Place-based risk factors for aggravated assault across police divisions in Little Rock, Arkansas

Abstract

Extant literature on crime and place recognizes crime does not occur randomly across space and is influenced by physical place features within the environment, such as alcohol outlets, restaurants, and public transportation stops, that generates and attracts crime. However, many analyses of these features’ influence on crime are conducted across entire jurisdictions which could potentially misrepresent the particular relationships within more localized areas. With a movement in police agencies towards proactive data-driven analyses of crime, the ability to better understand the environmental context of crime events continues to grow in prominence. Building from environmental criminology perspectives and concepts, the current study presents a framework police agencies can implement to more precisely identify the underlying place features influencing where crime occurs in different areas of their broader jurisdiction. Using risk terrain modeling, the current study examined variation in the risk factors for aggravated assault between a citywide model and three police division-specific models. Each police division was considered a unique environmental context that differentially affects crime, which necessitated site-specific analysis within the city to inform tailored police response. Findings are discussed in a policing framework with attention on practitioner implications and future directions of research.

Keywords:

• Policing
• crime and place
• risk terrain modeling
• aggravated assault

Notes

1. Policing in the United States is organized geographically. For example, many police departments divide their jurisdictions into various administrative units, such as sectors, districts, or divisions, as a means to allocate resources. Conceptually, each police division could be considered its own, distinct, environment with a unique set of places where crime is most likely to emerge as a result of the underlying criminogenic features at those locations. Although a strong argument could be made for the use of various areal political or administrative units across which the spatial dynamics of crime could be examined, we use police divisions because Little Rock Police Department (LRPD) organizes their CompStat program and meetings based on the distinct divisions. The goal of the current study is not to directly test theory, but to use the extant literature to guide expectations of how the concepts of crime and place research can be translated into a practical crime analysis method.

2. Aggravated assault incidents were not geocoded to the centerline file because some of the original districts’ boundaries were along streets. These districts were dissolved into a larger division area but the outer boundaries of the districts defined the area. Divisions were comprised of multiple districts. This was done to stay in line with how LRPD approaches their CompStat process by divisions. By using an offset, a better estimate of what division crimes occurred in was considered.

3. The LRPD provided police district shapefiles that were dissolved based on the divisions presented at CompStat meetings. There are differences between the Little Rock shapefile used for the citywide analysis and the three division shapefiles. The division shapefiles are about 119 square miles total while Little Rock is 121 square miles. While we acknowledge this difference, the researchers believed using LRPD files would better reflect real-world application. Little Rock does not have an island; it appears that way on the map because how LRPD separated Little Rock into districts around land features. There is a walking bridge across a section of the Little Maumelle River that falls within Little Rock city boundaries.

4. It is expected that police officers respond to crimes outside of their specific division based on the severity of the crime type being examined but a majority of their time is spent within their division.

5. For example, there are extant studies that do examine both of these aspects simultaneously and demonstrate that place features maintain a significant relationship with crime, even when controlling for various dimensions of community social characteristics (see Drawve, Thomas, and Walker 2016; Groff and Lockwood 2014; Piza et al. 2016).

6. The current study utilized RTMDx because of the diagnostic methodology the software employs. RTMDx, like many other spatial techniques, conducts analyses based on a study area parameter, creating issues with the MAUP (Bailey and Gatrell 1995). This issue can be more troublesome when using multiple administrative boundaries, in the current study police divisions, to determine study areas.

7. During the process of getting this study published, the city of Little Rock developed a data portal (data.littlerock.org/). This shows a progression of the city identifying a need to make data more readily available. Hopefully, future studies of Little Rock will contain a larger pool of potential risk factors/data to account for multiple elements of the environmental backcloth.

8. Liquor stores were a separate category in the treasury data for 2013 (off-site).

9. Calculating the average nearest neighbor allows the user to determine if density is an appropriate operationalization. If the observed distance is greater than the maximum influence of the model, a density operationalization is less meaningful. Testing such features as proximity only shortens the model run time by limiting the amount of variables created and tested.

10. There were a total of 192 variables created based on how each place feature was operationalized in RTMDx. With such a large number of variables, there is the possibility of spurious relationships. RTMDx runs a penalized Poisson regression model to identify all variables having non-zero coefficients (see Heffner 2013). This step results in a smaller number of variables that are then used in a stepwise regression to construct a ‘best’ model based on the Bayesian Information Criterion (BIC). The reduced set of variables often contains the same feature operationalized differently. For example, there could be two different spatial influences that have nonzero coefficients, but RTMDx tests to find the most appropriate one. That is achieved through two-step process of significance testing. Two separate stepwise regressions run, one assuming a Poisson distribution and a second assuming a negative binominal distribution. The BIC score of each model is used to determine the ‘best’ model. An improved BIC score is a lesser value when compared to another model.

11. Relative Risk Values (RRV) are calculated by exponentiating each significant risk factor’s coefficient.