Modeling the Accident Prediction for At-Grade Highway-Rail Crossings

Abstract

Since accidents at highway-rail at-grade crossings (HRGCs) are often catastrophic, safety prediction and evaluation at such locations are of great importance. In this paper, at-grade crossing inventory data and historical accident data were obtained from the Federal Railroad Administration (FRA’s) Office of Safety online databases. The HRGC railroad and highway characteristics were selected as the influencing variables. Considering HRGC accidents are over-dispersed count data with excessive zeros, six count data models, including the Poisson model, negative binomial model (NB), zero-inflated Poisson model (ZIP), zero-inflated negative binomial model (ZINB), hurdle Poisson (HP) model and hurdle negative binomial model (HNB) were investigated and developed for accident prediction. The ZINB model outperformed the other five models in terms of the goodness-of-fit, zero inflations, and statistical significance of factors. The most significant contributing factors in the ZINB model included the maximum timetable speed of train, exposure-related variables such as total through trains, highway traffic volume, rural or urban area, and the type of control devices at HRGCs, followed by the minimum speed of train, highway paved or not, and the number of traffic lanes. This study could assist decision-makers with a more robust safety evaluation at highway-rail at-grade crossings.