Abstract :
Although clinical trials report results in the aggregate, clinicians often wish to tailor treatments that are based on demographic, historic, clinical, or laboratory characteristics of their patients and are interested therefore in trial results that are presented separately according to such characteristics. Unfortunately, such subgroup analysis often are done incorrectly and often are misinterpreted, even when done correctly. Only subgroups that are defined by characteristics that are determined at or before the moment of randomization are valid. It is incorrect to determine whether treatment is effective among a subgroup of patients according to the probability value in that particular subgroup. The correct method uses a formal test for interaction. However, even when done correctly, most subgroup differences in treatment effectiveness prove to be spurious. A priori definition of the subgroup, strong supporting rationale, and, ultimately, replication in other studies increase confidence that subgroup differences are valid.
