Researchers in the field of accident analysis have long been aware uf the problems associated with drawing statistical inference on safety using accident data. Aside from the problems of accessibility and quality, accident data present a real challenge when it comes to statistical analysis. One of the most serious problems in accident analysis is the regression-to-the-mean bias which occurs due to the non-random site selection process in safety measure evaluation studies. This study presents a new empirical Bayes method (EBEST) which adjusts for regression-to-the-mean bias. Three typical applications in accident analysis are considered for regression-to-the mean bias, namely: 1. the evaluation of safety treatments, 2. the identification of high hazard locations, and 3. the assimilation of information from multiple safety measure studies (meta-analysis). A computer program was developed to execute these analyses as a part of this study. This manuscript describes the EBEST (Empirical Bayes Estimation of Safety and Transportation) methodology and presents examples of how the method works for each of the three accident analysis applications. This report appears in three volumes. Volume I: General methodology, FHWA-RD-90-091 , is a nonstatistical review of the study, Volume II: A users manual for BEATS, FHWA-RD-91-014, is a user's manual for the BEATS computer program, and Volume III: Theoretical development of new accident analysis methodology, FHWA-RD-91-015, contains the theoretical development of the procedure. (Author/publisher)
Abstract