer Behavior: Alternative Estimators . . . . . . . . . . . . . . . . . 23
C.8 O�cers’ Spanish Language Pro�ciency . . . . . . . . . . . . . . . . . . 29
C.9 O�cer Gender . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
C.10 Robustness Checks: Multiple Stopping O�cers . . . . . . . . . . . . . . 36
1
A Detailed Description of Data
A.1 CPD Data
The administrative data from CPD used in this study span multiple data sets collectedin collaboration with the Invisible Institute, Sam Stecklow, and Emma Herman over thecourse of three years (2016-2019). We obtained these records from the Chicago Police De-partment or Chicago Department of Human Resources via Freedom of Information Act(FOIA) or through court ordered releases stemming from requests made by Invisible In-stitute and Jaime Kalven. CPD provided the following data: rosters of all available currentand past o�cers up to 2018, data on o�cers allegation of misconduct from 2000 to 2016(complaints), unit movement data for individual o�cers from the 1930s to 2016, Tacti-cal Response Reports from 2004 to 2018 (i.e. use of force reports), and arrest data witharresting o�cers and arrestee demographic information from 2001 to 2017. The ChicagoDepartment of Human Resources provided data on o�cers’ language skills up to 2019 ando�cers’ home address in 2004, 2005 and in early 2019. We supplement our core data withdata on “Stop, Question and Frisk” (SQF) activity between 2012-2015, which was publicyshared by the Lucy Parson’s Lab. Finally, the Automated Daily Attendance and Assign-ment sheet data for each police district between 2012 and 2015 was obtained via a FOIArequest to the CPD and 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 arecommon, etc.), we constructed unique o�cer pro�les through a successive merge processdescribed here. Each �le contains some identifying information such as of demographicdata (birth year, race, gender) or other characteristics (name, start/badge number, ap-pointed date, resignation date, current unit). We used these identifying characteristics to�rst de-duplicate o�cers within a �le and to then merge to pre-existing o�cer data withinter-�le unique identi�ers. The merging process itself is an iterative-pairwise matchingmethod, where the o�cers in each data set are repeatedly merged on identifying charac-teristics and any successful 1-to-1 match in a round removes the matched o�cers fromthe next round of merging.
The resulting data contains records on 33,000 police o�cers appointed between Marchof 1936 to February of 2018. The number of years and o�cers varies across analyses inour paper due to missing data (for example, assignment data only exists for the years2012–2015).
2
A.2 Coding race
We determine race/ethnicity of CPD o�cers based on demographic data obtained fromthe CPD through FOIA. The CPD classi�es race/ethnicity in 6 mutually exclusive groups:white/Caucasian, Hispanic, Black/African American, Asian/Paci�c Islander, Native Amer-ican/Native Alaskan, and unknown/missing. If an o�cer has multiple races associatedwith them across di�erent data sets, we aggregate by most common non-missing races.
For Census and American Community Survey data, we construct corresponding racecategories as follows: any Hispanic individual is coded Hispanic; white and black arecomprised of individuals 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 andthe CPD’s pre-2012 beat map. The centroid of each census tract was identi�ed, then thedemographic information of all the centroids inside a beat were aggregated to determinethe beat’s population and demographic composition. District demographics were deter-mined by aggregating across all beats within that district. Post-2012 district and beatdemographics were constructed based on the pre-2012 beat data discussed previously andusing a crosswalk that maps pre-2012 beats to current (2018) beats and their respectivedistricts.
A.4 Voter File Merge
We merged CPD employee data with a nationwide commercial voter �le, maintained byL2 Inc., using the Fellegi-Sunter model as implemented in fastLink (Enamorado, Fi�eldand Imai, 2017). The merge was conducted using �rst, middle, and last name, along withyear of birth. Because a direct merge of the CPD roster with a voter �le of more than180 million observations would result in over 6 trillion pairwise comparisons, we employadditional restrictions for computational feasibility. We focus our attention on voter �leobservations that have a postal code in the greater Chicago area, reducing the number ofpairwise comparisons to slightly over 100 million. For string-valued �rst and last name,we evaluate the quality of a match by Jaro-Winkler similarity, a common string similaritymeasure widely used for name comparison that employs a character-wise comparison oftwo strings while placing more emphasis in their �rst four characters. This similarity is
3
discretized into three levels (di�erent, similar, and identical/almost identical) by thresh-olding at 0.85 and 0.94. For year of birth, the absolute value of the di�erence is thresholdedat 1 year (i.e., agreement in year of birth is declared for observations that are at most oneyear apart). For middle names, we evaluate whether candidate matches have an identi-cal value. Based on these discrete comparisons, the probabilistic model behind fastLink
assigns a probability of being a match for each pair of records. The intuition is simple:the more that a pair of records agree in the linkage �elds, the more likely that pair ofrecords is a match. In total, 21,276 out of 33,645 police o�cers were matched, for a matchrate just above the 63% mark; only pairs with a �tted match probability greater than 0.9are retained. We then use the resulting mapping to obtain L2’s proprietary predictionsof each matched o�cer’s current political orientation (Republican, Democratic, and non-partisan, with an extremely small number of third-party labels that are not depicted inanalysis), along with public records of general election and party primary turnout thatare aggregated and maintained by L2.
A.5 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 dueto small sample sizes. Within this subset, we further drop non-standard assignments (no-tably including “station supervisor” and “station security” assignments, as well as specialassignments for training, compensatory time, and excused sick leave). Patrol assignmentsin which o�cers are indicated as non-present are also dropped. These steps are intendedto ensure that o�cers nominally patrolling a beat are in fact actively circulating in theassigned geographic area, improving the plausibility of the CC assumption. For the samereason, we drop double shifts (patrol assignment slots in which the assigned o�cer servedfor more than one shift on the same day) to address the possibility that o�cers behavedi�erently due to fatigue in these circumstances. We also eliminate o�cers assigned tonon-standard watches (i.e., other than �rst through third watches). Finally, we drop o�-cers at ranks other than “police o�cer.” This step eliminates police sergeants, who servein 8% of beat assignments but make very few stops and arrests, as well as legal o�cers,helicopter pilots, explosives technicians, and canine handlers.
4
A.6 Preprocessing of police behavior
Events are merged to the remaining patrol assignments based on o�cer ID and date. Thisstep discards a large number of events, including those involving o�cers of higher ranksand incidents occurring on rest days. For stops, arrests, and uses of force, we drop allevents that occur outside of the reported patrol start/stop times. A key di�erence in thehandling of complaints, where retrospectively reported incident times are of extremelylow quality, is that we do not implement this step (i.e., all complaints about incidents onthe date of a recorded patrol assignment are assumed to be occur while the o�cer is onduty, an assumption that may potentially be violated).
Stops for “dispersal” and “gang and narcotics-related loitering” are coded as loiteringstops; those that are “gang / narcotics related” are coded as drug stops; “investigatorystops” and stops of “suspicious persons” are coded as suspicious behavior; and stops un-der the “Repeat O�ender Geographic Urban Enforcement Strategy (ROGUES)” programare combined with the “other” category. For stops, if a single o�cer is reported as bothprimary and secondary stopping o�cer, only one event is retained.
Arrests for municipal code violations and outstanding warrants are categorized as“other.”
Complaints initiated by other o�cers, which often relate to truancy or job perfor-mance rather than treatment of civilians, are discarded. Complaints can have multiplereported victims and complainants. When this is the case, each complaint is treated as asingle event, and all victims and complainants are aggregated. When multiple victims arereported, if any victims are black, the civilian race is coded as black. Otherwise, if anyHispanic victims are reported, civilian race is coded as Hispanic. Otherwise, if any whitevictims are reported, civilian race is coded as white. If and only if victim demograph-ics are unavailable, we fall back on complainant demographics. If multiple complainantsare reported, if any are black, the civilian race is coded as black. Otherwise, if any His-panic complainants are reported, civilian race is coded as Hispanic. Finally, if any whitecomplainants are reported, civilian race is coded as white.
B Causal Estimand in Behavioral Analysis
At a high level, the goal of our analysis is to evaluate the policy e�ect of a personnelreform that increases the representation of minorities in the CPD by assigning them to
5
positions that would otherwise be �lled by white individuals. The analysis is conductedat the level of the patrol assignment slot. Commanding o�cers are assumed to have a�xed set of patrol assignments that must be �lled, where each slot is associated with abeat (geographic patrol area) and shift time (temporal window). Multiple slots may beavailable for a particular beat and shift time, but each slot can be �lled by only one o�cer.
We organize beat assignments into groups, indexed by i, based on unique combi-nations of month (Mi), day group (weekday/weekend, Di), shift time (�rst/second/thirdwatch, Si), and beat (Bi), or unique MDSBs. Patrol assignment slots within a MDSB are in-dexed by j. For each slot, the realized pattern of o�cer behavior is denoted Yi,j(Ri,j), whereRi,j is the race of the o�cer assigned to a particular slot. Our notation implicitly makesthe stable unit treatment value assumption (SUTVA, Rubin, 1990), which requires that (1)there do not exist �ner gradations of o�cer identity (i.e., within the broad racial/ethniccategories used) that would result in di�ering potential o�cer behavior, and (2) that po-tential outcomes do not very depending on the racial/ethnic identity of o�cers assignedto other slots.1
The slot-level policy e�ect is the di�erence in potential outcomes (Rubin, 1974) Yi,j(r)−Yi,j(r ′), the change in behavior that would have realized if an o�cer of race r had beenassigned to the patrol assignment slot, rather than another o�cer of race r ′. These coun-terfactual di�erences are fundamentally unobservable; instead, we target the average pol-icy e�ect within the subset of F MDSBs for which policy e�ects can be feasibly estimated(i.e., for which variation in o�cer race 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 isthe average slot count across MDSBs. This can be rewritten as the weighted average of
1We explore the validity of this second assumption to the extent possible in SI Section C.10, in whichstops made by two o�cers are reanalyzed. In this section, we re-compute our estimates of di�erentialstopping behavior after excluding the second reporting o�cer from our analysis; the resulting estimatesare highly similar.
6
MDSB-speci�c e�ects, �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 2, 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 withinMDSBs, at least in ways that matter for potential o�cer behavior. (Hypothetically speak-ing, this independence condition could be achieved even without adjusting for MDSB ifwhite and nonwhite o�cers were randomly assigned locations and times to patrol.)
Our primary results estimate this quantity with an ordinary least squares (OLS) re-gression of the form Yi,j = �i +∑r �r1(Rij = r), where �i represents a �xed e�ect for MDSBi. Unbiasedness of this estimator requires the additional assumption that MDSB-speci�cpolicy e�ects are homogeneous, or that �i = � for all i. It is well known that whenthis assumption is violated, OLS recovers the weighted average of �is with weights cor-responding to the variance of o�cer race within strata. To allow for the possibility ofnon-homogeneous policy e�ects and other departures from our modeling assumptions,we therefore apply a number of alternative estimators, which are described in detail inSI Section C.7. As we show in SI Figures A6–A9, these alternative results are virtuallyidentical to our primary results.
C Additional Results
C.1 Race and District Assignment
The CPD currently subdivides Chicago into 22 policing districts which correspond to CPDunits, in which the majority of police o�cers work. There were 25 districts (numbered1 - 25) until 2012, at which time 3 smaller districts (ranking 18th, 21st, and 25th in landarea2) were eliminated and merged with other districts. Districts 23 and 21 and District 13
2See Chicago Police Department 2008 Annual Report, page 37, https://home.chicagopolice.org/wp-content/uploads/2014/12/2008-Annual-Report.pdf
7
Figure A1: Racial composition of o�cers’ assigned districts. Each panel depicts ageographic police district. Points represent the racial composition of district residents.Lines represent monthly proportions of o�cers assigned to a district that belong to eachracial group. Districts 21, 23, and 13 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
were eliminated and absorbed into neighboring districts in March and December of 2012,respectively. While District 23 was mostly absorbed by District 19 and most of District 13was absorbed by District 12, signi�cant parts of District 21 were absorbed by Districts 1,2, and 9.
Figure A1 displays signi�cant heterogeneity across districts in the racial compositionof their o�cers. There is a correlation between the racial composition of a district’s polic-ing unit and that of the population, and the heterogeneity across police units re�ectsChicago’s racial segregation across geographic areas.
8
C.2 O�cer Race and Patrol Assignments
Among o�cers assigned to a particular police district, considerable variation exists in theexact patrol assignments that o�cers receive. We examine each unit individually, tabu-lating o�cer race and shift time assignments (�rst, second and third watch, respectivelycorresponding to the nominal duty periods of midnight to 8 a.m., 8 a.m. to 4 p.m., and4 p.m. to midnight). Figure A2 depicts the frequency of each shift period, showing thatthe pattern of assignments di�ers dramatically by o�cer race. For example, white o�-cers in Wentworth (District 2) almost exclusively serve from 4 p.m. to midnight, whereasblack o�cers are more likely to be assigned to mid-day shifts. Similarly, Figure A3 showsthat relative to white o�cers, black o�cers are far more frequently deployed to assignedbeat 202—which roughly corresponds to a patrol area in the district’s southwest corner(depicted in Figure 1) that has extremely high police activity and a high concentrationof black residents. These results undermine analyses in a wide array of previous studiesthat aggregate at high levels of geography (for example, controlling for district or unitassignment) and which assume that o�cers face homogeneous conditions within thesecrude groupings.
9
Figure A2: Assigned shift time by o�cer race. Each panel depicts o�cers within a geographic police district. Within adistrict, each row shows the proportion of shift assignments for black, Hispanic, or white o�cers to �rst watch (midnight to8 a.m.), second watch (8 a.m. to 4 p.m.), and third watch (4 p.m. to midnight). Darker cells indicate a higher proportion ofassignments to a shift time, and entries in a row sum to unity. The �gure demonstrates that within any particular district,black and Hispanic o�cers are called to serve at very di�erent times of day (p < 0.001 for all within-district statistical tests ofindependence).
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
1st watch 2nd watch 3rd watch 1st watch 2nd watch 3rd watch 1st watch 2nd watch 3rd watch 1st watch 2nd watch 3rd watch 1st watch 2nd watch 3rd watch
1st watch 2nd watch 3rd watch 1st watch 2nd watch 3rd watch 1st watch 2nd watch 3rd watch 1st watch 2nd watch 3rd watch 1st watch 2nd watch 3rd watch
1st watch 2nd watch 3rd watch 1st watch 2nd watch 3rd watch 1st watch 2nd watch 3rd watch 1st watch 2nd watch 3rd watch 1st watch 2nd watch 3rd watch
1st watch 2nd watch 3rd watch 1st watch 2nd watch 3rd watch 1st watch 2nd watch 3rd watch 1st watch 2nd watch 3rd watch 1st watch 2nd watch 3rd watch
1st watch 2nd watch 3rd watch 1st watch 2nd watch 3rd watch 1st watch 2nd watch 3rd watch 1st watch 2nd watch 3rd watch 1st watch 2nd watch 3rd watch
Black
Hispanic
White
Black
Hispanic
White
Black
Hispanic
White
Black
Hispanic
White
Black
Hispanic
White
Black
Hispanic
White
Black
Hispanic
White
Black
Hispanic
White
Black
Hispanic
White
Black
Hispanic
White
Black
Hispanic
White
Black
Hispanic
White
Black
Hispanic
White
Black
Hispanic
White
Black
Hispanic
White
Black
Hispanic
White
Black
Hispanic
White
Black
Hispanic
White
Black
Hispanic
White
Black
Hispanic
White
Black
Hispanic
White
Black
Hispanic
White
Black
Hispanic
White
Black
Hispanic
White
Black
Hispanic
White
Assigned shift
0.00 0.25 0.50 0.75Proportion of officer-shifts assigned to shift
10
Figure A3: Assigned beat by o�cer race. Each panel depicts o�cers within a geographic police district. Within a district,each row shows the proportion of shift assignments for black, Hispanic, or white o�cers to beats, or geographic patrol areas.Darker cells indicate a higher proportion of assignments to a beat, and entries in a row sum to unity. The �gure demonstratesthat within any particular district, black and Hispanic o�cers are called to serve at very di�erent locations (p < 0.001 for allwithin-district statistical tests of independence).
District 25: Grand Central
District 22: Morgan Park District 23: Town Hall District 24: Rogers Park
District 19: Belmont District 20: Lincoln District 21: Prairie
District 16: Jefferson Park District 17: Albany Park District 18: Near North
District 13: Wood District 14: Shakespeare District 15: Austin
District 10: Ogden District 11: Harrison District 12: Monroe
District 7: Englewood District 8: Chicago Lawn District 9: Deering
District 4: South Chicago District 5: Calumet District 6: Gresham
District 1: Central District 2: Wentworth District 3: Grand Crossing
2500
2504
2549
2559
2574
2579
2556
2553
2547
2592
2599
2594
2578
2576
2543
2585
2506
2545
2584
2541
2570
2583
2590
2572
2551
2540
2501
2582
2575
2580
2586
2554
2548
2503
2507
2593
2577
2530
2552
2505
2550
2546
2510
2544
2534
2520
2513
2542
2531
2523
2525
2522
2515
2521
2512
2535
2532
2533
2524
2514
2511
2573
2502
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
2408
2470
2403
2458
2459
2406
2448
2457
2404
2456
2449
2471
2446
2440
2490
2400
2495
2494
2454
2455
2492
2453
2451
2452
2493
2442
2482
2407
2476
2447
2499
2483
2420
2410
2443
2441
2450
2430
2412
2481
2486
2487
2485
2445
2484
2424
2422
2444
2401
2411
2433
2405
2413
2431
2423
2432
2402
2472
1991
1997
1958
1977
1989
1987
1992
1996
1903
1957
1943
1952
1993
1982
1995
1942
1947
1974
1949
1904
1946
1970
1948
1954
1908
1955
1953
1956
1941
1940
1990
1950
1988
1945
1981
1994
1951
1984
1983
1944
1980
1985
1930
1900
1907
1906
1999
1986
1933
1905
1926
1921
1911
1915
1972
1910
1925
1920
1935
1914
1931
1932
1923
1912
1934
1913
1922
1971
1924
1901
1902
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
1000
1007
1070
1076
1079
1081
1091
1055
1003
1056
1086
1074
1053
1092
1006
1045
1054
1044
1077
1051
1065
1075
1084
1042
1096
1008
1072
1043
1088
1004
1090
1041
1046
1031
1050
1052
1032
1034
1082
1001
1095
1020
1023
1089
1083
1022
1093
1005
1010
1033
1040
1099
1030
1012
1013
1011
1024
1021
1014
1071
1002
1100
1107
1108
1147
1157
1174
1146
1179
1158
1177
1192
1173
1145
1175
1153
1178
1155
1193
1176
1189
1144
1154
1190
1165
1185
1106
1156
1194
1151
1170
1150
1184
1182
1103
1186
1142
1143
1152
1140
1181
1199
1172
1187
1195
1183
1159
1141
1105
1132
1134
1115
1110
1130
1101
1120
1125
1114
1188
1121
1113
1131
1122
1135
1124
1123
1112
1111
1133
1171
1102
1207
1273
1277
1288
1248
1295
1247
1274
1279
1256
1265
1280
1281
1246
1244
1206
1291
1294
1254
1285
1270
1250
1208
1251
1292
1201
1240
1255
1290
1299
1241
1283
1252
1209
1253
1214
1243
1200
1245
1293
1215
1221
1242
1234
1230
1235
1210
1233
1211
1271
1220
1205
1212
1223
1225
1272
1213
1232
1224
1222
1231
1202
703
707
708
775
776
7775 781
785
788
795
786
757
758
770
789
754
784
755
792
706
750
794
783
746
753
787
756
741
700
772
751
790
752
774
730
793
743
745
705
740
742
799
720
701
735
714
744
724
722
782
702
723
732
734
710
733
731
725
726
715
771
712
713
711
806
880
886
889
879
876
878
874
877
875
870
883
850
890
803
853
857
895
893
856
894
858
840
852
844
848
882
851
896
854
807
884
847
846
842
800
899
845
805
822
843
831
885
892
849
830
835
810
859
855
832
821
823
841
801
820
873
825
833
871
824
834
814
815
811
812
813
872
802
983
974
986
991
903
947
981
984
994
993
959
958
965
908
995
970
950
987
945
941
942
953
978
956
951
944
946
943
985
971
990
901
940
989
955
900
907
954
952
906
999
910
930
920
932
921
923
905
988
924
934
913
935
922
912
972
931
933
915
973
914
925
911
902
400
404
408
483
484
485
487
488
489
496
457
477
475
481
478
482
452
454
405
442
458
450
444
440
443
465
441
407
456
486
446
403
494
451
490
406
470
495
479
447
414
401
499
445
412
472
431
453
410
421
422
411
423
471
420
430
455
432
424
413
433
402
434
500
503
504
507
546
555
581
582
583
584
585
586
598
549
556
565
576
577
597
578
591
547
575
579
564
572
554
596
544
550
595
501
545
553
590
506
505
540
531
508
570
543
542
552
551
512
532
523
599
502
530
541
533
511
513
510
520
522
524
571
600
607
608
657
661
670
675
676
677
678
682
691
695
689
686
690
656
683
662
668
688
699
693
663
681
694
665
640
655
605
653
647
646
644
684
633
687
643
622
672
642
650
645
602
654
606
641
614
632
630
634
631
624
651
652
685
623
601
612
620
621
610
611
671
613
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
200
203
207
208
251
253
255
256
270
274
275
276
277
278
279
285
286
289
297
284
254
282
204
252
247
250
296
281
245
205
272
246
283
240
214
294
206
234
222
223
244
280
224
290
202
233
212
243
235
213
211
287
232
299
242
241
230
295
201
215
288
220
210
225
271
231
221
300
307
308
356
365
370
373
374
375
376
377
378
379
381
387
388
354
395
391
382
346
303
352
304
350
355
351
347
306
386
353
389
384
390
368
344
385
369
380
333
372
340
305
345
371
343
314
341
330
342
301
311
399
321
332
320
331
334
383
323
324
313
310
322
312
302
BlackHispanic
White
BlackHispanic
White
BlackHispanic
White
BlackHispanic
White
BlackHispanic
White
BlackHispanic
White
BlackHispanic
White
BlackHispanic
White
BlackHispanic
White
BlackHispanic
White
BlackHispanic
White
BlackHispanic
White
BlackHispanic
White
BlackHispanic
White
BlackHispanic
White
BlackHispanic
White
BlackHispanic
White
BlackHispanic
White
BlackHispanic
White
BlackHispanic
White
BlackHispanic
White
BlackHispanic
White
BlackHispanic
White
BlackHispanic
White
BlackHispanic
White
Assigned beat
0.0 0.1 0.2 0.3 0.4 0.5Proportion of officer-shifts assigned to beat
11
C.3 O�cers’ Political Preferences
12
Figure A4: Police are more likely to vote than co-racial and co-ethnic civilians.O�cer and city resident turnout rates by racial/ethnic groups, cohort, and election.
Black Hispanic White
20
00
20
02
20
04
20
06
20
08
20
10
20
12
20
14
20
16
1950 1960 1970 1980 1990 1950 1960 1970 1980 1990 1950 1960 1970 1980 1990
0.00.20.40.60.8
0.00.20.40.6
0.000.250.500.75
0.00.20.40.6
0.000.250.500.75
0.00.20.40.60.8
0.000.250.500.75
0.00.20.40.60.8
0.000.250.500.75
Birth year
Pro
port
ion
Race/Ethnicity Black Hispanic White
Group Chicago city population Chicago Police Department
13
Figure A5: Party identi�cation among o�cers using 2016 party primary partici-pation as proxy for party ID. O�cers aggregated to �ve-year windows on birth year(1950–1954, 1955–1960, etc.). For yearly bins and previous elections (2000–2016).
Black Hispanic White
Dem
ocra
tic prim
ary
Neith
er p
rimary
Republica
n p
rimary
1950 1960 1970 1980 1990 1950 1960 1970 1980 1990 1950 1960 1970 1980 1990
0.0
0.2
0.4
0.6
0.0
0.2
0.4
0.6
0.0
0.1
0.2
0.3
0.4
Birth year
Pro
port
ion
Race/Ethnicity Black Hispanic White
Group Chicago city population Chicago Police Department
14
C.4 O�cer Behavior: Overall Means
15
Table A1: Average events per shift. Mean number of stops, arrests, uses of force, andcomplaints without adjustment for time or location. Typical behavior is reported for 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�cers areexcluded due to small sample sizes. O�cer behavior toward Native American/Alaskanand Asian/Paci�c Islander civilians is not included for the purposes of computing totaland reason-speci�c events. Note that civilian race is often unknown in complaint records.Values are scaled for ease of interpretation.
Mean Mean Mean MeanBehavior (Pooled) (Black o�.) (Hisp. o�.) (White o�.)
Stops per shift:Civilian race: Black 0.303 0.261 0.309 0.325Civilian race: Hispanic 0.092 0.021 0.143 0.110Civilian race: White 0.057 0.021 0.061 0.075Reason: Other 0.131 0.116 0.135 0.139Reason: Loitering 0.006 0.003 0.009 0.007Reason: Suspicious 0.154 0.089 0.173 0.184Reason: Drug 0.048 0.018 0.063 0.058Reason: Tra�c 0.118 0.081 0.139 0.130Total 0.458 0.306 0.519 0.517
Arrests per 10 shifts:Civilian race: Black 0.564 0.495 0.593 0.591Civilian race: Hispanic 0.176 0.040 0.252 0.219Civilian race: White 0.089 0.032 0.098 0.117Reason: Other 0.310 0.204 0.346 0.354Reason: Drug 0.094 0.036 0.124 0.113Reason: Tra�c 0.070 0.035 0.078 0.087Reason: Property 0.152 0.115 0.166 0.168Reason: Violent 0.211 0.182 0.238 0.216Total 0.837 0.571 0.952 0.936
Uses of force per 100 shifts:Civilian race: Black 0.199 0.145 0.210 0.226Civilian race: Hispanic 0.041 0.009 0.058 0.050Civilian race: White 0.026 0.007 0.028 0.036Total 0.271 0.163 0.302 0.319
Complaints per 1,000 shifts:Civilian race: Black 0.542 0.604 0.493 0.530Civilian race: Hispanic 0.102 0.025 0.147 0.124Civilian race: White 0.084 0.048 0.081 0.106Reason: Other 0.988 1.181 0.852 0.941Reason: Arrest/Search 0.541 0.367 0.596 0.616Reason: Force 0.386 0.354 0.383 0.407Reason: Unknown 0.106 0.099 0.120 0.103Total 2.021 2.001 1.951 2.06716
C.5 O�cer Behavior: Shift Duration
We consider the possibility that stop, arrest, force, and complaint volume are driven bydi�erent amounts of time spent patrolling. Even among o�cers assigned to a particularshift time (a nominal eight-hour patrol period), minor variation exists in the precise startand end of the o�cer’s duty time. Of the o�cer-shifts analyzed, 85.5% are 9 hours induration, with 8.5- and 8-hour shifts making up an additional 7.8% and 5.1%, respectively(percentages are based on rounding shift duration to the nearest 6 minutes). In �xed-e�ect regression analyses that compare o�cers within unique MDSB combinations, weestimate that shifts of black o�cers are 0.0094 hours shorter (roughly 0.1% shorter) thantheir white counterparts assigned to the same MDSB, and Hispanic o�cer shift durationsare statistically indistinguishable from those of white o�cers. Because these di�erencesare two orders of magnitude smaller than reported di�erences in behavior, patrol timedisparities are unlikely to be a mechanism driving observed racial gaps in stop, arrest,force, and complaint volume.
17
C.6 O�cer Behavior: Regression Tables
18
Table A2: Models of Stops, by Civilian Race
Stops per shift, by civilian race:Black Hispanic White All races
O�cer black −0.133 −0.014 −0.013 −0.161[−0.156; −0.122] [−0.019; −0.011] [−0.016; −0.011] [−0.188; −0.148]
O�cer Hispanic −0.032 0.005 −0.004 −0.032[−0.047; −0.019] [−0.001; 0.010] [−0.008; −0.001] [−0.052; −0.016]
MDSB FE Yes Yes Yes YesN 2923458 2923458 2923458 2923458Ne�ective 2186703 2186703 2186703 2186703R2 0.295 0.230 0.183 0.276Speci�cation includes �xed e�ects for each unique combination of month, day group (weekday/weekend), shift time (�rst/second/third watch), and assigned beat. Cluster-robust95% con�dence intervals are computed by block bootstrap on o�cers and reported in brackets. N and Ne�ective indicate the number of o�cer-shifts analyzed and the number ofo�cer-shifts within MDSB containing variation on o�cer race, respectively.
Table A3: Models of Stops, by Reason for Stop
Stops per shift, by reason:Loitering Suspicious Drug Tra�c Other
O�cer black −0.004 −0.061 −0.017 −0.053 −0.026[−0.005; −0.003] [−0.071; −0.055] [−0.022; −0.014] [−0.067; −0.047] [−0.034; −0.017]
O�cer Hispanic 0 −0.019 −0.005 −0.005 −0.003[−0.001; 0.002] [−0.026; −0.011] [−0.010; −0.002] [−0.016; 0.002] [−0.010; 0.004]
MDSB FE Yes Yes Yes Yes YesN 2923458 2923458 2923458 2923458 2923458Ne�ective 2186703 2186703 2186703 2186703 2186703R2 0.170 0.232 0.228 0.268 0.206Speci�cation includes �xed e�ects for each unique combination of month, day group (weekday/weekend), shift time (�rst/second/third watch), and assigned beat. Cluster-robust95% con�dence intervals are computed by block bootstrap on o�cers and reported in brackets. N and Ne�ective indicate the number of o�cer-shifts analyzed and the number ofo�cer-shifts within MDSB containing variation on o�cer race, respectively.
19
Table A4: Models of Arrests, by Civilian Race
Arrests per 10 shifts, by civilian race:Black Hispanic White All races
O�cer black −0.153 −0.031 −0.019 −0.203[−0.192; −0.129] [−0.043; −0.022] [−0.027; −0.011] [−0.249; −0.175]
O�cer Hispanic −0.034 −0.008 −0.009 −0.050[−0.059; −0.013] [−0.021; 0.004] [−0.016; −0.004] [−0.085; −0.020]
MDSB FE Yes Yes Yes YesN 2930957 2930957 2930957 2930957Ne�ective 2193343 2193343 2193343 2193343R2 0.183 0.160 0.123 0.185Speci�cation includes �xed e�ects for each unique combination of month, day group (weekday/weekend), shift time (�rst/second/third watch), and assigned beat. Cluster-robust95% con�dence intervals are computed by block bootstrap on o�cers and reported in brackets. N and Ne�ective indicate the number of o�cer-shifts analyzed and the number ofo�cer-shifts within MDSB containing variation on o�cer race, respectively.
Table A5: Models of Arrests, by Reason for Arrest
Arrests per 10 shifts, by reason:Tra�c Drug Property Violent Other
O�cer black −0.019 −0.034 −0.030 −0.022 −0.098[−0.029; −0.012] [−0.046; −0.027] [−0.038; −0.023] [−0.033; −0.013] [−0.123; −0.079]
O�cer Hispanic −0.009 0.001 −0.008 0.002 −0.037[−0.017; −0.002] [−0.010; 0.013] [−0.015; −0.003] [−0.006; 0.011] [−0.056; −0.020]
MDSB FE Yes Yes Yes Yes YesN 2930957 2930957 2930957 2930957 2930957Ne�ective 2193343 2193343 2193343 2193343 2193343R2 0.161 0.173 0.128 0.114 0.157Speci�cation includes �xed e�ects for each unique combination of month, day group (weekday/weekend), shift time (�rst/second/third watch), and assigned beat. Cluster-robust95% con�dence intervals are computed by block bootstrap on o�cers and reported in brackets. N and Ne�ective indicate the number of o�cer-shifts analyzed and the number ofo�cer-shifts within MDSB containing variation on o�cer race, respectively.
20
Table A6: Models of Force, by Civilian Race
Uses of force per 100 shifts, by civilian race:Black Hispanic White All races
O�cer black −0.093 −0.009 −0.006 −0.112[−0.126; −0.068] [−0.019; −0.002] [−0.013; −0.001] [−0.149; −0.085]
O�cer Hispanic −0.032 0.006 −0.006 −0.034[−0.062; −0.012] [−0.005; 0.016] [−0.014; 0.002] [−0.068; −0.010]
MDSB FE Yes Yes Yes YesN 2933232 2933232 2933232 2933232Ne�ective 2195410 2195410 2195410 2195410R2 0.094 0.095 0.096 0.095Speci�cation includes �xed e�ects for each unique combination of month, day group (weekday/weekend), shift time (�rst/second/third watch), and assigned beat. Cluster-robust95% con�dence intervals are computed by block bootstrap on o�cers and reported in brackets. N and Ne�ective indicate the number of o�cer-shifts analyzed and the number ofo�cer-shifts within MDSB containing variation on o�cer race, respectively.
Table A7: Models of Complaints, by Civilian Race
Complaints per 1,000 shifts, by civilian race:Black Hispanic White All races
O�cer black −0.066 0.007 −0.012 −0.058[−0.213; 0.062] [−0.039; 0.047] [−0.055; 0.026] [−0.323; 0.153]
O�cer Hispanic −0.099 −0.012 −0.020 −0.163[−0.206; −0.012] [−0.066; 0.032] [−0.065; 0.013] [−0.383; 0.017]
MDSB FE Yes Yes Yes YesN 2933278 2933278 2933278 2933278Ne�ective 2195437 2195437 2195437 2195437R2 0.091 0.087 0.089 0.092Speci�cation includes �xed e�ects for each unique combination of month, day group (weekday/weekend), shift time (�rst/second/third watch), and assigned beat. Cluster-robust95% con�dence intervals are computed by block bootstrap on o�cers and reported in brackets. N and Ne�ective indicate the number of o�cer-shifts analyzed and the number ofo�cer-shifts within MDSB containing variation on o�cer race, respectively.
21
Table A8: Models of Complaints, by Reason for Complaint
Complaints per 1,000 shifts, by reason:Arrest/Search Force Unknown Other
O�cer black −0.222 −0.030 0.019 0.175[−0.342; −0.108] [−0.146; 0.070] [−0.035; 0.073] [−0.001; 0.338]
O�cer Hispanic −0.012 −0.087 −0.003 −0.061[−0.112; 0.096] [−0.176; −0.005] [−0.049; 0.034] [−0.180; 0.056]
MDSB FE Yes Yes Yes YesN 2933278 2933278 2933278 2933278Ne�ective 2195437 2195437 2195437 2195437R2 0.104 0.087 0.076 0.087Speci�cation includes �xed e�ects for each unique combination of month, day group (weekday/weekend), shift time (�rst/second/third watch), and assigned beat. Cluster-robust95% con�dence intervals are computed by block bootstrap on o�cers and reported in brackets. N and Ne�ective indicate the number of o�cer-shifts analyzed and the number ofo�cer-shifts within MDSB containing variation on o�cer race, respectively.
22
C.7 O�cer Behavior: Alternative Estimators
23
Our primary analysis of o�cer behavior uses OLS regression with MDSB �xed e�ects,of the form Yi,j = �i +∑r �r1(Rij = r), where �i represents a �xed e�ect for MDSB i. As wediscuss in SI Section B, 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 assign-ment slot with a minority o�cer on resulting stop, arrest, use-of-force, and complaintvolume) if MDSB-speci�c policy e�ects are highly variable in a way that is associatedwith the proportion of minority o�cers that are assigned to MDSBs (in this case, it is wellknown that OLS recovers the weighted average of MDSB-speci�c policy e�ects, whereweights are determined by variance of o�cer race within the MDSB).
To gauge robustness of our results to the violation of this assumption, we presentalternative estimates in SI Figures A6–A9 below. The �rst alternative estimator takes thewithin-MDSB di�erence in behavior between average minority- and white-o�cer patrolassignments, then aggregates these according to the number of patrol assignment slots ineach MDSB. Following the notation de�ned in SI Section B, this estimator can be writtenas
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 computethe unweighted 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 racial/ethnic di�erences in o�cerbehavior are driven by di�erences in experience between minority and white o�cers. Ifthis were the case, it would undermine the applicability of our results to the e�ect of ahiring reform that brought in additional minority rookie o�cers. To examine whetherthese di�erences impact our results, we extend the regression speci�cation by addingadditional linear and quadratic terms for each o�cer’s length of service.
24
Figure A6: Stops. Points (error bars) depict estimated di�erences (95% block-bootstrap con�dence intervals) in the number ofcivilian stops conducted by black and Hispanic o�cers per shift, relative to white o�cers patrolling in the same month, daygroup (weekday/weekend), shift time (�rst/second/third watch), and beat. Results presented using four di�erent estimators.
Black officers (vs white) Hispanic officers (vs white)
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
-0.20
-0.15
-0.10
-0.05
0.00
Diffe
rence
s in
sto
ps
per
shift
by
off
icers
of
each
raci
al/eth
nic
gro
up,
vers
us
whit
e o
ffic
ers
faci
ng C
C
Count-weighted mean of within-MDSB differencesOLS with MDSB FE
OLS with MDSB FE and experience controlsSimple mean of within-MDSB differences
25
Figure A7: Arrests. Points (error bars) depict estimated di�erences (95% block-bootstrap con�dence intervals) in the numberof arrests conducted by black and Hispanic o�cers per shift, relative to white o�cers patrolling in the same month, day group(weekday/weekend), shift time (�rst/second/third watch), and beat. Results presented using four di�erent estimators.
Black officers (vs white) Hispanic officers (vs white)
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
Diffe
rence
s in
arr
est
s per
10
shifts
by
off
icers
of
each
raci
al/eth
nic
gro
up,
vers
us
whit
e o
ffic
ers
faci
ng C
C
Count-weighted mean of within-MDSB differencesOLS with MDSB FE
OLS with MDSB FE and experience controlsSimple mean of within-MDSB differences
26
Figure A8: Uses of Force. Points (error bars) depict estimated di�erences (95% block-bootstrap con�dence intervals) in thenumber of uses of force by black and Hispanic o�cers per shift, relative to white o�cers patrolling in the same month, daygroup (weekday/weekend), shift time (�rst/second/third watch), and beat. Results presented using four di�erent estimators.
Black officers (vs white) Hispanic officers (vs white)
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
-0.15
-0.10
-0.05
0.00
Diffe
rence
s in
use
s of
forc
e p
er
10
0 s
hifts
by o
ffic
ers
of
each
raci
al/eth
nic
gro
up,
vers
us
whit
e o
ffic
ers
faci
ng C
C
Count-weighted mean of within-MDSB differencesOLS with MDSB FE
OLS with MDSB FE and experience controlsSimple mean of within-MDSB differences
27
Figure A9: Complaints. Points (error bars) depict estimated di�erences (95% block-bootstrap con�dence intervals) in thenumber of civilian complaints registered against black and Hispanic o�cers per shift, relative to white o�cers patrolling inthe same month, day group (weekday/weekend), shift time (�rst/second/third watch), and beat. Results presented using fourdi�erent estimators.
Black officers (vs white) Hispanic officers (vs white)
Civ. r
ace:
Bla
ck
Civ. r
ace:
Hisp
anic
Civ. r
ace:
Whi
teRea
son:
Oth
er
Reaso
n: A
rrest
/sea
rch
Reaso
n: F
orce
Reaso
n: U
nkno
wn
Tota
lCiv
. rac
e: B
lack
Civ. r
ace:
Hisp
anic
Civ. r
ace:
Whi
teRea
son:
Oth
er
Reaso
n: A
rrest
/sea
rch
Reaso
n: F
orce
Reaso
n: U
nkno
wn
Tota
l
-0.6
-0.3
0.0
0.3
Diffe
rence
s in
com
pla
ints
per
1k
shifts
by o
ffic
ers
of
each
raci
al/eth
nic
gro
up,
vers
us
whit
e o
ffic
ers
faci
ng C
C
Count-weighted mean of within-MDSB differencesOLS with MDSB FE
OLS with MDSB FE and experience controlsSimple mean of within-MDSB differences
28
C.8 O�cers’ Spanish Language Pro�ciency
29
Figure A10: Comparison of Spanish-speaking Hispanic O�cers to White O�cers. Points (error bars) depict estimateddi�erences (95% block-bootstrap con�dence intervals) in the number of civilian stops, arrests, uses of force, and civilian com-plaints generated by white o�cers per shift, relative to Spanish-speaking o�cers patrolling in the same month, day group(weekday/weekend), shift time (�rst/second/third watch), and beat.
Force uses per 100 shifts Complaints per 1,000 shifts
Stops per shift Arrests per 10 shifts
Civ. r
ace:
Bla
ckCiv
. rac
e: H
ispan
icCiv
. rac
e: W
hite
Tota
l
Civ. r
ace:
Bla
ck
Civ. r
ace:
Hisp
anic
Civ. r
ace:
Whi
te
Reaso
n: O
ther
Reaso
n: A
rrest
/sea
rch
Reaso
n: F
orce
Reaso
n: U
nkno
wnTo
tal
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
s
Reaso
n: D
rug
Reaso
n: T
raffi
cTo
tal
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
ntTo
tal
0.00
0.05
0.10
0.15
0.20
-0.2
-0.1
0.0
0.1
0.2
0.3
0.00
0.05
0.10
0.00
0.03
0.06
0.09
Diffe
rence
s in
behavio
rs f
or
whit
eoff
icers
, vers
us
SS
H o
ffic
ers
faci
ng
CC
30
C.9 O�cer Gender
31
Figure A11: Stops. Points (error bars) depict estimated di�erences (95% block-bootstrap con�dence intervals) in the numberof stops of civilians by black and Hispanic male and female o�cers per shift, relative to white o�cers patrolling in the samemonth, day group (weekday/weekend), shift time (�rst/second/third watch), and beat.
Black officers (vs white) Hispanic officers (vs white)
Civ. r
ace:
Bla
ck
Civ. r
ace:
Hisp
anic
Civ. r
ace:
Whi
te
Reaso
n: L
oite
ring
Reaso
n: S
uspi
ciou
sRea
son:
Dru
g
Reaso
n: T
raffi
c
Reaso
n: O
ther
Tota
lCiv
. rac
e: B
lack
Civ. r
ace:
Hisp
anic
Civ. r
ace:
Whi
te
Reaso
n: L
oite
ring
Reaso
n: S
uspi
ciou
sRea
son:
Dru
g
Reaso
n: T
raffi
c
Reaso
n: O
ther
Tota
l
-0.20
-0.15
-0.10
-0.05
0.00
Diffe
rence
s in
sto
ps
per
shift
by
off
icers
of
each
raci
al/eth
nic
gro
up,
vers
us
whit
e o
ffic
ers
faci
ng C
C
Gender Female officers Male officers
32
Figure A12: Arrests. Points (error bars) depict estimated di�erences (95% block-bootstrap con�dence intervals) in the numberof arrests by black and Hispanic male and female o�cers per shift, relative to white o�cers patrolling in the same month, daygroup (weekday/weekend), shift time (�rst/second/third watch), and beat.
Black officers (vs white) Hispanic officers (vs white)
Civ. r
ace:
Bla
ck
Civ. r
ace:
Hisp
anic
Civ. r
ace:
Whi
teRea
son:
Dru
g
Reaso
n: T
raffi
c
Reaso
n: P
rope
rty
Reaso
n: V
iole
nt
Reaso
n: O
ther
Tota
lCiv
. rac
e: B
lack
Civ. r
ace:
Hisp
anic
Civ. r
ace:
Whi
teRea
son:
Dru
g
Reaso
n: T
raffi
c
Reaso
n: P
rope
rty
Reaso
n: V
iole
nt
Reaso
n: O
ther
Tota
l
-0.2
-0.1
0.0
Diffe
rence
s in
arr
est
s per
10
shifts
by o
ffic
ers
of
each
raci
al/eth
nic
gro
up,
vers
us
whit
e o
ffic
ers
faci
ng C
C
Gender Female officers Male officers
33
Figure A13: Uses of force. Points (error bars) depict estimated di�erences (95% block-bootstrap con�dence intervals) in thenumber of uses of force by black and Hispanic male and female o�cers per shift, relative to white o�cers patrolling in thesame month, day group (weekday/weekend), shift time (�rst/second/third watch), and beat.
Black officers (vs white) Hispanic officers (vs white)
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
-0.15
-0.10
-0.05
0.00
Diffe
rence
s in
use
s of
forc
e p
er
10
0 s
hifts
by o
ffic
ers
of
each
raci
al/eth
nic
gro
up,
vers
us
whit
e o
ffic
ers
faci
ng C
C
Gender Female officers Male officers
34
Figure A14: Complaints. Points (error bars) depict estimated di�erences (95% block-bootstrap con�dence intervals) in thenumber of complaints registered against by black and Hispanic male and female o�cers per shift, relative to white o�cerspatrolling in the same month, day group (weekday/weekend), shift time (�rst/second/third watch), and beat.
Black officers (vs white) Hispanic officers (vs white)
Civ. r
ace:
Bla
ck
Civ. r
ace:
Hisp
anic
Civ. r
ace:
Whi
te
Reaso
n: A
rrest
/sea
rch
Reaso
n: F
orce
Reaso
n: O
ther
Reaso
n: U
nkno
wn
Tota
lCiv
. rac
e: B
lack
Civ. r
ace:
Hisp
anic
Civ. r
ace:
Whi
te
Reaso
n: A
rrest
/sea
rch
Reaso
n: F
orce
Reaso
n: O
ther
Reaso
n: U
nkno
wn
Tota
l
-0.5
0.0
0.5
Diffe
rence
s in
com
pla
ints
per
1k
shifts
by o
ffic
ers
of
each
raci
al/eth
nic
gro
up,
vers
us
whit
e o
ffic
ers
faci
ng C
C
Gender Female officers Male officers
35
C.10 Robustness Checks: Multiple Stopping O�cers
Data on stops of civilians indicate that 87% of stops report two o�cers as participatingin the stop. In our main analysis, we treat a stop by two o�cers as two incidents in thedata, as both o�cers contribute to the decision to engage a civilian. To gauge the extent towhich this decision drives our results, we present an alternative analysis of stops in whichwe use only data on �rst reporting o�cers. As Figure A15 shows, results are essentiallyunchanged when imposing this restriction.
36
Figure A15: Stops. Points (error bars) depict estimated di�erences (95% block-bootstrap con�dence intervals) in the numberof civilian stops conducted by black and Hispanic o�cers per shift, relative to white o�cers patrolling in the same month,day group (weekday/weekend), shift time (�rst/second/third watch), and beat, using only data generated by the �rst reportingo�cer.
Black officers (vs white) Hispanic officers (vs white)
Civ. rac
e: Blac
k
Civ. rac
e: Hisp
anic
Civ. rac
e: Whit
e
Reaso
n: Lo
iterin
g
Reaso
n: Sus
piciou
sRea
son:
DrugRea
son:
Traffic
Reaso
n: Othe
rTota
lCiv.
race:
Black
Civ. rac
e: Hisp
anic
Civ. rac
e: Whit
e
Reaso
n: Lo
iterin
g
Reaso
n: Sus
piciou
sRea
son:
DrugRea
son:
Traffic
Reaso
n: Othe
rTota
l
-0.12
-0.08
-0.04
0.00
Diff
eren
ces
in s
tops
per
shi
ft by
offic
ers
of e
ach
raci
al/e
thni
c gr
oup,
ver
sus
whi
te o
ffice
rs fa
cing
CC
37