Modeling temporal correlation and heterogeneity in real-time conflict rates using Bayesian Tobit models for signalized intersections.

The use of real-time traffic conflicts for safety studies provide more insight into how important dynamic signal cycle-related characteristics can affect intersection safety. However, such short-time window for data collection raises a critical issue that the observed conflicts are temporally correlated. As well, there is likely unobserved heterogeneity across different sites that exist in conflict data. The objective of this study is to develop real-time traffic conflict rates models simultaneously accommodating temporal correlation and unobserved heterogeneity across observations. Signal cycle level traffic data, including traffic conflicts, traffic and shock wave characteristics, collected from six signalized intersections were used. Three types of Tobit models: conventional Tobit model, temporal Tobit (T-Tobit) model, and temporal grouped random parameters (TGRP-Tobit) model were developed under full Bayesian framework. The results show that significant temporal correlations are found in T-Tobit models and TGRP-Tobit models, and the inclusion of temporal correlation considerably improves the goodness-of-fit of these Tobit models. The TGRP-Tobit models perform best with the lowest Deviance Information Criteria (DIC), indicating that accounting for the unobserved heterogeneity can further improve the model fit. The parameter estimates show that real-time traffic conflict rates are significantly associated with traffic volume, shock wave area, shock wave speed, queue length, and platoon ratio.