Severity analysis of property damage in highway accidents

Accident data collected from 2014 to 2019 for the Dongguan section of the Guangshen Yanjiang Expressway in China were utilized to investigate the key factors influencing the severity of road property damage in highway traffic accidents. The spatial correlation among adjacent accidents was addressed using a spatial generalized ordered Probit model, which employed varying association distance thresholds. An XGBoost machine learning algorithm was developed to estimate the model parameters, and the SHAP (SHapley Additive exPlanations) method was employed to elucidate the model outputs. The results show that significant spatial correlations are present within the accident data. The spatial generalized ordered Probit model demonstrated superior performance compared to the conventional generalized ordered Probit model, with the model based on a 200 m association distance threshold yielding the best results. The SHAP method significantly enhanced the interpretability of the XGBoost machine learning model. Parameter estimation revealed that variables such as single-vehicle accidents, passenger cars, lorries, heavy tractors, nighttime occurrences, early morning periods, cloudy conditions, rainy conditions, and bridge locations were significantly associated with the severity of road property damage resulting from traffic accidents.