Ask Mallard

Effect modification and interaction

Effect modification means that the magnitude of an exposure or treatment effect differs across levels of another variable; statistical interaction is how such heterogeneity is represented on a chosen effect scale.

What question does this help answer?

“Does the association or treatment effect differ meaningfully across subgroups or values of another variable?”

Why it matters

Average effects can hide clinically important heterogeneity. However, subgroup analyses are easy to overinterpret, especially when many are examined after results are known.

How to think about it

Effect modification depends on the effect measure. A factor may modify an effect on the risk-difference scale but not on the relative-risk scale. Interaction terms in regression models test a specific scale and model form.

The same effect on two scales

Two groups with the same risk ratio and different risk differences. In the lower-risk group risk falls from 10 to 5 in 100; in the higher-risk group from 40 to 20 in 100. The risk ratio is 0.50 in both, but the risk difference is 5 in 100 in one and 20 in 100 in the other.No modification on the ratio scale, clear modification on the difference scaleLower-risk group10 in 100Untreated5 in 100TreatedRisk ratio 0.50Risk difference 5 in 100Higher-risk group40 in 100Untreated20 in 100TreatedRisk ratio 0.50Risk difference 20 in 100Whether a factor modifies an effect depends on the scale; say which scale an interaction term tests.
Worked arithmetic. The inputs are round illustrative numbers, not results from any study; the outputs are calculated from them.

How Mallard may use this concept

Mallard may recommend a prespecified interaction when the clinical question explicitly asks whether an effect differs by age, sex, baseline risk, site, disease severity, or another scientifically motivated characteristic.

Good practice

  • Define important subgroups before looking at results when possible.
  • Use an interaction test rather than comparing “significant” in one group with “not significant” in another.
  • Report stratum-specific estimates with confidence intervals.
  • State the effect scale.
  • Limit exploratory subgroup fishing.
  • Consider power: interaction tests often require larger samples than main-effect tests.

Common mistake

A p-value below 0.05 in one subgroup and above 0.05 in another does not by itself demonstrate effect modification.

Clinical example

A treatment may reduce absolute cardiovascular risk more among patients with high baseline risk even if the relative effect is similar across risk groups.

Recent studies that used this
Searching PubMed…
Related Learn pages
Sources & further reading
  1. Effect modification, interaction and mediation: an overview of theoretical insights for clinical investigators (opens in a new tab) · Clin Epidemiol · doi 10.2147/CLEP.S129728
  2. Recommendations for presenting analyses of effect modification and interaction (opens in a new tab) · Int J Epidemiol · doi 10.1093/ije/dyr218
Last source review: 2026-09-05
AI-generated educational content
Mallard uses reputable methodological and reporting sources, but this material may contain errors or omit important context. Verify important methodological decisions with the cited sources and, when appropriate, a qualified research methodologist or statistician.