How the choice of safety performance function affects the identification of important crash prediction variables

•We compared safety performance functions for crash rate modeling at intersections.•Boosted regression trees had the best predictive performance.•All models differed on the identification of important intersection characteristics.•We found strong interaction and nonlinear effects.•Ignoring interactions and nonlinearities can lead to sub-optimal treatment strategies. Across the nation, researchers and transportation engineers are developing safety performance functions (SPFs) to predict crash rates and develop crash modification factors to improve traffic safety at roadway segments and intersections. Generalized linear models (GLMs), such as Poisson or negative binomial regression, are most commonly used to develop SPFs with annual average daily traffic as the primary roadway characteristic to predict crashes. However, while more complex to interpret, data mining models such as boosted regression trees have improved upon GLMs crash prediction performance due to their ability to handle more data characteristics, accommodate non-linearities, and include interaction effects between the characteristics. An intersection data inventory of 36 safety relevant parameters for three- and four-legged non-signalized intersections along state routes in Alabama was used to study the importance of intersection characteristics on crash rate and the interaction effects between key characteristics. Four different SPFs were investigated and compared: Poisson regression, negative binomial regression, regularized generalized linear model, and boosted regression trees. The models did not agree on which intersection characteristics were most related to the crash rate. The boosted regression tree model significantly outperformed the other models and identified several intersection characteristics as having strong interaction effects.