关键词: 运营车辆安全, 严重程度致因分析, 不可观测异质性, 随机参数Logit模型, SHAP可解释分析

Abstract:

With the rapid development of the road transport industry, the number of accidents involving operational vehicles has been continuously increasing, and the severity of these accidents has become a widespread concern. Especially, operational vehicles, due to their unique operational characteristics, face more complex traffic environments and risks. This study analyzes the impact of "unobserved heterogeneity" on the severity of operational vehicle accidents. Based on traffic accident data for operational buses and trucks in China, relevant variables are selected from four aspects: driver behavior, vehicle type, road characteristics, and environmental conditions, and a random parameters Logit model is constructed. By introducing random parameters, the model can effectively capture heterogeneity and uncertainty between individuals, improving its explanatory power and predictive performance. The SHAP method is further applied to analyze the direction, importance, and non-linear interactions between variables. The results show that, for operational buses, complex road shapes and vehicle types significantly increase the severity of accidents, especially the interaction effect between complex road conditions and improper operations, which notably raises the accident severity. For operational trucks, the interaction effect between hazardous material transport vehicles and complex road conditions is stronger, and speeding behavior significantly increases the probability of major accidents. The SHAP analysis quantifies the contribution of 10 multidimensional factors to accident severity, revealing that bus accidents are mainly influenced by road environment factors, while truck accidents are more significantly related to vehicle attributes. This further quantifies the differing impacts of human factors and environmental factors on the severity of accidents.

Key words: operational vehicle safety, severity causality analysis, unobserved heterogeneity, random parameter Logit model, SHAP interpretability analysis