Understanding Causality of Intersection Crashes

Intersection crashes in the United States account for more than one-fifth of all highway fatalities. Crash, geometric, and roadway information can help identify crash causes. How such elements can serve this function, however, may not be clear to database users because of the random variation inherent in crashes. For this reason, classification trees and crash estimation models (CEMs) were developed from a 6-year data set, which contained more than 70,000 crashes that occurred at more than 6,000 intersections in Northern Virginia. The trees showed that specific causal factors, such as surface condition, indicated whether a given crash was rear-end or angle. Because such trees suggested that intersection crashes were not purely random, CEMs for 17 intersection classes were developed on the basis of traffic control, number of approaches and lanes, and rural versus urban area to predict four crash frequencies: rear-end, angle, injury, and total. The 68 CEMs showed deviance-based, pseudo–R-squared values between .07 and .74 and varied by intersection class. Of the nine angle crash models in which risk increased by making the approaches undivided, the increase varied between 43% and 154%. Two lessons emerged. First, the small proportion of variables that successfully classified most rear-end and angle crashes should be given increased attention to ensure that these data elements are recorded accurately at the crash scene. The methodology used in this study showed that much of the tree factor space (81%) was composed of only 10 variables. Second, facility-specific intersection CEMs should be developed because a geometric variable may be a surrogate for other phenomena.