Re-evaluating the Use of Near-Crashes as Surrogates for Crashes

Abstract

This study aimed to evaluate whether near-crashes, stratified by severity, can reliably serve as surrogates for crashes. Using a subset of 2,000 near-crashes from the Second Strategic Highway Research Program Naturalistic Driving Study (SHRP 2 NDS) dataset, manual severity ratings were applied following a structured protocol. Existing four-level crash severity classifications were used to provide a comparative basis for evaluating the rated near-crashes, while baseline events served as a reference to assess whether each severity level more closely resembled normal driving conditions or safety-critical events. The study further assessed the predictive power of key variables using random forest models and compared near-crashes and crashes to identify similarities and differences in event characteristics, contributing factors, and predictor effects. Results showed that near-crashes offer a practical and validated alternative for studying crash risk in naturalistic driving datasets. By combining rigorous manual rating, machine learning validation, and comparative analysis with crash events, this study demonstrated that near-crashes can effectively inform safety modeling specifically for high-severity and moderate-severity events. Future research should focus on improving surrogate validity for extreme crashes, incorporating additional predictors, and leveraging automated video and sensor data for scalable severity assessment. Effective use of near-crash data can accelerate identification of risk factors, support intervention design, and ultimately reduce crash-related injuries and fatalities.