ABSTRACT
Adaptive methods have become both popular and controversial. We cover the full range of methods, from blinded and other procedures that do not use information about the treatment effect to those that use the treatment effect to modify design features, predominantly sample size. Classic methods for continuous outcomes, such as that of Stein (1945) and the naïve method of pretending the revised sample size was fixed in advance, are included. Blinded methods are considered for both continuous and binary outcomes. Methods based on examining the treatment effect are more controversial and have received criticism. Nonetheless, studying the early papers like Bauer and Kohne (1994) and Proschan and Hunsberger (1995) fosters understanding of the key adaptation principle underlying several methods. The conditional error principle of Muller and Schafer (2004) is a very important improvement whose full potential has yet to be appreciated. The appealing method of Chen, DeMets, and Lan (2004) allows a change of sample size based on examination of the treatment effect, provided that the observed trend is promising. We end with an emergency method that can be used to salvage a trial requiring unplanned changes. The key tool is what we have emphasized throughout the book, the re-randomization test. Although the re-randomization test provides valid inference, important caveats apply.
Lan and DeMets, 1983 missing, as is Dunnett reference.
