The evaluation of the effectiveness of safety improvements using accident data has been a difficult task complicated by anomalies such as the regression-to-the-mean phenomenon, lack of control through traditional experimental design considerations, and insufficient sample sizes (e.g., accident frequencies and/or number of experimental units - sites). Statistical methods have been proposed to remove subjective bias which may confound conclusions in such cases. This study, funded by the Federal Highway Administration, examines various statistical procedures under a variety of applications to determine which procedures are most effective under what circumstances. Examples range from the simplest before/after design with no comparison group to multiple observations in time on a before/after with comparison group study. These procedures were applied to situations with few sites and high accident frequencies (such as changing posted speed on multilane high volume highway sections) as well as situations with many sites but low accident frequencies (such as signalization at intersections). Both actual and simulated data is used in these comparison. The sophistication of the statistical methods range from the most elementary procedures currently in use to more recently proposed sophisticated methods of Empirical Bayes estimation and time series analyses. Recommendations are made regarding which procedures are optimal for specific study designs. Also, design considerations are proposed which may alleviate problems and simplify the statistical analysis apriori to data collection.
Abstract