ABSTRACT
The present research examines the long-term, bidirectional relationships between calls for service, crime, and two police patrol strategies in Santa Monica, California: foot patrol and police stops. Using nine years of monthly data (2006–2014), we estimate two sets of block-level, longitudinal models to tease apart these relationships. In our first set of models, we use police actions and calls for service in the preceding month(s) to predict crime in the subsequent month. In our second set of models, we use calls for service and crime in the preceding month(s) to predict police actions in the subsequent month. We find that while changes in calls for service and crime often precede changes in police action, changes in crime also tend to follow them. For example, police stops appear to be particularly receptive to burglary: blocks with more burglaries receive greater numbers of police stops, and blocks with more police stops have reduced odds of experiencing burglary. We also find that the length of effects of predictors varies as a function of predictor and outcome: whereas some predictors exhibit short temporal effects (e.g. one month), other predictors exhibit much longer temporal effects (e.g. twelve months). Our results thus provide important insight into the spatial and temporal relationships between police actions and police incidents. Police actions must be neatly tailored to police incidents at precise levels if long-term deterrent effects at these levels are to be achieved.
KEYWORDS:
Disclosure statement
No potential conflict of interest was reported by the authors.
Notes
1 Note that in this context, ‘longer’ does not necessarily imply consistent police activity throughout the entire period of analysis. It is possible that police activity ebbs and flows at a given location over time, and thus using a longer temporal unit of analysis helps to capture blocks that potentially have systematically sporadic activity over an extended period of time.
2 These years of data have been publically released by the SMPD.
3 All spatial buffer computations were performed in Stata. We used the latitude/longitude of the centroid for each block (rather than block polygons), given that their small size relative to the buffer size results in a very small proportion that would be bisected by the buffer boundary. Our looping code: (1) takes the latitude/longitude for a block centroid, (2) computes the distance to all other block centroids in the study area, (3) keeps those within the specified distance, and (4) computes the average value of the variable of interest for these blocks in the buffer weighted by inverse distance to the focal block. The looping code operates similarly for the remaining blocks in the study area.
4 As noted in Appendix 2, we distinguish between similar-appearing events, like ‘Living in a Vehicle’ and ‘Abandoned Vehicle’, as part of our disorder coding process because different events imply different types of public concern. For example, an abandoned vehicle in and of itself is asocial: the vehicle may be left in a block for a variety of reasons (e.g. no insurance, mechanical issues, etc.). In this sense, the call regards a concern about the vehicle, rather than the person associated with the vehicle, and therefore we code it as physical disorder. In contrast, living in a vehicle implies social consequences: the call regards a concern about the person occupying the vehicle, rather than the vehicle's presence in the block itself, and therefore we code it as social disorder.
5 Due to limitations in our data, we were unable to differentiate between onview and citizen-reported disorder events.
6 The correlation between the count and duration (i.e. length of stop) variables for police stops was 0.93, suggesting they provide similar information. We therefore use count variables.
7 The correlation between the count and duration (i.e. length of patrol) variables for foot patrol was 0.87, suggesting they provide similar information. We therefore use count variables.
8 Due to limitations in our data, we were unable to differentiate between onview and citizen-reported crime events.
9 Although we would have preferred to estimate the models with larceny and police stops as negative binomial regression models, or zero-inflated negative binomial models, these models encountered estimation difficulties. We therefore estimated OLS models and log transformed our outcome for the police stops model. Note that as the mean of a count gets large enough the distribution will become normal.
10 We did not include disposition data in our analyses given that (1) approximately half of these data were missing for foot patrol (and approximately 83% of events with data were coded as ‘Checks Okay’) and (2) there was little variance in the disposition data for police stops (approximately 74% of events resulted in an ‘Advisal’ or ‘Citation/Other Enforcement’).