A Novel Cluster-Based Multinomial Logit Modeling for Crash Severity Prediction in Collisions Involving Automated EV-Only Manufacturer (AEVOM) Vehicles

The increasing prevalence of Automated Electric Vehicles-Only Manufacturer (AEVOM) vehicles underscores the need for better understanding of crash severity under partial automation. Utilizing police-reported crash data from Texas between 2017 and 2024, this study applies a two-stage analytical framework to capture heterogeneity in crash outcomes. Cluster Correspondence Analysis (CCA) is first used to classify crashes into four distinct typologies: high-speed highway collisions, intersection-related events, low-speed impacts with fixed objects, and parked-vehicle crashes in non-trafficway environments. Variable selection and clustering validity are supported by XGBoost (Extreme Gradient Boosting) feature importance metrics and the Creemer’s V statistic. Within each cluster, Random Parameter Logit (RPL) and RPL with Heterogeneity in Means (RPLHM) models are estimated to account for unobserved heterogeneity in crash severity determinants. The analysis reveals that key variables such as lighting conditions, road classification, driver age, seatbelt use, and vehicle type influence severity outcomes differently across clusters. Notably, severe injuries are observed even in low-speed or seemingly controlled environments, highlighting functional limitations in current AEVOM vehicles automation systems. This framework improves model fit and interpretability relative to aggregate models and provides actionable insights for the advancement of advanced driver-assistance systems, infrastructure design, and policy strategies aimed at enhancing the safety of semi-AEVs.