Comparison of Crash Modification Factors for Engineering Treatments Estimated by Before–After Empirical Bayes and Propensity Score Matching Methods

Abstract

Cross-sectional and the empirical Bayes (EB) before–after are two of the most common methods for estimating crash modification factors (CMFs). The EB before–after method has now been accepted as one way of addressing the potential bias caused by the regression to the mean problem. However, sometimes before–after methods may not feasible because of the lack of data from before and after periods. In those cases, researchers rely on cross-sectional studies to develop CMFs. However, cross-sectional studies may provide biased CMFs through confounding. The propensity score (PS) matching method, along with cross-sectional regression models, is one of the methods that can be used to address confounding. Though PS methods are widely used in epidemiology and other studies, there are only a few studies that have used PS matching methods to estimate CMFs. The intent of this study is to evaluate and compare the performance of cross-sectional regression models using PS matching methods with the results from the EB and traditional cross-sectional methods. The comparisons were conducted using two carefully selected simulated datasets. The results indicate that optimal propensity score distance (PSD) matching with maximum variable ratio of 5 performed quite well compared with the EB before–after and the traditional cross-sectional methods.