Kernel Space-Time Interaction Tests for Identifying Leading Indicators of Crime

• Activity Number: 214
• Type: Topic Contributed
• Date/Time: Monday, August 4, 2014 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Seth Flaxman*+ and Daniel Neill and Alex Smola
• Companies: and Carnegie Mellon and Carnegie Mellon
• Keywords: spatial ; spatiotemporal ; criminology ; kernel methods ; machine learning ; independence testing

Abstract

The CityScan software (Neill, 2013) can accurately predict geographic hot-spots of violent crime by detecting emerging clusters of crimes. CityScan has been used for targeted policing by the Chicago Police Department, but its current use relies on monitoring a manually chosen set of leading indicators. To better identify the most predictive leading indicators, we use a geocoded, time-stamped dataset from Chicago of ~9 million calls to 911 between 2007 and 2010, and search for space-time associations with shootings. Correlations can be confounded in the spatio-temporal setting by underlying spatial effects (e.g., "bad" neighborhoods) and temporal effects (e.g., more crimes in the summer). To address this issue, we propose a new kernel-based statistical test for space-time interaction. Like the classical Knox and Mantel tests, our test explicitly controls for separable spatial and temporal dependencies, but unlike the classical tests it can capture non-linear dependencies. We demonstrate its applicability to questions in criminology and predictive policing by using it to test for space-time interactions between disorder and crime and between 911 call types and shootings/homicides.