Evaluation of Signalized-Intersection Crash Screening Methods Based on Distance from Intersection

Abstract

A probability sample of signalized intersections was collected from Oregon to review the performance of various strategies for identifying intersection-related crashes. Data from 73 intersections were collected with satellite imagery and databases from the Oregon Department of Transportation. Crashes from 2010 to 2012 were identified and manually classified as either intersection-related or not on the basis of geolocation, field codes in the crash databases, and geometric characteristics. This evaluation identified two sets of crashes: one with 964 intersection-related and another with 571 non-intersection-related crashes. The authors then studied how the probability of a crash's being intersection-related changes with the distance to the intersection and other relevant variables. Statistical model development used 55 intersections; the remaining 18 were left for comparison of various classification methods. Findings of this comparison were confirmed with data from all 73 intersections. Results indicated that regression models yielded results comparable with those from simpler methods based on distance only. The popular method of selecting crashes that are within a 250-ft radius tends to identify fewer intersection-related crashes than actually exist. A threshold of 250 ft, therefore, would result in underprediction of intersection crashes when safety performance functions are developed. This research found that a threshold of 300 ft potentially minimized such a risk of underestimation; however, using 250 ft might still be appropriate if minimizing the frequency of false positive identifications was of interest.