Hot-Spot-Based Predictive Policing in Pittsburgh: A Controlled Field Experiment
Hot-spot-based policing programs aim to deter crime through increased patrols to high-crime street corners or city blocks, but optimal strategies for identifying hot spots and the potential crime reductions from proactive patrols to these areas are not well understood. We present findings from an empirical comparison of crime forecasting methods and a controlled experiment evaluating a hot-spot-based predictive policing program in Pittsburgh, PA. We compared the performance of several place-based forecasting models on predicting historical crime data and selected those that demonstrated high predictive accuracy and spatial dispersion of forecasted areas in Pittsburgh. Over 16 months, weekly hot spots were selected across all police zones in Pittsburgh, and hot spots were targeted with additional patrols (on foot and in patrol cars) by the Pittsburgh Bureau of Police (PBP). Areas exposed to treatment (increased patrols in the predicted hot spots) changed on a weekly basis, allowing for a statistical comparison of crime counts in treated and control areas across the city. We find statistically and practically significant reductions in crime counts within predicted hot spots, with observed reductions of 17.2% in Part 1 violent (P1V) crimes and 11.0% in Part 1 property (P1P) crimes in patrolled areas. We also examine potential displacement of crime to locations (1) adjacent to hot spots and (2) not adjacent to hot spots, and find no evidence of crime displacement to those areas resulting from increased patrols to predicted hot spots.
Committee: Daniel Neill (New York University), Wilpen Gorr, Amelia Haviland, George Chen
Copy of DAP Paper
For More Information, Please Contact: