Addressing data limitations in autonomous vehicles crash severity analysis: A hybrid framework for small samples and sample imbalance.
OBJECTIVE This study aimed to address key data limitations in autonomous vehicle (AV) crash-severity analysis, including small samples and sample imbalance, and to identify interpretable risk factors associated with injury outcomes in AV crashes. METHODS A verified dataset of 2,946 AV crash events from 2015 to 2024 was constructed, with crash records spatially matched to variables related to road geometry, traffic control, roadway attributes, and the built environment. A hybrid analytical framework was developed by combining sample-balancing methods, feature-selection techniques, and random-parameter logit modeling. Three balancing strategies, ROSE, SMOTE, and ROSE+SMOTE, were compared with three feature-selection methods, mutual information (MI), random forest (RF), and XGBoost, under both balancing-first and feature-selection-first workflows. Nineteen model specifications were evaluated using stratified five-fold cross-validation. Further causal analysis of AV crash severity was conducted based on the parameter estimates and marginal effects of the random-parameter logit model. RESULTS All hybrid specifications improved performance relative to the unprocessed baseline. The combined ROSE+SMOTE strategy produced stronger performance than single balancing methods. Model 18, which applied ROSE+SMOTE before XGBoost feature selection, achieved the highest Macro-Recall, indicating strong balanced sensitivity across severity classes. Model 19, which applied XGBoost feature selection before ROSE+SMOTE, achieved the best overall and safety-oriented performance, with the highest Accuracy, Macro-Precision, Injury-class Recall, Macro-F1, ROC-AUC, PR-AUC, the lowest FNR-injury, and the lowest AIC. Significant injury-risk factors included side-impact crashes, restaurant density, school and metro stop, commercial and mixed-use/public land use, lane-markings-only, nighttime, rush hour, crosswalks, T/Y-intersections, expressways and arterials, intersection-related locations, and crash lanes ≤ 2. Greater road width was associated with a lower injury probability. CONCLUSIONS The proposed hybrid framework provides a stable and interpretable approach for analyzing small and imbalanced AV crash datasets, and offers evidence to support scenario-based ODD testing, targeted infrastructure improvement, curbside management, and context-sensitive AV safety governance.