Abstract
One of the most used methods to examine sources of heterogeneity in meta-analyses is the so-called 'subgroup analysis'. In a subgroup analysis, the included studies are divided into two or more subgroups, and it is tested whether the pooled effect sizes found in these subgroups differ significantly from each other. Subgroup analyses can be considered as a core component of most published meta-analyses. One important problem of subgroup analyses is the lack of statistical power to find significant differences between subgroups. In this paper, we explore the power problems of subgroup analyses in more detail, using 'metapower', a recently developed statistical package in R to examine power in meta-analyses, including subgroup analyses. We show that subgroup analyses require many more included studies in a meta-analysis than are needed for the main analyses. We work out an example of an 'average' meta-analysis, in which a subgroup analysis requires 3-4 times the number of studies that are needed for the main analysis to have sufficient power. This number of studies increases exponentially with decreasing effect sizes and when the studies are not evenly divided over the subgroups. Higher heterogeneity also requires increasing numbers of studies. We conclude that subgroup analyses remain an important method to examine potential sources of heterogeneity in meta-analyses, but that meta-analysts should keep in mind that power is very low for most subgroup analyses. As in any statistical evaluation, researchers should not rely on a test and p-value to interpret results, but should compare the confidence intervals and interpret results carefully.