Kernel Density Bandwidth Specification in Neighborhood Violence Prevention Research

Abstract

Place-based interventions may reduce violence, but approaches for capturing nearby incidents using kernel density estimation (KDE) vary. KDE smooths geospatial point data, like crime incidents, using a user-specified bandwidth often selected through data-driven approaches that rely on the underlying point pattern. Because point patterns vary by outcome, time, and context, data-driven methods can produce bandwidth sizes that are misaligned with the spatial extent of a place-based intervention, potentially limiting the ability to detect its effect. To illustrate the inferential challenges associated with data-driven bandwidth selection approaches, this study aimed to (1) quantify variability in bandwidths selected through data-driven methods and (2) examine the impact of bandwidth size on simulated intervention effects. We used violent crime data for Philadelphia (2013-2023). For Aim 1, we calculated bandwidth sizes for each crime-year combination using two default data-driven selection criteria and compared selected sizes across crime types and years. For Aim 2, we used a hypothetical place-based intervention with a known effect (30% reduction in nearby assaults) and ran simulations to examine how the intervention effect, estimated using Poisson regression, changed based on the bandwidth size used to estimate the crime density surface. Bandwidth sizes varied significantly by data-driven selection method, crime type, and year (range: 45.9-48,450 ft). For the simulated intervention, "true effects" (i.e., the reduction of nearby assaults attributed to the intervention) were only detectable at bandwidths between 200 and 2900 ft. Larger bandwidths resulted in estimates that incorrectly suggested the intervention was ineffective or increased crime. Data-driven bandwidth selection can obscure or distort intervention effects. Researchers should be critical and transparent when selecting KDE parameters in place-based violence prevention research.

Keywords: Kernel density estimation; Neighborhood violence prevention; Place-based epidemiology.