ABSTRACT
Chapter 6 explains how to use the Backdoor Theorem to adjust for confounding via standardization, thereby estimating a population averaged causal effect. It presents the outcome-modeling and the exposure-modeling approaches to standardization. It also presents outcome- and exposure-modeling approaches to estimating the average effect of treatment on the treated. The outcome- and exposure-modeling approaches are shown to agree when both models are nonparametric, but not when they are parametric. Doubly robust estimation is introduced to reconcile disagreement; it is shown to be a valid approach when at least one of the two models is correct. Propensity scores and prognostic scores are defined, and used to verify a violation of faithfulness in a data example. The outcome- and exposure-modeling approaches are compared to each other and to the doubly robust approach with a simulated example and with real data examples. R code is also provided.
