Problems associated with comparisons of response-defined subsets of patients in randomized trials. Treatment-related bias and response migration.

Subset analysis may be justified between various arms of randomized trials as long as subsets are defined by variables which do not cause distortion of other prognostic factors. However, bias will occur when treatment response is used to define a subset of patients in which the results of the same treatment are analyzed. Thus, comparisons between groups of responders in randomized studies are just as inherently biased as comparisons between responders and non-responders. Using a constructed example the effects of treatment-related bias on the interpretation of trial results are demonstrated, and the concept of response migration is introduced. It is shown that in randomized trials the only unbiased measures of treatment efficacy are response rate, overall time to progression, and overall survival.