Recent studies have shown that more than 200,000 red light running crashes occur annually in the United States. It can be hypothesised that the majority of these accidents result from "driver error", whether intentional or inadvertent. However, the purpose of this research is the possible contribution of the geometric characteristics of intersections to RLR crash risk - such factors as entering or cross-street traffic volumes and intersection width. The focus of this current paper is the statistical modelling strategy chosen from many possible strategies to answer the questions using data from the Federal Highway Administration's the Highway Safety Information System (HSIS). The HSIS contains multi-year, multi-state data on accidents, roadways, and traffic volumes. Poisson and Negative Binomial regression models in PROC GENMOD were utilised to test the hypotheses. Using this example, the authors are attempting to demonstrate the use of this strategy in terms of when to model the dependent variable for different conditions, how to select predictors if multiple years involved, and how to approach interaction effects. Above all, is the model parsimonious, consistent, or stable? In short, when do we have enough confidence to reject (or fail to reject) the research hypotheses of any association between dependent and independent variables. (A)
Samenvatting