Crash modeling for intersections and segments along corridors: A Bayesian multilevel joint model with random parameters

Previous highway safety studies have focused on either intersections or roadway segments while some researchers have analyzed safety at the corridor-level. The corridor-level analysis, which aggregates intersections and roadway segments, may allow us to understand the safety problems in the wider perspective. However, it would result in losing some of the specific characteristics of intersections or roadway segments. Therefore, we proposed a multilevel joint model that explores traffic safety at the segment/intersection level, with the consideration of corridor-level variables. In addition, the variations in the roadway characteristics and/or traffic volumes across corridors have been considered using random parameters model. Nevertheless, sometimes corridors are excessively long and, thus, it is uncommon to find corridor-level variables that have fixed values for the entire length of corridors. Therefore, current corridors were divided into sub-corridors, which have similar traffic volumes and roadway characteristics, and constructed another multilevel structure based on the sub-corridor. As a result, four Bayesian models have been estimated, and these models are multilevel Poison-lognormal (MPLN) joint models with spatial corridor and sub-corridor random effects terms and MPLN joint models with random parameters, which vary across corridors and sub-corridors. Based on a 3-years crash data from 247 signalized intersections and 208 roadway segments along 20 corridors in two counties, results showed that four-roadway segment, five-intersection, and three-corridor/sub-corridor variables were significant, and they include exposure measures and some geometric design variables. With respect to model performance, it was found that the MPLN joint model with random sub-corridor parameters provides the best fit for the data. Lastly, it is suggested to consider the proposed multilevel structure for corridor safety studies. (C) 2017 Elsevier Ltd. All rights reserved.