Evaluating Spatial-Proximity Structures in Crash Prediction Models at the Level of Traffic Analysis Zones

Abstract

Zonal crash prediction that considers cross-zonal spatial correlation is a frontline research topic, especially in the context of transportation safety planning. This study presents an evaluation of crash prediction models at the level of traffic analysis zones with four types of spatial-proximity structures: 0–1 first-order adjacency, common-boundary length, geometry-centroid distance, and crash-weighted centroid distance. Bayesian spatial analysis with conditional autoregressive priors was successfully applied, and Hillsborough data were used to compare the model-fitting and predictive performance associated with the proposed models. The results confirmed the extensive existence of cross-zonal spatial correlation in crash occurrence. The best predictive capability, relatively, was associated with the model that considered proximity of neighboring zones by weighing their common-boundary lengths. Furthermore, full consideration of all possible spatial correlations for all zones significantly increased model complexity, which probably resulted in reduced predictive performance.