Using multivariate adaptive regression splines (MARS) to develop crash modification factors for urban freeway interchange influence areas

Abstract

Crash modification factors (CMFs) are used to measure the safety impacts of changes in specific geometric characteristics. Their development has gained much interest following the adoption of CMFs by the recently released Highway Safety Manual (HSM) and SafetyAnalyst tool in the United States. This paper describes a study to develop CMFs for interchange influence areas on urban freeways in the state of Florida. Despite the very different traffic and geometric conditions that exist in interchange influence areas, most previous studies have not separated them from the rest of the freeway system in their analyses. In this study, a promising data mining method known as multivariate adaptive regression splines (MARS) was applied to develop CMFs for median width and inside and outside shoulder widths for “total” and “fatal and injury” (FI) crashes. In addition, CMFs were also developed for the two most frequent crash types, i.e., rear-end and sideswipe. MARS is characterized by its ability to accommodate the nonlinearity in crash predictors and to allow the impact of more than one geometric variable to be simultaneously considered. The methodology further implements crash predictions from the model to identify changes in geometric design features. Four years of crashes from 2007 to 2010 were used in the analysis and the results showed that MARS's prediction capability and goodness-of-fit statistics outperformed those of the negative binomial model. The influential variables identified included the outside and inside shoulder widths, median width, lane width, traffic volume, and shoulder type. It was deduced that a 2-ft increase in the outside and inside shoulders (from 10ft to 12ft) reduces FI crashes by 10% and 33%, respectively. Further, a 42-ft reduction in the median width (from 64ft to 22ft) increases the rear-end, total, and FI crashes by 473%, 263%, and 223%, respectively.