Assessing the interdependence of rider fault-status and injury severity in motorcycle rear-end crashes: insights from bivariate probit and XGBoost-SHAP models.

Abstract

This study examines the interdependent relationship between fault status and injury severity in motorcycle rear-end crashes in Thailand using data from 1,549 crashes (2011-2015) integrated from the Department of Highway's Accident Information Management System and Traffic Information Movement System. This article employs a bivariate probit model alongside various boosting techniques for simultaneous estimation of injury severity and at-fault status. Among the tested models (AdaBoost, CatBoost and LightGBM), both the bivariate probit and XGBoost-Endogenous models demonstrate superior performance in accuracy and F1-score. The bivariate probit model reveals that injury severity is significantly influenced by rider characteristics (age, gender), road features, and traffic conditions. Riders under 55 years old, female riders and those on roads with depressed medians or higher traffic volume show lower injury severity risk. Conversely, drunk riding, nighttime crashes on unlit roads, and higher truck traffic percentages increase severe injury likelihood. The XGBoost model corroborates these findings, identifying traffic volume, truck percentage and nighttime conditions on unlit roads as the most crucial predictors of injury severity. Regarding fault status, younger riders and those using safety equipment show a higher probability of being at-fault. This novel analytical approach provides valuable insights for motorcycle safety policy development and future research directions.