Stochastic Models Relating Crash Probabilities With Geometric And Corresponding Traffic Characteristics Data

Several kinds of models have been developed in modeling the occurrence of crashes, but most of these models have deficiencies and lack good results. Single and multivariate deterministic models have illustrated some influences of causal factors on crashes, but the inherent deterministic characteristic of these models makes explaining crash events a difficult task for these kinds of models. Stochastic regression models, such as Poisson, Negative Binomial, and Zero Inflated Poisson (ZIP) have been explored to account for the discrete and stochastic characteristics of crashes. However, no consistent results have been illustrated yet. The possible reasons for the deficiencies of existing models could be attributed to the modeling methodologies or the data set used. We have opportunities to obtain corresponding traffic data related to the time when crashes take place from the Smart Travel Lab at University of Virginia. Those data include volume, speed, and occupancy. Based on the review of the existing models, data obtained from the Smart Travel Lab were used in the application of several stochastic regression models including Poisson, negative binomial, zero inflated Poisson, and zero inflated negative binomial regression models. The selected variables include crash counts, volume, speed, occupancy, curvature, exposure, and standard deviation of speed. Negative binomial and ZIP were shown to be preferred modeling methods for this study. Significant positive relationships were also identified between the occurrence of crashes and volume, standard deviation of speed, and exposure. These relationships could be applied to provide powerful support to the decision making of incident management in Intelligent Transportation Systems.