Integrating machine learning into path analysis for quantifying behavioral pathways in bicycle-motor vehicle crashes.

The behavioral pathways in traffic crashes describe the chained linkages among contributing factors, pre-crash road user behaviors, and crash outcomes. Bicyclists are more vulnerable than motorists on road and their pre-crash behaviors play an essential role in the pathways leading to injuries. The objective of this study is to develop a methodological framework that integrates machine learning with path analysis to quantify behavioral pathways in bicycle-motor vehicle crashes. Specifically, two sets of models are developed for predicting: 1) pre-crash behaviors given contributing factors and 2) bicyclist injury severity given contributing factors including pre-crash behaviors. The path analysis chains machine learning models to establish the indirect linkages between contributing factors and injury severities through correlates of pre-crash behaviors. This study explored five machine learning methods, including Random Forest (RF), Categorical Naive Bayes (CNB), Support vector machine (SVM), AdaBoost (Boost), and Neural network (NN). To reduce the bias of any single model, this study proposes a technique to combine model estimates by averaging marginal effects. This study used a dataset containing 9,296 bicycle-motor vehicle crashes to demonstrate the application of the framework. Across five machine learning models, the signs of marginal effects generally agree but their magnitudes vary substantially. The pre-crash behavior of "bicyclist failed to yield" increases bicyclist injury severity by 1.11%. The path analysis results highlighted contributing factors related to risky pre-crash behaviors that lead to severe injuries, such as bicyclist intoxication. The framework is expected to support agencies' decision-making to improve cycling safety by reducing unsafe behaviors on roads.