Permisos de trabajo

Is there evidence of racial bias among police officers in Dallas? In this section, I test for the presence of racial bias by investigating relationships between officer race and arrestee race.

I adapt the test of racial bias used in work by [6] examining officer bias in traffic stops. [6] develop a model of officer interactions that allows officers of different races to have differing arrest behavior as long as these differences are independent of suspect race. Specifically, they test whether the relative rank order of arrest rates across officer race groups is the same within each suspect race. For example, if White officers are more likely than Black officers to arrest a suspect from any race group, then this reflects the

total arrest preferences of White and Black officers but does not imply that either group is racially biased. Alternatively, if Black officers have higher arrest rates than White officers for White suspects and White officers have higher arrest rates than Black officers for Black suspects, either Black officers or White officers are racially biased (or both). Critically, the test does not find evidence of racial bias if officers statistically discriminate against suspects, or if officers of all races use suspect race as a signal of offending characteristics that are correlated with race. More generally, the test allows suspect race groups to have different compositions or unobservable characteristics that may cause differences in arrest rates for suspects of different races across all officers. Arrest rates can differ across suspect race groups in the test because differences in officer race arrest rates are always measured as a relative ranking within a suspect race group.

AppendixA.4 presents the economic framework for the racial bias test, adapted from [6] to the call for service setting. In the call for service setting, officers maximize their expected benefit of making an arrest and face costs of exerting effort that may differ by officer and suspect race. Officers make effort choices after viewing the suspect race and a signal of whether the arrest is feasible or whether their will be a sufficient basis for an arrest if the officer exerts effort. The distribution of signals of feasibility and underlying feasibility of arrests are allowed to differ by suspect race, and these differences enable officers to statistically discriminate against suspects of a particular race. A key difference between the call for service setting and the traffic stop setting

is that I do not directly observe suspect race for all incidents.25 _{Instead, I}

execute this test by comparing officer arrest propensities for different arrestee race outcomes.

I consider unconditional arrestee race outcomes that take a value of 1 if an individual is arrested and is a particular race, either Black, Hispanic, or White, and 0 otherwise. As before, I estimate ✓i,r terms for each arrestee

race outcome, r, using the model outlined in Section 1.3. Next, I regress the ✓i,rterms on the full set of officer demographic characteristics and examine the

impact of officer race on officer arrest propensity. This regression framework allows me to directly test whether the rank of arrest rates across officer race changes for different arrestee race outcomes.26

The test used in this paper offers three new advantages. First, I am able to test for racial bias among officers in a setting that is not affected by officers electing to initiate interactions. Prior work studying racial bias in policing has examined interactions between officer and suspect race in officer-

25_{I have limited information on suspect identification in the data. There are two sets of}

records of suspects in the data: (1) I observe suspects identified and demographic infor- mation for these suspects in a subset of cases (limited to data prior to 2017), (2) I observe demographic characteristics of suspects that are unknown to officers at the conclusion of the response. I treat both of these records as outcomes in the data because they are entered by responding officers and may be a function of officer effort. Arrest records can be considered a subset of (1) and are available for the duration of the sample period.

26_{My preferred specification uses the two stage regression in order to be consistent with the}

analysis in Section1.5.2. I also conduct the test by inserting officer demographic variables

directly into the first stage while omitting the officer fixed effects in Section 1.6.2.2. The

initiated incidents, such as traffic stops [e.g. 56, 4, 5, 6, 48].27 _{These papers}

consider suspect race as a given characteristic of a traffic stop; however, in reality, suspect race is also a choice variable of the officer, who chooses to stop a particular individual. In this paper, I am able to test for the presence of racial bias in a setting that is not affected by this form of selection because calls are initiated by complainants and not officers.

Second, I use a regression framework to control for a large array of observable incident characteristics. [6] address the fact that different officers may face different types of incidents by re-sampling their data to create com- parable incident sets across officer race. In my setting, I use regression models to measure officer race arrest effects adjusted for observable differences in the composition of incidents across officers.

Lastly, I leverage officer identifiers to better understand relation- ships between officer behavior and arrestee race. I am able to use officer identifiers to trace the distribution of officer effects by race for each of the ar- restee race outcomes and find that most variation in officer behavior is within rather than across officer race.

The racial bias test used in this paper cannot detect cases when all officers exhibit similar racial bias toward a particular group, a limitation of prior tests as well. As discussed above, Black arrestees are markedly overrepre-

27_{An exception in this literature is [104], which studies racial bias of state troopers who}

sented in the sample, making up 40% of total incident arrests and only 24% of the Dallas population.28 _{If arrests with Black suspects result in higher arrest}

rates of Black individuals across all officers, this pattern could be consistent with statistical discrimination, or institutional discrimination. Institutional racial bias will occur when the organizational priorities of the department di- rect resources toward policing one race group relative to others, and all officers behave similarly given these priorities. Likewise, the higher representation of Black arrestees could also be consistent with uniform police officer attitudes of taste-based racial bias against Black suspects. Recent evidence on the im- portance of implicit racial bias in decision-making could be consistent with uniform taste-based discrimination against minority groups [e.g.32].

1.6.2 Results of the Test of Racial Bias