Predictive Crash Risk Model for Network Screening on Rural Two-Lane Highways

Network screening is a critical step in the ongoing Highway Safety Improvement Programs (HSIPs). The overall effectiveness of these programs largely relies on using a robust method for identifying sites with high potential for safety improvements. This study proposed a new crash risk model designed for rural highways, where access to extensive roadway, crash, and traffic data may be limited. Specifically, the proposed model does not require exact measurement of roadway geometric features as all geometric variables were classified into categories that are easy to compile by highway agencies. The dependent variable used in this model was crash frequency while the independent variables included roadway and roadside characteristics besides traffic exposure.  The performance evaluation was conducted utilizing a dataset encompassing 1495 miles of rural two-lane highway segments in Oregon. This dataset included information on roadway geometry, roadside characteristics, traffic conditions, and crash records for ten years. The model development phase involved the utilization of a training dataset spanning five years, (2011-2015). Subsequently, an independent dataset spanning the subsequent five-year period, from 2016 to 2020, was employed for the purpose of examining the performance of the proposed model in network screening. To evaluate the effectiveness of the proposed methodology, network screening using the proposed model was compared to that using two established methods: the Empirical Bayes (EB) method and the Potential for Safety Improvement (PSI) method. Using crash frequencies for highway segments, the proposed method demonstrated comparable performance to the well-established EB and PSI methods. The difference in the identification of true positive segments was 8% with the EB method and 10% with the PSI method. Similarly, the proposed method exhibited a correlation coefficient of 0.864 between the segment-predicted number of crashes and those observed using the 5-year crash history.