Intersection crashes cost the nation more than $40 billion annually, account for more than one-fifth of all highway crash fatalities nationally, and totaled almost 75,000 in the Virginia Department of Transportation’s (VDOT) Northern Virginia District for the period 2001 through 2006. Although VDOT maintains several databases containing more than 170 data elements with detailed crash, driver, and roadway attributes, it was not clear to users of these databases how these data elements could be used to identify causal factors for these intersection crashes for two reasons: (1) the quality of some of the data elements was imperfect, and (2) and random variation is inherent in crashes. This study developed an approach to address these two issues.
To address the first issue, the completeness and accuracy of the 179 data elements that comprise the VDOT CRASHDATA database were assessed. For the 76 data elements for which the quality of the data was imperfect, eight rules for using these elements were developed. The rules indicate which data elements should be used in certain circumstances; which data elements are incomplete; and how to manipulate the data for certain applications.
To address the second issue, classification trees and crash estimation models (CEMs) were developed. The trees showed that specific causal factors, such as the approach alignment or surface condition, successfully indicate whether a given crash was a rear-end or angle crash. By extension, the trees suggested that intersection crashes were not purely random. Accordingly, it was feasible to develop CEMs that for 17 intersection classes predicted the number of crashes for a 1-year period for four crash types: rear-end, angle, injury, and total. The 68 CEMs showed deviance-based pseudo R-square values between 0.07 and 0.74, suggesting that the causal factors explained some, but not all, of the variation in intersection crashes. The CEMs varied by intersection class.
Two actions with regard to crash data analysis may be taken as detailed in this report. First, the eight crash data quality rules developed in this study should be considered for use on a case-by-case basis for studies requiring intersection crash data. Second, when they are collected at the crash scene, the variables that successfully classified rear-end and angle crashes may be given increased attention such that every effort is made to ensure these data elements are accurately recorded.