[Durgical](https://durgical.com/en/) › [DurgiPedia](https://durgical.com/en/durgipedia/) › [Epidemiology](https://durgical.com/en/epidemiology/) › [Bias, confounding and causation](https://durgical.com/en/epidemiology/areas/bias/) › Effect modification (interaction)

# 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.

Updated Nov 5, 2026

## Key points

- The effect of the exposure differs between strata of a third factor.
- It is not a bias: report by stratum; do not adjust it away or average it.
- Confounding: strata alike, different from the crude. Modification: strata different from each other.
- The same factor can be both a confounder and a modifier.
- It depends on the scale: there can be interaction on the multiplicative scale (RR) and not on the additive scale (risk difference), or the reverse.

## 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.

## Memory aid

With confounding the strata agree with each other against the crude; with interaction the strata argue with each other.

## 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.

[Add to my cards (3)](https://durgical.app/estudiar/?add=/en/epidemiology/effect-modification/)

## Connections

## Related 5

- [Confounding](https://durgical.com/en/epidemiology/confounding/)
- [Controlling confounding](https://durgical.com/en/epidemiology/controlling-confounding/)
- [Chance, bias, confounding or cause](https://durgical.com/en/epidemiology/chance-bias-confounding-or-cause/)
- [Heterogeneity (I²)](https://durgical.com/en/epidemiology/heterogeneity/)
- [Multiple comparisons](https://durgical.com/en/epidemiology/multiple-comparisons/)

[← All of epidemiology](https://durgical.com/en/epidemiology/)

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](https://www.cochrane.org/authors/handbooks-and-manuals/handbook/current/chapter-10) (Cochrane, 2024); [Chapter 25: Assessing risk of bias in a non-randomized study](https://www.cochrane.org/authors/handbooks-and-manuals/handbook/current/chapter-25) (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.

---
Source: https://durgical.com/en/epidemiology/effect-modification/
License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Cite as: DurgiPedia (durgical.com), https://durgical.com/en/epidemiology/effect-modification/