Modeling rear-end crash risk by injury severity using evasive action-based conflicts: a random parameters bivariate extreme value model

Conflict-based safety assessments have increasingly incorporated unobserved heterogeneity in crash risk estimation; however, the explicit integration of evasive actions into rear-end crash risk estimation remains limited. While ignoring unobserved heterogeneity can bias parameter estimates, overlooking evasive maneuvers may undermine the behavioral validity of conflict-based safety assessments. This study proposes a non-stationary bivariate random parameters extreme value model that integrates conflict measures conditioned on evasive actions to estimate rear-end crash frequencies by severity at signalized intersections. Video data were collected from five signalized intersections in Queensland, Australia. The proposed model was developed using Modified Time-to-Collision and Deceleration Rate to Avoid a Crash as crash proximity measures and the expected post-collision velocity difference (Delta-V) as a severity measure. The random parameters bivariate extreme value model has been found to outperform the fixed parameters model in terms of goodness-of-fit statistics and out-of-sample predictions. The results indicate that the rear-end crash risk is positively associated with the frequency of rear-end conflicts per signal cycle, while the proportion of right-turning vehicles influences both crash risk and injury severity. Models developed using evasive-action-based conflicts produced crash frequency estimates that more closely matched the observed total and non-severe crash frequencies than those developed using all conflicts. The proposed framework enhances the behavioral realism of conflict-based safety evaluations and supports the development of time-sensitive countermeasures at signalized intersections.