Exploring Risk Factors for Crash Severity During Thailand’s Holiday Travel: Machine Learning Exploration Compared to Heterogeneity Modeling

Thailand’s Western New Year and Songkran festivals witness a surge in traffic crashes due to increased travel volume. This study explores risk factors influencing crash injury severity during these holidays (2017–2019). Crash data is analyzed using both the Random Parameters Ordered Logit Model with Means Heterogeneity (RPOLHM) and machine learning techniques (Multilayer Perceptron (MLP), Adaptive Boosting (AdaBoost), Extreme Gradient Boosting (XGBoost), Random Forest). Crash severity is categorized as property damage only (PDO), minor injury, and severe/fatal injury. The results reveal that factors like motorcycle or pedestrian involvement, adverse weather, speeding, drunk driving, fatigue, nighttime conditions, improper overtaking, and urban location all significantly increase the risk of severe/fatal crashes. Notably, the XGBoost model outperforms both RPOLHM and other machine learning methods, achieving a validation accuracy of 83.8%. While machine learning approaches demonstrate superior predictive capability, RPOLHM provides interpretable coefficient estimates and marginal effects essential for understanding causal mechanisms. This complementarity suggests that concurrent application of both paradigms offers comprehensive insights: machine learning for prediction-oriented objectives and econometric models for policy formulation. These findings provide valuable guidance for policymakers, highway engineers, and researchers to develop targeted road safety interventions during these high-risk periods.