It is common practice, at least in large-scale trials, to conduct subgroup analyses to determine treatment effect differences for various subsets of the entire study population. They are very common in trials that do not demonstrate an overall favorable trend. Such post-hoc explorations of the data are sometimes employed to find one or more subgroups in which the treatment "really works." What is often ignored is the mathematical fact that in any real or constructed data set with a trend in either direction, it is easy to single out at least one subgroup for which the difference reaches a conventional level of statistical significance.3 The medical literature is replete with unconfirmed subgroup findings. Such post-hoc results should not be interpreted as conclusive. They may, however, be of value for hypothesis generation.
As an example of post-hoc analyses, take the case of a placebo-controlled trial of a calcium channel blocker in patients with acute myocardial infarction.2 Although no overall mortality benefit from the active intervention was observed, a positive "trend" in the findings persuaded the investigators to perform subgroup analyses, which showed a significant reduction in mortality for infarction patients with normal myocardial function. The publication failed to report that mortality correspondingly increased in patients with impaired myocardial function, suggesting a negative inotropic drug effect. Additionally, this analytic exercise was not defined a priori. It should not be surprising that this post-hoc finding has yet to be confirmed in another calcium channel blocker trial.
The ISIS-2 trial tested the effects of streptokinase and aspirin, individually and in combination, on short-term mortality in patients admitted with acute myocardial infarction. The trial demonstrated mortality benefits for both active interventions. In a study of the potential fallacy of subgroup analyses, the investigators observed that patients born under the Zodiac signs of Gemini and Libra exhibited a 5% higher mortality on aspirin compared to placebo, whereas those born under the other Zodiac signs had a 30% lower mortality on aspirin compared to placebo.1 There is no plausible biological explanation for this observation, which nicely illustrates the pitfalls of post-hoc subgroup analysis.
Was this article helpful?