ABSTRACT
Chapter 13 introduces time-dependent confounders as a variable that serves as an intermediate variable for a previous treatment and as a confounder for a subsequent treatment. It shows how to estimate the joint effects of those two treatments using marginal structural models, structural nested mean models, and optimal dynamic treatment regimes. The three different approaches enlist three different sets of potential outcomes. Corresponding sequential randomization assumptions are required for valid estimation via a given approach; these replace the sufficient set of confounders assumption used for time-invariant confounders. Proof of validity under the assumptions is provided for each approach. Examples and R code are also provided.
