Purpose: Interaction and subgroup analyses remain controversial topics in epidemiology. A recent theoretical paper suggested that a combination of no overall treatment-outcome association and treatment effect limited to a single subgroup would imply a clinically implausible interaction, with opposite treatment effects in the two subgroups. However, this argument was based entirely on point estimates and ignored sampling error and statistical inference.
Methods: We simulated hypothetical studies in which treatment truly affected the outcome in only one subgroup, with no effect in the other subgroup. We generated 1000 random samples for three study designs (small clinical study, case-control, and large cohort), and different values of total sample size (N), relative size of the affected subgroup, and treatment effect. We estimated the frequency of significant results for tests of overall and subgroup-specific treatment effects, and treatment-by-subgroup interaction.
Results: Combination of statistically non-significant overall treatment effect and significant treatment-by-subgroup interaction occurred frequently, especially if the affected subgroup was proportionally smaller, even in studies with high power to detect the overall effect (e.g. in 37.1% of samples with N = 20 000, with 600 outcomes, and an effect (odds ratio of 1.5) limited to 30% of subjects). Furthermore, in most samples with a significant interaction, subgroup analyses correctly indicated that the significant effect was limited to one subgroup.
Conclusion: In studies where the treatment truly affects the risks in only one subgroup, a non-significant overall effect will often coincide with a statistically significant treatment-by-subgroup interaction. Thus, a non-significant overall effect should not prevent testing plausible interactions.
Keywords: interaction test; pharmacoepidemiology; simulations; subgroup analysis.
Copyright © 2013 John Wiley & Sons, Ltd.