Effect modification (interaction)
- Concept
The exposure has a different effect depending on a third factor (age, sex, genes). It is not an error: it is a finding to report.
What it is
Effect modification (interaction) is present when the size of the association between exposure and outcome changes across levels of a third factor. Unlike confounding, it is not a bias to remove but a real feature, described by reporting the measure in each stratum. It is recognised because the stratum estimates clearly differ from each other; with confounding, the strata resemble each other and differ from the crude estimate.
Worked example
- Worker cohort: night work (exposed) and hypertension over 5 years, with 500 people per group in each age stratum.
- Under 45: 15 of 500 exposed (3%) and 15 of 500 unexposed (3%): RR = 1.0.
- 45 or older: 75 of 500 exposed (15%) and 25 of 500 unexposed (5%): RR = 3.0.
- Crude: 90 of 1 000 exposed (9%) versus 40 of 1 000 (4%): RR ≈ 2.25.
- The stratum RRs (1.0 and 3.0) differ sharply: age modifies the effect.
- Reading: a single RR of 2.25 would hide that the risk appears only in people aged 45 or older.
Confounding versus effect modification
| Feature | Confounding | Effect modification |
|---|---|---|
| What it is | Distortion by a third factor | A real effect that differs by the third factor |
| Strata | Alike, different from the crude | Different from each other |
| What to do | Adjust and give one summary measure | Report each stratum separately |
| Is it an error? | Yes, a bias to control | No, a finding to describe |
Markers
- Interaction
- Strata
In practice
- A treatment can help one subgroup and not another; that is why trials define their subgroups in advance.
- Unplanned subgroup analyses generate hypotheses; many ‘interactions’ found that way are chance.
Common pitfalls
- Pooling with Mantel-Haenszel strata whose effects are very different.
- Declaring interaction because one stratum is significant and another is not: the estimates must be compared directly with a test for interaction.
Test yourself
Think of the answer, then tap the question to see it.
What is effect modification?
When the size of the association between exposure and outcome changes across levels of a third factor.
Crude RR 2.25; RR 1.0 in the young and 3.0 in the old. Confounding or effect modification?
Effect modification: the strata clearly differ from each other; each is reported separately.
Should effect modification be adjusted for or removed?
No. It is not a bias but a finding; it is described by reporting the measure in each stratum.
Connections
Educational content on epidemiological and statistical methods; the examples use made-up numbers to teach the calculation. Not a clinical or legal guide: current Colombian rules are in the official sources.
Sources: Chapter 10: Analysing data and undertaking meta-analyses (Cochrane, 2024); Chapter 25: Assessing risk of bias in a non-randomized study (Cochrane, 2024).
Spotted a mistake?We are always improving for the community. Tell us what is wrong and our team will review it.Report a mistake
Representative model, for educational purposes only.