Patient subsets and variation in therapeutic efficacy.

1 The determination of what treatment is best for what kinds of patients is a general objective of clinical research. We consider here the extent to which this objective can be accomplished reliably in a single clinical trial. 2 The reliability of subset analyses is often poor due to problems of multiplicity and limitations in numbers of patients studied. Implications for the design of clinical trials are presented. 3 Statistical approaches to subset analysis are reviewed in a general manner. In order to obtain the degree of reliability usually demanded of clinical therapeutic evaluations, "statistically significant' interactions between relative treatment efficacy and subsets should be demonstrated. Exploratory analyses of subset differences are important but should be reported as hypotheses to be tested in separate studies.