Steve Ross

Noise or Signal? Decomposing Sources of Officer Racial Disparities in Traffic Stops

Abstract:  Many states and localities benchmark racial differences in police stops across officers to identify officers who are significant outliers. However, we have little evidence on how important these outliers are in their contribution to overall racial disparities, and to understand their importance we need to examine the broader distribution of officer stop behavior. We estimate Bayesian shrunk officer propensities for their traffic stops to be of minority, as opposed to white, motorists for Connecticut, Massachusetts and New Jersey state troopers, and describe the distribution of officer propensities for each state highway patrol. We utilize detailed information on the location and timing of traffic stops to compare officers that made stops at similar times and locations, and we utilize location data from telemetric devices to evaluate the effectiveness of our controls in capturing officer exposure to minority motorists when the officers were making stops. Conditional on highway location and shift (day, evening and night), we obtain estimates of the standard deviation of the distribution of officer propensities for black versus white and Hispanic versus white stops in the three states of between 3.6 and 6.4 percentage points increased likelihood of a stop being of a minority motorist. We then simulate the officer distribution of stop propensities using the distribution of stops at the highway location by shift level and calculate a counterfactual standard deviation under the null of no systematic heterogeneity. Under the null of no heterogeneity, the finite sample of stops within our location by shift clusters can explain between 30 and 45 percent of the estimated officer stop propensity standard deviations, so that the remaining unexplained standard deviation across officers’ stop propensity ranges between 2.0 and 3.2 percentage points. The substantial standard deviation estimates under the null of no heterogeneity arise even though 1. Adding additional controls for time of day, month of year and day of week has virtually no effect on the individual officer propensity estimates and 2. Officer propensity to be exposed to a high share of motorists from minority neighborhoods is uncorrelated with officer propensity for police stops to be of minority motorists after conditioning on stop location and shift. Finally, we find minimal evidence of skewness in the distribution of officer propensities.