A trial misses its primary endpoint but one subgroup shows a large benefit. How do you interpret it?

Why

With many subgroups, some will show large effects by chance, and the smaller the subgroup, the wider the swing. EMA's 2019 guideline on subgroups in confirmatory trials describes this case as using a subgroup to rescue a trial that has formally failed. It says that from a formal statistical point of view no further confirmatory conclusions are possible, and that one or more additional trials should usually be conducted.

Credibility rests on three things. Was the subgroup pre-specified with a stated reason? Is there a biological mechanism, such as a biomarker linked to the drug's target?

Is the effect replicated in another trial? Also check the complementary group: if the rest of the population shows harm, the pattern may just be noise splitting around zero. Look at the interaction test rather than the subgroup p value alone.

The guideline allows rare exceptions, such as very large cardiovascular trials where a subgroup holds a lot of randomized evidence and repeating the trial is not feasible. Even then the bar is high and the rationale must be explicit.