Generalized Linear Models to Identify the Impact of Road Geometric Design Features on Crash Frequency in Rural Roads

Abstract

One of the first steps in improving the road network is to implement a traffic safety program and identify problematic road segments where local risk factors increase the likelihood of vehicular accidents. Numerous studies have attempted to identify both the geometric and traffic features of road segments that may lead to increased risk. The statistical models linking the number of accidents at road sites over a unit of time and the characteristics of these sites contribute to highlighting and quantifying the impact of the development of road infrastructure on the risk of accidents. The evaluation and selection of the best-fit model are usually carried out using an information criterion. The main purpose of this research is to propose a comparative assessment framework to support the evaluation task during the process of modeling crash frequency. Performance evaluation was conducted using additional approaches, specifically data splitting and rootograms. In this study, we focus on Korean rural road accident records. Six generalized linear models were applied, and their performance was compared to determine the best-fitting statistical model. According to each model assessment method used, the negative binomial hurdle model showed the best performance among all the investigated models.