Comparison of designs for adaptive treatment strategies: baseline vs. adaptive randomization

Clinicians apply adaptive treatment strategies when treating chronic disease: at each patient visit, the clinician decides whether to switch a patient's treatment at that time, given the patient's outcomes to date. Thus, each clinical strategy is based on an adaptive decision rule that determines when (if ever) a patient's treatment is switched. Previously, we defined an adaptive randomization scheme for evaluating such strategies. In this scheme, a patient's treatment may be switched at one of the follow-up times with probability that depends on the patient's observed outcomes under the initial treatment. These randomized switches generate treatment sequences that provide complete or partial data on adaptive treatment strategies of interest. We show that this design equivalently generates strategy-specific studies with drop-out mechanisms determined by the randomized switches. We compare the efficiency of this approach to baseline randomization, where each patient is assigned to one adaptive strategy, and prove that there exists an adaptive design with the same efficiency as baseline randomization. We further exploit the idea of randomized drop-out to obtain causal comparisons beyond the original class of strategies assigned by baseline randomization, illustrating the greater flexibility of the adaptive design for evaluating adaptive treatments. We use the theory of multiple imputation to characterize the potential loss of precision due to drop-out induced by the more flexible design. (C) 2002 Elsevier B.V. All rights reserved.