Background: Although randomized clinical trials are usually planned for an overall comparison of treatment groups, sometimes unplanned subgroup analyses will be conducted subsequently. In extreme situations, the population is divided into many different subgroups defined by prognostic factors, and treatment groups are compared with respect to the outcome of interest within each subgroup. The risks of this procedure are well known. The probability of erroneously judging a non existing difference between treatment groups as significant is no longer controlled and the results will be misleading. One may think that subgroup analysis in meta-analysis is less error-prone than in individual trials because identifying an effect in a subgroup of one trial may be outweighed by no effect in the same subgroup of other trials.
Objectives: It will be investigated whether the multiplicity problem of subgroup analyses, i.e. the danger of wrongly judging non existing differences between treatment groups as significant, is reduced when subgroup analyses are performed in meta-analyses instead of individual clinical trials.
Methods: The consequences of subgroup analyses in individual clinical trials and in meta-analyses will be investigated by simulating clinical trials. Starting with real data of a clinical trial on high-dose chemotherapy in breast cancer, a set of clinical trials will be generated artificially by randomly allocating treatment labels A or B to the patients.
Results: Subgroup analyses lead to an increased probability of falsely judging a non existing difference between treatment groups as significant also in meta-analyses of randomized clinical trials. The multiplicity problem of subgroup analyses is not reduced in meta-analyses as compared to individual clinical trials.
Conclusions: Results of subgroup analyses in meta-analyses of randomized clinical trials have to be interpreted with caution, and methods for correcting the type I error are necessary.