An Application of Decision Tree Models toExamine Motor Vehicle Crash Severity Outcomes

Abstract

Classification and Regression Tree (CART) and chi-square automatic interaction detection (CHAID) decision tree models are estimated and compared to examine the effect of driver characteristics and behaviors, temporal factors, weather conditions, and road characteristics on motor vehicle crash severity levels using Missouri crash data from 2002 to 2012. The CHAID model is found to significantly better discriminate among severity outcomes, and results suggest that the presence of alcohol, speeding, and failing to yield lead to many fatalities each year and likely have interactive effects. Decision rules are used to identify changes in driving policies expected to reduce severity outcomes.