XGBoost-SHAP and Unobserved Heterogeneity Modelling of Temporal Multivehicle Truck-Involved Crash Severity Patterns

This paper aims to address the critical issue of multivehicle truck crashes in developing regions, with a focus on Thailand, by analyzing the factors that influence injury severity and comparing the effectiveness of predictive models. Utilizing advanced random parameters and the XGBoost machine learning algorithm, we conducted a comprehensive analysis of injury severity factors in multivehicle truck-involved accidents, contrasting weekdays and weekends. Our findings reveal that the XGBoost model significantly outperforms the heterogeneous logit model in predicting crash severity outcomes, demonstrating superior accuracy, sensitivity, specificity, precision, F1 score, and area under the curve (AUC) in both model training and testing phases. Key risk factors identified include motorcycle involvement, head-on collisions, and crashes occurring during late night/early morning hours, with environmental elements like road lane numbers and weekend hours also playing a significant role. The study introduces XGBoost as a novel and improved method for truck safety analysis, capable of capturing the complex interactions within multivehicle crash data and offering actionable insights for targeted interventions to reduce crash severity. By highlighting specific risk factors and the effectiveness of XGBoost, this research contributes to the development of data-driven strategies for enhancing truck safety in developing countries.