Abstract
The objective of this research was to develop prediction models for total crashes and fatal or injury crashes for rural horizontal curves on undivided roads, with a focus on three distinct aspects. The first was an emphasis on assembling a large, high-quality data set. Crash prediction models were developed by using a data set of 11,427 rural horizontal curves on Wisconsin state trunk network roads with more than 13 parameters and four distinct types of crash data sets. The second focus area was to use regression tree analysis in creating a simple model of horizontal curve safety aimed at practitioners of systemic road safety management and creating subsets of data that warranted further analysis. Regression tree results identified the curve radius of approximately 2,500 ft as a significant point below which there is a marked increase in crashes on horizontal curves. The third focus area was to research the effect on horizontal curve crash prediction models of different selection criteria to assemble the crash data set. Models (total and fatal or injury) based on a crash data set with and without crashes in the proximity of intersections were compared. The results show that when crashes on horizontal curves are selected where crash report forms indicate the presence of a horizontal curve, crashes that occur in the proximity of intersections do not affect model results significantly; therefore, the inclusion of such crashes would increase the size of the data set and benefit model development.