Police Officers on Drug Corners in Philadelphia, Drug Crime, and Violent Crime: Intended, Diffusion, and Displacement Impacts

1. According to the 2000 Census, the population of the Philadelphia primary metropolitan statistical area (PMSA) was 5.1 million, smaller than the New York City–New Jersey, Chicago, and Los Angeles PMSAs but larger than Dallas–Forth Worth at 3.5 million.

2. “(A) police‐led effort to fight drugs and blight in Kensington and North Philadelphia, in 1998 … The campaign, in which police cooperated with other city agencies cleaning up abandoned houses and towing abandoned vehicles, has been widely credited with improving the quality of life in those neighborhoods. Despite a record number of arrests, police were careful in choosing their targets, avoiding the mass sweeps that provoked civil‐liberties questions about some previous police crackdowns” (Marcus Citation2002).

3. Although Philadelphia was widely known to have misreported UCR crime data in 1997–1999, reform efforts spearheaded by Timoney during his tenure as Commissioner appears to have improved the reliability of police crime reporting. We are not aware of major, systemic complaints about Philadelphia Police Department crime reporting from this data period. Further, for such a potential threat to internal validity to provide a plausible counter‐explanation for the impacts we observe here, one would have to presume that whatever department‐level reporting problems existed dramatically intensified just when Operation Safe Streets was launched, and especially in the specific intervention areas. This seems less than plausible because it would require coordinated, systematic collusion among a very large number of officers. .

4. The intervention started 2 days later on May 1. Additional analyses were conducted leaving the week of 4/29 as a pre‐treatment week; results were not markedly different. To test for a lag in benefits or an anticipatory effect of the program, the startup date was allowed to vary for several weeks both before and after the initial startup date, with no enhancement of impact coefficients.

5. Repeated requests made by the Philadelphia Daily News to the Philadelphia Police Department to obtain additional data and extend the observation period were denied.

6. One reviewer of an earlier draft suggested the traditional ARIMA two‐step model specification and estimation is no longer a favored method of analysis, because the two‐step procedure relies too heavily on model specification by the researchers and can be viewed as a form of data mining. To allay these concerns, we conducted additional analyses using the simplest ARIMA model (1,0,0), a simple serial auto‐correlation model. In this way, we were still able to focus on the longitudinal aspect of the data with minimal researcher specification. Results were very similar despite the lack of model specification, with the exception of an observed county‐wide reduction in violent crime and a localized reduction in drug crimes in the adjoining areas.

7. Additional ARIMA models were run in which a time variable was included in lieu of the temperature variable. Monthly dummy variables failed to improve the model’s explanatory power. Although a model with weekly dummy variables did result in a marginally better fit, it failed to significantly alter the coefficients of the independent variables. For parsimony’s sake, we report here only the model using the temperature variables.

8. Chi‐square analyses showed no significant relationships between geocoded status and crime type.

9. The correlation between drug crimes in a blockgroup and the number of Operation Safe Streets intervention sites was .506 (p < .01). This suggests planners were relatively successful at targeting high‐drug‐crime areas for program sites.

10. Using Crimestat’s hierarchical nearest‐neighbors hot‐spot identifier, we set an alpha of .05 and a minimum of 10 drug crimes to categorize a “hot spot.” The center of the hot spot was identified and then buffered within 1/10 of a mile. Only those buffers which did not intersect the Operation Safe Streets buffers were included in the comparison‐site analyses.

11. Although the comparison areas were identified using “hot spot” analyses, we also compared differences between the target areas and the comparison areas on demographic data. Using 2000 US Census data, we compared blockgroups in which target areas were nested with blockgroups in which comparison areas were nested. Comparing on demographics such as percent African American, percent female‐headed households, percent single‐parent households with children, household median income, household median value, percent owner‐occupied households, percent of the population below the poverty line, and percent population 16 years and up unemployed, the target and comparison areas only differed significantly on percent owner‐occupied households. Target blockgroups reported a higher percentage of owner‐occupied households.

12. Since the outcomes in question are based on count data, concern was raised about whether the residual variance was constant across the time period, an assumption made by the ARIMA analyses. We performed a series of analyses in which the 139‐week period was broken down into three (46–47 weeks), four (34–35 weeks), and finally five (27–28 weeks) time periods. An ANOVA for each of the breakdown arrangements tested the homogeneity of the variance. For both the three and four time period frameworks, there was no significant variation. Only for the five time periods, and only for drug crime, was there a significant difference in the homogeneity of the variance. We concluded from these analyses that heteroskedasticity was not a pervasive problem. We also tried a number of models estimating a 9/11 effect, with both a steady and a declining impact. The parameters were not significant.

13. The average week in Philadelphia resulted in approximately seven homicides. We are unable to perform any meaningfully localized analyses with these few cases.

14. As with the violent crime, the crimes are reported per km² per week.