Bayesian Spatial Models of Crash Frequency at Highway–Railway Crossings

Abstract

Though rare, crashes at highway–railway crossings usually result in severe injuries and fatalities. Intersections along railway corridors share many conditions that support the use of spatial analyses, but so far the spatial effects are essentially unknown. Little research has been conducted about highway–railway crossings using spatial correlation. The study presented in this paper analyzed spatial correlation structures on crash frequency at railway crossings, particularly conditional autoregressive (CAR) and joint specification, with the use of Gaussian Kriging models. Full Bayesian Poisson–lognormal approaches were used to compare the effects of various models, including heterogeneity-only, spatial-only, and heterogeneity–spatial models. These methods were estimated with crash data from a low-speed passenger train service in Costa Rica. The deviance information criterion was used to compare the models. Heterogeneity–CAR models show a better goodness of fit than heterogeneity-only, joint Kriging, and CAR-only definitions. The second-order neighboring model (CAR-2) yielded the best fit and related to an average neighbor distance of about 700 m (0.44 mi). Robust semi-variograms based on joint Kriging models and CAR methods showed similar results. The proportion of variation in the data explained by spatial correlation methods was similar (29% of the total variation). These findings suggest that spatial correlation at highway–railway crossings should be considered in modeling of crash frequencies.