Probabilistic classification of hazardous materials release events in train incidents and cargo tank truck crashes

As two of the main modes of hazmat surface transportation, quantifying conditional probability of release of hazmat from trains in rail incidents and Cargo Tank Trucks (CTTs) in highway crashes is an important component of risk assessment. The objective of this study was identifying computational tools with reliable performance for quantifying probability of hazmat release in train incidents and CTT crashes, based on the type of the decision-making problem. Events of release of hazmat were probabilistically classified by statistical and machine learning methods (Mixed Logistic Regression, Naïve Bayes, Random Forests, and Support Vector Machines) using relevant explanatory variables. The datasets were Federal Railroad Administration rail equipment incident data, and combined Nebraska and Kansas police-reported traffic crash data. Given the rarity of these events, the classification performance of different methods was compared based on precision, recall and area under ROC curves (AUC). Random Forests had the best performance in classifying hazmat release for trains and railcars, based on different criteria. For CTTs, Support Vector Machines and Random Forests had the highest precision, while logistic regression and naïve Bayes performed better based on recall and AUC. The research provides recommendations regarding usage of the methods depending on the purpose of analysis.