Examining a result separately in several subgroups means several opportunities for a difference to appear by chance.
Examining a result separately in several subgroups means several opportunities for a difference to appear by chance. With enough subgroups, one will reach significance in a trial with no real effect at all, and this is arithmetic rather than suspicion.
A trial sized for an overall comparison is too small for any slice of it. Subgroup findings therefore combine a high chance of false positives with a high chance of missing real differences, which is the worst available combination.
Subgroups named in the protocol before data collection are a different proposition from subgroups found afterwards. The first is a hypothesis being tested, the second is a hypothesis being generated, and papers should say which.
Showing an effect is significant in one group and not another is not evidence they differ. The correct analysis tests whether the effect sizes themselves differ, and that test is frequently absent where the claim is made.
Subgroup findings replicate at low rates. Several well-known subgroup claims that shaped practice for years disappeared when tested directly in a trial designed for the question.
A post-hoc subgroup difference can almost always be explained after the fact, because plausible mechanisms are easy to construct. Explanation offered after a finding adds much less than it feels like it does.
For generating the next trial's primary question, and for checking whether an overall result is driven entirely by one group. Both are real uses and neither supports acting on the subgroup result alone.
Whether the subgroup was pre-specified, how many were examined, whether an interaction test was reported, and whether anyone has since tested it directly.
Published checklists assess whether a subgroup finding is believable: was it one of few, pre-specified, supported by an interaction test, consistent across studies, and biologically plausible in advance. Running a claimed subgroup effect through those five questions usually resolves it quickly.
Unless a subgroup finding meets those criteria, the overall estimate remains the best prediction for every subgroup including the one that looked different. That is counterintuitive and it is the position the statistics support.
The claims on this page about slicing a trial after the fact, and why the slices so often fail to replicate can be checked against sources outside this site: