Developing Safety Performance Functions for Two-Lane Two-Way Rural Highways Using Multiple Regression Techniques - A Comparative Analysis

Developing crash prediction models, also known as Safety Performance Functions (SPFs), for low-volume rural two-lane two-way highways is often a challenging task. The crash data for these roadways are characterized by low sample mean and small sample size. To this point, there has not been a proper guidance on developing SPFs for such roadways in rural states. This study fills this gap by developing 36 crash prediction models using four different types of regression models, namely the Negative Binomial (NB), Zero-Inflated Negative Binomial (ZINB), Tobit, and Conway-Maxwell-Poisson (COM-Poisson). Models were developed using three sets of crash data comprising of crashes between 2007 and 2011: the US-287 roadway segments in Wyoming, Colorado, and the whole corridor combined. The purpose of including Colorado and the combined dataset is to draw a comparison of the model performance on datasets with different sample sizes. The results showed that overall COM-Poisson outperformed other models in terms of model fit. However, for datasets with relatively lower sample means such as in the case of Fatal and Injury (FI) crash models, ZINB can be a suitable alternative. The prediction capability of the developed models was also evaluated with a validation dataset prepared from Wyoming’s 2016- 2019 crash data. The analysis further confirms that COM-Poisson is a better performing model. The transferability of Colorado and the combined SPFs to Wyoming data was tested. It can be concluded that the neighboring jurisdictions' SPFs can be borrowed without scarifying the accuracy by a large margin.

Language

  • English

Media Info

  • Media Type: Digital/other
  • Features: Figures; References; Tables;
  • Pagination: 17p

Subject/Index Terms

Filing Info

  • Accession Number: 01764945
  • Record Type: Publication
  • Report/Paper Numbers: TRBAM-21-03597
  • Files: TRIS, TRB, ATRI
  • Created Date: Dec 23 2020 11:25AM