|
Abstract:
|
We consider sequentially randomized trials in which treatments are assigned in response to observed intermediate outcomes. Patients may, for example, be randomized to the remaining treatment options if they have responded poorly to their initial treatment. Conventional marginal structural models (MSMs) based on static treatment regimens often fail in this context due to a clear violation of the Experimental Treatment Assignment assumption. We show that MSMs can be used, however, to model mean counterfactual outcomes corresponding to realistic individualized treatment rules. Such rules may, for example, start patients on a particular treatment a0 and then switch them to another treatment a1 if their intermediate outcome becomes unfavorable. We present IPTW estimators for this problem and illustrate the methodology through a simulation study.
|
