Generalized Linear Modeling of Crashes on Urban Road Links

Abstract

As crash data have distinctive behavior like over-dispersion, researchers have used statistical methods to deal with this unique behavior of crash data specifically. This study employed generalized linear modeling techniques to develop the model. It was assumed that the accident counts followed negative-binomial distribution, and the link function was chosen to be the log link function. Negative-binomial modeling technique was chosen over Poisson distribution because it is the most used technique by many researchers as crash data may encounter over-dispersion. The accident data set showed greater variability between its variance and mean. The accident frequency distribution is shown in this study that it is highly skewed, with a very high number of road segments registering zero accidents. Negative binomial distribution was chosen over Poisson distribution after comparing Akaike’s Information Criterion (AIC) and Bayesian Information Criteria (BIC). The method is widely applied to count data. Twenty-two parameters were estimated in the model. Since p < 0.05 in the omnibus test, the null hypothesis is rejected, which indicates that the model is reasonably fit. The strongest variables in the model were witnessed to be the length of the links, number of lanes, average daily traffic, bus lane, number of buses and trolleys, and HGVs.