Analysis of real-time crash risk for expressway ramps using traffic, geometric, trip generation, and socio-demographic predictors

Abstract

There have been numerous studies on real-time crash prediction seeking to link real-time crash likelihood with traffic and environmental predictors. Nevertheless, none has explored the impact of socio-demographic and trip generation parameters on real-time crash risk. This study analyzed the real-time crash risk for expressway ramps using traffic, geometric, socio-demographic, and trip generation predictors. Two Bayesian logistic regression models were utilized to identify crash precursors and their impact on ramp crash risk. Meanwhile, four Support Vector Machines (SVM) were applied to predict crash occurrence. Bayesian logistic regression models and SVMs commonly showed that the models with the socio-demographic and trip generation variables outperform their counterparts without those parameters. It indicates that the socio-demographic and trip generation parameters have significant impact on the real-time crash risk. The Bayesian logistic regression model results showed that the logarithm of vehicle count, speed, and percentage of home-based-work production had positive impact on crash risk. Meanwhile, off-ramps or non-diamond-ramps experienced higher crash potential than on-ramps or diamond-ramps, respectively. Though the SVMs provided good model performance, the SVM model with all variables (i.e., all traffic, geometric, socio-demographic, and trip generation variables) had an overfitting problem. Therefore, it is recommended to build SVM models based on significant variables identified by other models, such as logistic regression.