non-experimental research, including collapsibility and exchangeability. It will then introduce the propensity score as an efficient method to achieve exchangeability. Propensity scores have additional advantages for non-experimental comparative effectiveness research and pharmacoepidemiology. These include the ability to estimate causal parameters in the presence of heterogeneous treatment effects. Heterogeneity of treatment effects can also be an indication of unmeasured confounding, however. Sensitivity analyses excluding various proportions of those treated contrary to prediction can help to detect and reduce unmeasured confounding in such settings. The presentation will be based on concepts, results from simulation studies, and empirical examples.