Appropriate Regression Model Types for Intersections in SafetyAnalyst

Abstract

Highway agencies in the United States have started implementing SafetyAnalyst, highway safety management software. Some agencies have been developing or considering developing safety performance functions (SPFs) using local data to maximize the benefits of the software. With multiple years of data on hand, several models are available for developing the SPFs, yet there is no reliable guide to selecting appropriate models. To this end, this study examined 11 models, including 8 panel and 3 cross-sectional models, to recommend appropriate model types for intersection SPFs of SafetyAnalyst. Data collected from 18,356 intersections in Virginia over the period 2003–2008 were used. A cross-sectional model with summed or averaged crash frequencies was found to severely underestimate a dispersion parameter that plays a critical role in the empirical Bayes method. Panel models with independent correlation and autoregressive Order 1 correlation and a pooled cross-sectional model were found to be appropriate across all 8 subtypes. Agencies unable to construct panel data and develop panel models can apply the pooled cross-sectional model, which does not require matching the same intersections across different years.