Dynamic treatment regimes: practical design considerations

Background Clinical management of chronic disease requires a dynamic treatment regime (DTR): rules for choosing the new treatment based on the history of response to past treatments. Estimating and comparing the effects of DTRs from a sample of observed trajectories of treatment and outcome depends on the untestable assumption that new treatments are assigned independently of potential future responses to treatment, conditional on the history of treatments and response to date (“sequential ignorability”). In longitudinal observational studies, sequential ignorability must be assumed, while randomization of dynamic regimes can guarantee it. Methods Using several clinical examples, we describe the simplest randomized experimental designs for comparing DTRs. We begin by considering an initial treatment A and a second treatment B, and discuss how a dynamic treatment regime that starts with A and leads (sometimes) to B, might be compared to either fixed treatment A or B. We also illustrate the problem of finding the optimal sequence of treatments in a DTR, when there are several choices. We describe and contrast two ways of incorporating randomization into studies to compare such regimes: baseline randomization among DTRs versus randomization at the decision points (sequentially randomized designs). Conclusions We discuss estimation and inference from both baseline randomized and sequentially randomized designs and conclude with a discussion of the differences between the experimental and observational approaches to optimizing and comparing dynamic treatment regimes.