Exploring Statistical Methods in Developing Safety Performance Functions

Safety performance function (SPF) is a tool to identify the contributing factors prone to traffic crashes in safety systemic analysis. It helps transportation planners and engineers to better understand the reasons causing crashes, identify the most severe hazardous locations, and propose the countermeasures and improvements. The traditional method of developing SPFs is the negative binomial regression in the family of the generalized linear model (NBGLM), very easy to interpret because of its linearity assumption. However, this traditional statistical method would have an inaccurate performance with more variables included in SPFs.

This study explored the statistical methods to develop the SPFs for curves. Horizontal curves are a basic element in the roadway network, also one of the dangerous locations on roads. In the US, horizontal curve alignments have three times higher crash rate than other types of roadways, much higher than the tangent sections on the same road. Florida has more than 200,000 curves on public roads. Currently, there is no updated information on curve safety statewide, thus no clear safety guidance for curve safety throughout the state.

Using the independent horizontal curve data in the state of Florida, this study analyzed the contributing factors for curve crashes. This paper explored the methods for developing SPFs such as the best subset selection method, the shrinkage method, the generalized additive model (GAM) method, and the tree-based method. These methods are more complex than the negative binomial so that they have more capabilities to handle more variables in SPFs.

Using curve crashes in Florida as an example, we examined the impacts of different variables including traffic volume, demographic, roadway characteristics, curve characteristics, and spatial relation to intersections. Then we developed the SPFs for curves via each method. The paper compared the contributing factors and the model performance of each SPF. In general, traffic volume, curve length, curve radius, speed limit and the number of lanes were the variables selected by most of the methods for estimating the number of crashes on curves. The results showed the complex methods achieved better SPFs than the traditional method and the traditional negative binomial method could be substituted.

This is important since the Federal Highway Administration (FHWA) suggests the traditional NBGLM for developing SPFs, which are only capable of a few contributing factors and can only interpret the linear relationships between the number of crashes with those factors. When the more complex methods included in the safety systemic analysis, as well as more in-detailed well-informed transportation-related data, people will have more powerful methods to understand the safety issue in a systemic approach.