A Markov switching regression analysis of freeway crash risks considering spatial effect

Abstract

A Markov switching logit model with spatial dependencies for real-time crash risk assessment is proposed, with the purpose of identifying hazardous traffic-flow conditions with high crash potential. The Markov switching process assumes that freeway segments can switch between two unobserved safety states over time, and that parameter estimates may vary between these two states. The spatial simultaneous autoregressive process was used to account for possible dependencies in crash likelihood between neighbouring freeway segments. The proposed model was used to link crash likelihood with real-time traffic, weather and roadway geometry data. The Bayesian inference method based on Markov chain Monte Carlo simulations was used for model estimation. Bayes factor analysis suggested that the proposed model produces a better fit than other alternatives that ignore temporal or/and spatial effects. The estimation results revealed that two states with regard to crash likelihood exist and that freeway segments can switch between these states over time. The spatial autocorrelation coefficient indicated that crash likelihood on a freeway segment is interlinked with those on neighbouring segments over space. The key traffic variables contributing to crash likelihood are detector occupancy, occupancy difference between adjacent lanes, speed variance and occupancy difference between upstream and downstream detector stations. Moreover, three geometric variables and weather conditions are significantly related to crash likelihood in the model. The receiver operating characteristic curve showed that the predictive performance of the proposed model is satisfactory.