Macroscopic Models for Accident Prediction at Railroad Grade Crossings

Accident prediction and ranking of high-accident locations at railroad grade crossings is often performed with the U.S. Department of Transportation (DOT) accident prediction formula, as described in the FHWA Railroad–Highway Grade Crossing Handbook. However, the current version of the model was developed in the 1980s, and all model coefficients remain unchanged except for a normalizing constant that FRA updates every few years to reflect recent nationwide accident trends. This paper presents accident prediction models for the same warning device categories defined in the U.S. DOT model, but it uses a zero-inflated negative binomial form. Ten years of data from the state of Illinois were used to compare the accuracy of one of the models with the U.S. DOT accident prediction formula. Five years of data were used to create the model and estimate predictions, and the following 5 years were used to evaluate the predictions. Prediction accuracy is measured in terms of the cumulative accident frequency and the accuracy for ranking high-accident locations. Results highlight advantages of a model built with recent data to predict the overall accident trends and the absolute accident frequencies, as well as the benefits that the U.S. DOT prediction formula still may provide for ranking high-accident locations.