Finding a New Safety Performance Function for Two-Way, Two-Lane Highways in Rural Areas

Abstract

For over 30 years, crash prediction models (CPMs) have been created and analyzed, with the objective being to find the best way to predict where crashes will occur and how to prevent them in the future. This has recently become a popular discussion and reality since the release of the Highway Safety Manual (HSM) and its CPM in 2010. However, many are still hesitant to begin implementing these methods as the accuracy can vary. This is a study testing the original HSM's CPMs to state-specific calibrated CPMs, and new, independent CPMs to find the best model for rural, two-lane highways in Kansas. Almost 300 miles of highway geometric data were collected to create these new models using negative binomial regression. The most significant variables in each model were found to consistently be lane width and roadside hazard rating. These models were compared against CPMs calibrated to be used on the HSM using nine validation segments. A difficulty to overcome was the large amount of animal-related crashes, as they account for 58.9 percent of crashes on Kansas highways. Removing those from the equation showed a large improvement in accuracy compared to other models created.