Clustering-Based Roadway Segment Division for the Identification of High-Crash Locations

This article introduces a clustering approach to roadway segment division, in place of the traditional fixed-length and variable-length division methods, to improve the calibration of safety performance functions (SPFs) for the purpose of identifying high-crash locations. The clustering approach helps to reduce crash heterogeneity for within-group elements by grouping roadway segments with similar crash distributions into homogeneous groups. For comparison purpose, all three segment division methods were applied to a 142.6-kilometer (88.6-mile) stretch of freeway on Interstate 95 that spans three counties in southern Florida in the United States. Using 5 years of crash data occurring on segments generated from each of the three division methods, the corresponding SPFs were calibrated using the negative binomial model. The calibrated SPFs were then used in the empirical Bayes approach of identifying high-crash locations. The results showed that clustering method produced a much better-fitted SPF than that produced by using the traditional division methods. Furthermore, the site screening for high-crash locations on segments divided by the clustering method improved upon the shortcomings of that using the existing sliding window method.