Coping with endogeneity and unobserved heterogeneity in real-time clustering critical crash occurrences nested within weather and road surface conditions.

The knot of endogeneity and unobserved heterogeneity are causes of rendering parameter estimates inconsistent in real-time crash prediction. This study intends to alleviate the potential sources of these issues in detecting critical crashes, involving fatal or incapacitating injuries, versus non-critical crashes through a 402-mile Interstate-80 in Wyoming. Among different types of endogeneity, the problem of errors-in-variables and simultaneity was respectively mitigated by conflating disaggregated real-time traffic observations with aggregated environmental conditions and removing secondary crashes from the dataset. The possibility of omitted variables and unobserved heterogeneity were dealt by using random intercepts in hierarchical modeling under Bayesian inference. Three models were calibrated. Model-1 treated all predictors as fixed parameters. Model-2 and Model-3, respectively, considered weather and road surface conditions as random intercepts. Model-2 outperformed the others where the Intraclass Correlation Coefficients confirmed that the crash dataset is more nested within weather conditions. Results indicated that critical crashes require more interaction between vehicles, and shaping backward shockwave reduces their risk on Interstate-80 with a comparatively less traffic volume. Furthermore, considering different categories of weather and road surface conditions, critical crashes are more likely to occur on dry pavement and cloudy conditions compared to the wet surface and rainy days.