Randomized Selection Designs

The general goal of a randomized selection design is to select one or more treatments from several competing candidates to which patients are randomly assigned, in such a way that selected treatment(s) are likely to be better than those not selected. For example, if one treatment is clearly superior to all the others, we may demand that the procedure select that treatment with high probability. The experimental treatments could be different doses of a drug or intensities of a behavioral intervention, different treatment schedules, modalities, or strategies, or different combinations of treatments. The hallmark feature of a selection design is its ability to achieve its stated goals with surprisingly fewer participants compared with traditional “phase III” trials, precisely because it eschews the formal hypothesis test paradigm with its tight control over type 1 error rates. These designs can be used in clinical research to screen for treatments that are worthy of further evaluation in a subsequent confirmatory clinical trial and to discard unpromising treatments. Thus, they are ideal for middle development settings where we are interested in selecting promising treatments under circumstances typically limited by smaller sample sizes. In this chapter, we discuss the randomized selection designs of Simon, Wittes, and Ellenberg, Steinberg and Venzon, and the Levin-Robbins-Leu family of sequential subset selection procedures. The first two designs select a single treatment, while the latter allows for sequential elimination of inferior treatments, sequential recruitment of superior treatments, and may be used to select treatments with fixed or variable subset sizes.