Detecting Bias in Traffic Searches: Examining False Searches of Innocent Drivers

We test current models of racial bias in policing, identify limitations, and propose a test of racial bias, that does not depend on unknown population contraband rate. We conceptualize police officer search decisions as a 2 (search/no search) by 2 (contraband present/absent) table, with missing data (if the police did not search, the presence of contraband is unknown). We constrain the feasible problem space using properties of a 2 x 2 contingency table. Then we examine all possible feasible 2 x 2 tables to identify instances of racial differences in police officer hit and false alarm rates. To do this, we develop a new test of racial bias, the Overlapping Condition Test. We analyze state and county data across 25 United States police departments. These departments have an observable racial difference in false alarm rate regardless of the true value of missing data (under every feasible 2 x 2 table there is a racial difference). This effect is found in 10 out of 14 state police departments and 9 out of 11 local departments across the United States. That is, for every feasible real world scenario police officers have lower false alarm rates for White drivers than Black drivers. We interpret this difference in false alarm rate as a threshold bias. That is, officers use different criteria for searching Black drivers than White drivers and this conclusion is not qualified by the unknown contraband rate. Future directions should explore how police officers make the decision to search drivers and develop interventions to address the racial bias in search rate.