Spatial Effects on Zone-Level Collision Prediction Models

Abstract

A recent study developed a set of zone-level negative binomial collision prediction models to investigate the relationship between various transportation and sociodemographic characteristics and overall roadway safety. The developed models used data from Metro Vancouver, British Columbia, Canada, and considered the effect of Poisson variations and heterogeneity (extra variation) on collision occurrence. This study aims to evaluate spatial effects on the occurrence of collisions and to check whether the inclusion of spatial variables can improve the goodness of fit and inference capability of those previously developed prediction models. Transit-reliant and application-based collision prediction models with spatial correlations were developed by using the WinBUGS software. The convergence of the developed models was tested by trace plots of the parameter estimated, the Brooks–Gelman–Rubin statistics, and ratios of Monte Carlo errors relative to the standard deviations of the estimates. The results showed that incorporation of the spatial correlations affected the parameter estimates, the values of dispersion parameters and intercepts, and also the t-statistics. The effect of the main exposure variable on all of the models for total, severe, and property-damage-only collisions was found to be smaller under spatial models. The smaller values of the exponents of the main exposure variable confirmed the assumption that spatial effects need to be considered in collision prediction models to mitigate any potential bias associated with model misspecification.