Modeling injury severity in intercity bus crashes in a developing country: analyzing contributing factors and heterogeneity using statistical and machine learning approaches

Intercity bus crashes are among the crashes associated with a relatively high number of fatalities. Investigating the factors influencing injury severity in intercity bus crashes can support more precise policymaking to enhance road safety and reduce crash severity. The primary objective of this study is to identify factors affecting injury severity in such crashes. Additionally, recognizing heterogeneity in injury severity outcomes, which arises from diverse contributing factors, enables a more nuanced assessment of crash dynamics. Thus, a secondary goal is to examine the systematic and random heterogeneity in injury severity. Using data from 3,030 intercity bus crashes, this study employs six analytical approaches: the Random Parameter Ordered Logit (RPOL), Random Parameter Multinomial Logit (RPML), Support Vector Machine (SVM), Convolutional Neural Network (CNN), eXtreme Gradient Boosting (XGBoost), and Random Forest (RF) models. These models were used to analyze factors influencing injury severity and assess heterogeneity. Key findings from the RPOL model indicate that the following factors increase the likelihood of severe injuries: run-off collisions, driver drowsiness, the involvement of motorcycles, bicycles, or pedestrians, head-on collisions, bus rollovers, crashes occurring on holidays, midnight crashes, winter and autumn crashes, and intersection-related crashes. Furthermore, injury severity in run-off collisions varies significantly among buses with mechanical defects, highlighting systematic heterogeneity due to interactions between these variables. Moreover, model comparisons revealed that the XGBoost, RF, SVM, and CNN outperform RPOL and RPML in predictive accuracy for injury severity assessment. Finally, some recommendations were made to improve the safety of intercity buses.