Corrigendum to “When police pull back: Neighborhood-level effects of de-policing on violent and property crime, a research note”

Nix, J., Huff, J., Wolfe, S. E., Pyrooz, D. C., & Mourtgos, S. M. (2024). When police pull back: Neighborhood-level effects of de-policing on violent and property crime, a research note. Criminology, 62(1), 156–171. https://doi.org/10.1111/1745-9125.12363

In the published version of this article (Nix et al., 2024), a coding error resulted in incorrect census data and spatial weights being matched to 54 of 78 neighborhoods (these data were used to create Level 2 controls for population, racial composition, immigration, and disadvantage, and a Level 1 control for spatial lags). Crime data was incorrectly matched to 2 of the 78 neighborhoods.

Upon correcting these errors, we identified two additional, albeit minor, mistakes. First, our AQI variable only captured daily levels of carbon monoxide. We have corrected this measure so that it includes particulate matter, NO₂, and ozone, which more closely reflects the way we described the variable in the article. Second, our weather data included Colorado weather stations outside of the City and County of Denver. In this correction, we have excluded those stations.

Our findings are substantively similar upon making these corrections (see Table 1 below). The relationship between pedestrian stop deviations and violent crime remains statistically significant (b = −0.009, SE = 0.002, p < 0.001). The relationship between vehicle stop deviations and violent crime is no longer statistically significant, but the magnitude of the coefficient is strikingly similar (b = −0.00175, SE = 0.001, p = 0.137; previously b = −0.00245, SE = 0.001, p = 0.042). The relationship between drug arrests and property crime is no longer statistically significant (b = −0.027, SE = 0.006, p < 0.001 in the original model, compared to b = −0.012, SE = 0.009, p = 0.181 in the corrected model). Finally, in the corrected model, the relationship between pedestrian stops and property crime is statistically significant (b = −0.005, SE = 0.001, p < 0.01; previously b = 0.002, SE = 0.003, p = 0.51).

One notable difference is that in the corrected analysis, we observe that the relationship between reduced pedestrian stops and property crime was more pronounced in neighborhoods with higher levels of disadvantage (see Table S14 and Figure S4).

We have updated our Harvard Dataverse and replication files to reflect these corrections. In addition to minor changes to the results presented in Table 1 of the article, there are also various small differences in the tables and figures included in the Supplemental Materials. Updated Supplemental Materials are available under the “Supporting Information” tab of the online version of the article.

Below is a revised results section with updated point estimates:

Panel 1 of Table 1 displays the results of four mixed effects models that regressed police discretionary behaviors on variables reflecting the start of the COVID and Floyd periods, respectively (along with controls). The constant captures the neighborhood-week mean deviation in the outcomes pooled across the pre-COVID period, when all control variables are set to zero. Pedestrian stops, vehicle stops, and drug arrests differed statistically in the pre-COVID period, each in a negative direction. Disorder arrests, meanwhile, were indistinguishable from zero. This period also serves as the reference category for the pooled-periods of exogenous shocks. During the COVID period, there was a much greater reduction in police activity than in the pre-COVID period, relative to the prior 4-year weighted average. Police made roughly 3.10 fewer pedestrian stops across neighborhood-weeks (constant −0.809 + COVID −2.295 coefficient), 9.89 fewer vehicle stops, 0.92 fewer drug arrests, and 0.24 fewer disorder arrests each week. These reductions persisted during the Floyd period, when again using the pre-COVID period as a reference category, the police made roughly 3.18 fewer pedestrian stops, 8.87 fewer vehicle stops, 0.94 fewer drug arrests, and 0.19 fewer disorder arrests each week.

These trends in policing are not explained by climate variation or population mobility but reflect shifts that are timed with two exogenous shocks that defined the experiences of citizens and police alike in 2020. The variance components—as reflected in standard deviations—also reveal there was significant variation across neighborhoods in the pre-COVID period, and there was dramatically more variation in the COVID and Floyd periods for pedestrian stops (+169% and +226%), vehicle stops (+14% and +34%), and drug arrests (+76% and +100%), but not disorder arrests. Thus, even while there were wholesale reductions in policing across Denver, neighborhoods were experiencing them very differently.

Panels 2 and 3 provide the results of mixed-effects negative binomial models predicting violent and property crime, respectively. In both, the pre-COVID period continues to serve as the reference category for the COVID and Floyd periods, with the expectation that the inclusion of the policing mediators should attenuate the relationship between these periods and crime. For violent crime, the naïve coefficients (not reported in tabular form) for the period effects are 0.066 (p = 0.248) and 0.326 (p < 0.001) for the COVID and Floyd periods, respectively. Once accounting for between- and within-neighborhood differences, the Floyd coefficient reduces 35% in magnitude, with the COVID period indistinguishable statistically from the pre-COVID period, while there were still 0.212 more violent crimes per neighborhood-week in the Floyd period.

The introduction of discretionary policing behavior mediators revealed a mixed story. Pedestrian and vehicle stops behaved consistent with theoretical expectations. Upward deviations in pedestrian stops from the prior 4-year weighted average were associated with 0.009 (p < 0.001) lower units of violent crimes per neighborhood-week, a 0.9% reduction. The Floyd coefficient was reduced by about 11% in the pedestrian stops model. Vehicle stops, along with both indicators of arrests, were statistically null.

Panel 3 displays the results of negative binomial models predicting property crime. The coefficients for the COVID and Floyd effects were 0.180 (p < 0.001) and 0.428 (p < 0.001), respectively (though not shown, the naïve coefficients were 0.166 and 0.422, respectively). The mediator analysis revealed that pedestrian stops were associated with property crime. A one-unit upward deviation in pedestrian stops corresponded with 0.005-unit reductions in property crime. In contrast, there was no evidence that deviations in vehicle stops, drug arrests, or disorder arrests were associated with variations in property crime across neighborhoods.

Panels 2 and 3 also reveal that the effects of policing were not universal across neighborhoods. The variance components reported at the bottom of each panel revealed significant variation. The standard deviation for the policing coefficients ranged from as little as 0.001 (vehicle stops) to as much as 0.041 (disorder arrests). Stronger effects of de-policing on property crime were found in more disadvantaged neighborhoods in Denver (see Table S14).