ABSTRACT

A rigorous account of a statistical framework for the multiple decision setting is presented, targeted to readers who are interested in the technical underpinnings. Precise descriptions of the potential outcomes framework and required assumptions (SUTVA, sequential randomization, positivity) are given based on a formal definition of the sets of treatment options that are feasible for patients with a given history at each decision point, which implies a formal definition of regimes themselves. A formal proof of the g-computation algorithm is presented. Methods for estimation of the value of a fixed, given treatment regime based on regression modeling, inverse and augmented inverse probability weighting, and marginal structural models are described in detail.