Heterogeneity in crash patterns of autonomous vehicles: The latent class analysis coupled with multinomial logit model

Understanding the heterogeneity in autonomous vehicle (AV) crash patterns is crucial for enhancing the safety and public acceptance of autonomous transportation systems. In this paper, 584 AV collision reports from the California Department of Motor Vehicles (CA DMV) were first extracted and augmented by a highly automatic and fast variable extraction framework. Crash damage severities, classified as none, minor, moderate, and major, were set as the dependent variables. Factors including crash, road, temporal, vehicle, and environment characteristics were identified as potential determinants. To account for the heterogeneity inherent in crash data and identify key factors influencing the damage severity in AV crashes, a methodology integrating the latent class analysis and multinomial logit model was employed. Two heterogeneous clusters were determined based on the skewed distributions of vehicle status and driving mode. The model estimation results indicate a positive association between severe crash damage and some risk factors, such as head-on, intersection, multiple vehicles, dark with street lights, dark without street lights, and early morning. This study also reveals significant differences among the variables influencing the damage severity across two distinct subclasses. Moreover, partitioning the AV crash dataset into heterogeneous subsets facilitates the identification of critical factors that remain obscured when the dataset is analyzed as a whole, such as the evening indicator. This paper not only enhances our understanding of AV crash patterns but also paves the way for safer AV technology.