Crash risk estimation in freeway exit influence areas considering exit configurations: An interpretability-informed Bayesian hierarchical extreme value model

Abstract

Freeway exit influence areas concentrate mandatory diverging maneuvers and exhibit substantially higher crash risk than basic freeway segments. Although direct-type and parallel-type exits differ in their diverging characteristics, most existing studies treat them as homogeneous objects, leaving their crash risk and influencing mechanisms insufficiently understood. This study proposes an interpretability-informed Bayesian hierarchical extreme value framework for crash risk estimation across different exit configurations. A Quasi-Kinetic Energy Risk Indicator (QKERI) is first developed to jointly characterize potential collision damage and the spatiotemporal proximity of interacting vehicles. Quantile regression is then used to determine the threshold for extreme conflicts, while the Conditional Shapley Value Index (CSVI) provides an interpretable basis for covariate selection. Finally, a peaks-over-threshold generalized Pareto model is estimated within a Bayesian hierarchical framework. The proposed approach is evaluated using 3027 conflict events extracted from more than 1200 min of trajectory data collected at 15 consecutive exit influence areas along a 200 km freeway. The estimated annual daytime crash frequencies are evaluated against three years of observed daytime crash records using Poisson confidence intervals, MAE, RMSE, and MAPE. The results indicate that the threshold for extreme conflicts is primarily determined by short-term traffic stream characteristics, whereas crash risk is mainly associated with adjacent lane density difference, heavy vehicle proportion, and conflict location. Among these factors, heavy vehicle proportion remains statistically significant throughout the modeling process. And the three scale covariates retain consistent coefficient signs across all 15 leave-one-site-out refits. Further discussion finds consistency between the CSVI results and the extreme value estimation, and the effective diversion length of a site is more decisive than the exit configuration in explaining differences in crash risk. These findings provide insights for the geometric design and proactive safety management of freeway exit areas.