Last, we used propensity score (PS) methods to examine the combined effects of all candidate mediators selected via prior models in contributing to racial/ethnic differences in effects for alcohol use on alcohol problems. We selected PS methods as a final step in testing mediation because PS methods do not rely on the many often untenable assumptions of standard covariance adjustment, and as a result, may offer more reliable inferences (Rubin, 1997). PS methods have historically been used predominantly to account for differences in probability of treatment exposure, but can also be used to address causal pathways/confounding in other circumstances. Rosenbaum and Rubin showed that at a given PS, the conditional distribution of confounders is independent of the treatment assignment/exposure, and the PS adjustment is thus sufficient to remove bias due to observed covariates (Rosenbaum and Rubin, 1983). Here, we used PS weighting (Robins et al., 2000) and “standardized mortality ratio” (SMR) weights, which assign a weight of 1 to the “exposed” group (Black or Latino) and the propensity odds (i.e., p/(1−p), where p is the predicted PS) to the comparison
