A New Accident Prediction Model for Highway-Rail Grade Crossings Using the USDOT Formula Variables

Abstract

This paper presents the ZINDOT model, a methodology utilizing a zero-inflated negative binomial model with the variables used in the United States Department of Transportation (USDOT) accident prediction formula, to determine the expected accident count at a highway-rail grade crossing. The model developed contains separate formulas to estimate the crash prediction value depending on the warning device type installed at the crossing: crossings with gates, crossings with flashing lights and no gates, and crossings with crossbucks. The proposed methodology also accounts for the observed accident count at a crossing using the Empirical Bayes method. The ZINDOT model estimates were compared to the USDOT model estimates to rank the crossings based on the expected accident frequency. It is observed that the new model can identify crossings with a greater number of accidents with Gates and Flashing Lights and Crossbucks in both Illinois (data which were used to develop the model) and Texas (data which were used to validate the model). A practitioner already using the USDOT formulae to estimate expected accident count at a crossing could easily use the ZINDOT model as it employs the same variables used in the USDOT formula. This methodology could be used to rank highway-rail grade crossings for resource allocation and safety improvement.