Do Simulated Traffic Conflicts Predict Crashes? An Investigation Using the Extreme Value Approach*

There has been significant recent interest in using microscopic traffic simulation in road safety analysis. However, whether the simulated traffic conflicts can predict crashes reliably is still an open question. This study aims at investigating the prediction validity of simulated traffic conflicts using the extreme value approach. The traffic simulation models were developed based on real-world data from four approaches of a signalized intersection and calibrated using a two-step calibration procedure that aimed at enhancing the correlation between simulated conflicts and actual field-measured conflicts. Extreme value models were developed for field conflicts and simulated conflicts of three scenarios (i.e., default, first-step calibration and second-step calibration) separately, and then the model estimated crashes were compared to observed crashes for the purpose of performance evaluation. The results show that after the second-step calibration, a high correlation between simulated conflicts and field conflicts is achieved, but the estimated crashes from these simulated conflicts are not as good as field conflicts and are with systematic underestimation. Nevertheless, each effort of simulation model calibration has shown progressive improvement in terms of crash estimation accuracy and similarity of extreme value distributions between simulated conflicts and field conflicts. This finding implies the possibility of predicting crashes reliably from simulated conflicts with proper calibration.