Capturing and Exploring Unobserved Heterogeneity in Traffic Crash Injuries in China: A Deep Learning Approach

Abstract

Traffic crash injuries have often been considered a leading threat to human lives. To alleviate this problem, crash data are used to explore various factors of injuries which resulted in accidents with both statistical models and deep learning models. Unobserved heterogeneity, as a result of the complexity of crashes, is inevitable in injury severity analysis. However, it is overlooked in most deep learning models. In this study, TabVAE, a deep learning model which accounts for unobserved heterogeneity, is proposed to analyze the severity of injuries. Data from a total of 3,445 crashes in China, involving 5,384 passengers, are collected and analyzed. The experimental results suggest that the proposed TabVAE model can better capture unobserved heterogeneity. Model comparison demonstrates that TabVAE outperforms classical statistical methods and deep learning models. Additionally, the proportion of unobserved heterogeneity within data and the importance of explanatory variables are identified. The results show that energy dissipation is the key factor to the injury of passengers, and that vehicle damage serves as the most important indicator of energy dissipation. Furthermore, Advanced Driver Assistant System contributes significantly to the unobserved heterogeneity of crashes since it alters the way of energy dissipation. Therefore, addressing unobserved heterogeneity caused by ADAS systems is essential for future accident analysis works.