On problems related to subgroup analyses in clinical trials

Background:The intention of randomized clinical trials is usually the comparison of two or more treatments in a population of patients with a specific disease. Most often, the trial is planned for an overall comparison of the treatment groups. During data analysis of the trial often further questions emerge for treatment differences in certain subgroups of patients defined by prognostic factors. In extreme situations, this may lead to a division into many different subgroups. The treatment groups are then compared with respect to the outcome of interest in each subgroup. Material and Methods:The probable consequences of this procedure are illustrated by means of a 'hypothetical clinical trial' in breast cancer patients. Results: It is shown that 'significant' differences between treatment groups can emerge just by chance. Conclusion: With an extensive subgroup analysis the probability of falsely judging a non-existing difference as significant is no longer controlled and the results may be misleading.