The Impact of Geocoding Uncertainty on Crime Hotspot Prediction

Geocoding uncertainty in police incident records can distort crime hotspot prediction. Using theft incidents in Washington, D.C., this study quantifies how three common preprocessing strategies—record deletion, centroid aggregation, and random perturbation—affect kernel density estimation (KDE) hotspot prediction under varying data quality and spatial supports. Monte Carlo simulations introduce uncertainty into 5%–50% of training incidents and apply each strategy across six preprocessing units (street, tract, neighborhood, and 100/300/500 m grids) under long‐ and short‐horizon rolling‐window validation. Performance is assessed with the Prediction Accuracy Index (PAI) and a pixel‐wise F1 score. Deletion is scale‐invariant but lowers performance through information loss. Relocation‐based strategies perform well under fine supports but deteriorate as preprocessing units become coarser relative to the KDE bandwidth. Sensitivity tests using 100/300/500 m bandwidths confirm these patterns, with marked degradation around 𝜔≈30% in the present setting.