The purpose of this study was to identify contributory factors to severity of rollover crashes in the mountainous state of Wyoming. These crashes account for more than half of all roadway fatalities in Wyoming, compared with the average of the U.S. rollover-related fatality crashes, which stands at 33%. In this study, the standard generalized linear model (GLM) was extended to the method of generalized additive model (GAM) to determine if giving more flexibility provides more realistic point estimates of the factors to the rollover crash severity. The results highlighted the superiority of the GAM compared with the GLM in terms of confusion matrix accuracy and Akaike Information Criterion (AIC). The results of the GAM highlighted that the majority of important factors that contribute to rollover crash severity are related to drivers' characteristics such as driving while under influence of drugs, being under an emotional condition, driving with no valid driver license, and driving with suspended drivers' license. Also, it was found that the impact of passenger vehicles on the severity of rollover crashes is not stable and varies based on the gender of drivers. Only two predictors were considered based on the smooth functions including posted speed limit and drivers' age. We accounted for non-linearity of those two predictors by means of cubic spline smooth function.
Keywords: Generalized additive model; crash severity; rollover crashes; smooth function.