With the use of a general classification regression model, this study investigated the causal relationship between time to train arrival (TTA) and crash frequency at highway-railroad grade crossings. In particular, a stratified structure in the explanatory variables was used to avoid the collinearity problem generally confronted in linear regression models. TTA is a good estimate of rail sight distance and time to collision, and it could be used to predict crash frequency at a grade crossing. A 14-year crash data set accompanied by crossing inventory data including TTAs was collected for the empirical study. Study results indicated that a negative relationship between TTAs and crash frequencies was generally found for all types of trains. Similar causal relationships were also found in various combinations of both crossing attributes and crash characteristics. Sensitivity analysis on the variable combinations was also conducted to investigate the key risk factors that might result in traffic collisions at grade crossings. Policy implications based on the empirical study are discussed, and future research directions are recommended.