Modeling the Impact of Subject and Opponent Vehicles on Crash Severity in Two-Vehicle Collisions

The main goal of this research was to develop models for crash severity prediction in a subject vehicle while taking into account the impact of both vehicles involved (Vehicles V1 and V2). Three binary targets were modeled: FataISIK (to predict overall severity), FataISIKV1 (to predict the probability of a serious injury, fatality, or both in Vehicle V1), and FatalSIKV2 (to predict the probability of a serious injury, fatality, or both in Vehicle V2). For the period from 2006 to 2010, 874 collisions involving injuries, fatalities, or both were analyzed. However, the crash sample included few severe events. Because imbalanced data introduce a bias toward the majority class (nonsevere crashes), in predictive modeling, they would result in less accurate predictions of the minority class (severe crashes). For the challenge imposed by small sample size and imbalanced data to be overcome, an important methodology was developed on the basis of a resampling strategy by using 10 stratified random samples for model evaluation. The effect of vehicle characteristics such as weight, engine size, wheelbase, and registration year (age of vehicle) were explored. Logistic regression analysis for FataISIK suggested that the age of the vehicle and the type of collision were significant predictors (p < .0084 and .0346, respectively). Models FataISIKV1 and FatalSIKV2 showed that the engine size of the opponent vehicle was statistically significant in predicting severity (p < .0762 and p < .03875, respectively). Models for FataISIKV1 and FatalSIKV2 yielded satisfactory results when evaluated with the 10 stratified random samples: 61.2% [standard deviation (SD) = 2.4] and 61.4% (SD = 3.1), respectively.