Crash-type heterogeneity and cross-year transfer stability in injury severity modeling.
Abstract
Objectives
To compare crash-type differences in injury severity factors across car-car (CC), motorcycle/electric bicycle-vehicle (MR), and pedestrian-vehicle (PV) crashes, examine whether risk-factor associations vary across years, and evaluate later-year predictive performance and probability calibration.
Methods
Using 5,185 crashes from the China National Automobile Accident In-Depth Investigation System (NAIS) for 2018-2022, this study first tested heterogeneity across the three crash types and assessed temporal variation using predictor-by-year tests with Benjamini-Hochberg adjustment. Fixed-parameter Logit models were then compared with eligible random-parameter models. Later-year performance was evaluated using a strict rolling-origin approach in terms of discrimination, probability error, calibration, and high-risk-set overlap.
Results
Significant heterogeneity was observed across crash types, supporting separate modeling. The 54 estimable predictor-by-year blocks all had BH-adjusted q values at or above the prespecified q < 0.05 criterion, and the q < 0.10 sensitivity rule yielded the same classification. Fixed-parameter Logit models were selected for CC, MR, and PV. The largest positive within-type risk differences (RDs) were observed for Older Drivers in CC (RD = 0.306), the Non-Small-Vehicle Category in MR (RD = 0.375), and High Pre-Crash Speed in PV (RD = 0.381). Under strict rolling-origin evaluation, AUC values were 0.737, 0.641, and 0.638 for CC, MR, and PV, respectively, with corresponding calibration slopes of 0.581, 0.875, and 0.548. PV had the highest log loss, whereas CC had the highest expected calibration error (ECE), indicating that performance patterns differed across evaluation metrics.
Conclusions
The major risk factors and their relative importance differed across crash types, supporting crash-type-specific interpretation of injury severity. The predictor-by-year screening produced BH-adjusted q values above the prespecified threshold for all estimable blocks, while later-year discrimination, calibration, and high-risk case identification varied across models. Cross-year model performance should therefore be evaluated using probability calibration and high-risk-set stability in addition to AUC. The findings characterize the analyzed NAIS sample and provide model-conditioned evidence for crash-type-specific risk assessment.