RACE ON POLICING: Results of Estimation
The set of covariates included in the regressions is constrained by the lack of data available on an annual basis at the city level. While some variables, such as city population and the presence of a black mayor, are available annually for cities, in other cases compromises must be made. We attempt to deal with these data limitations in three ways. First, where annual data for larger geographic areas exist, we use the most disaggregated data series available. Thus, SMSA-level unemployment rates, state per capita income, and state measures of the age distribution are included as regressors. Second, where city-level measures are critical, as with the percent black, we linearly interpolate between decennial censuses.
Finally, as a substitute for effective covariates, we include year dummies, city-fixed effects, and, in some specifications, region-year interactions using the nine U. S. census regions. These variables absorb much of the variation in the data, particularly for demographic and socio-economic factors which tend to change slowly over time. For instance, year and city dummies alone eliminate over 95 percent of the variation in the demographic variables and over 90 percent of the variation in per capita income. To the extent that other (unmeasured) demographic and socio-economic factors exhibit a similar pattern, the use of these indicator variables will reduce any omitted-variable bias from this source.
Results of Estimation
Table 2 presents the results from estimation of the relationship between arrest patterns by race of suspect and the racial composition of the police force. The specifications estimated are of the form
where с indexes cities, t corresponds to years, and X is the vector of economic, socio-economic and demographic controls described above. The arrest variables are arrests in a given crime category per member of the racial classification (e.g. violent crime arrests per non-white resident). The police variables are per capita using the city population as the denominator. This specification assumes that any given individual’s probability of arrest is a linear function of the number of white and non-white police per capita (all of the results are also robust to estimation in logs).
One advantage of this choice of specification is that the interpretation of the p coefficients does not depend on racial composition of a city’s population. The literal interpretation of the p coefficients is the change in the number of arrests when one officer of a given race is added to the police force. Our primary focus is on the difference in p by race of police, i.e. how much do arrests of suspects of a given race change as a function of the racial composition of the police force holding the total number of police constant. The у and X terms represent city-fixed effects and year dummies respectively.
The eight columns of Table 2 correspond to results by race of arrestee for four different arrest categories: total arrests (including the three other categories examined, as well as public order offenses, prostitution, drunk driving, and a wide range of other generally minor crimes), property crime (burglary, larceny, and auto theft), violent crime (murder, rape, robbery, and aggravated assault), and drug offenses (both possession and distribution). Almost two-thirds of all arrests are for offenses not covered by the last three categories. Year dummies and city-fixed effects are included in the regressions, but are not reported in the table. All specifications are estimated using weighted least squares with weights proportional to city population.
We turn our attention first to the top two rows of Table 2 which present the parameters of primary interest. The top row contains estimates of the change in the number of arrests of each type with respect to the number of white police officers. The second row presents that same coefficient for non-white officers.