Effect modification, explained
Effect modification explained: why one exposure can carry different risks in different groups, how it differs from confounding, and how to report it.
By Manouchehr Hessabi, MD, MPH
Effect modification is present when the effect of an exposure on an outcome differs across levels of a third variable, such as age, sex, an existing illness or a genotype. Unlike confounding, it is not an error to be removed from a study. It describes something real about how an exposure relates to an outcome in different groups, and the methods literature treats it as a finding to report carefully.
Readers meet the idea often without the name: a medicine that appears to help one group more than another, or a gene variant that changes how a metal exposure relates to a health outcome. This explainer covers what effect modification is, how it differs from confounding and from interaction, why the answer depends on the scale used to measure it, and how a careful paper should present it. It is educational and not a substitute for personal medical advice.
What is effect modification?
The European Network of Centres for Pharmacoepidemiology and Pharmacovigilance (ENCePP) puts it plainly in its methodological guide: "Effect modification occurs when the effect of a single exposure on an outcome depends on the values of another variable" (ENCePP Methodological Guide, Chapter 7, read November 17, 2026). The other variable is called the effect modifier.
The guide also explains why the concept matters in practice: effect modification "is used to identify subpopulations that are particularly susceptible to the exposure of interest." In other words, the question is not only whether an exposure is associated with an outcome, but for whom the association is stronger or weaker.
Two pieces of vocabulary help here. The risk ratio (RR) divides the risk of the outcome in an exposed group by the risk in an unexposed group. The risk difference (RD) subtracts one from the other. Both appear below, because effect modification can look different depending on which one is used.
How is effect modification different from confounding?
Confounding occurs when a third factor is linked to both the exposure and the outcome and distorts the estimate of their relationship. It is a bias. Epidemiologists try to remove it by design, for example through randomization or matching, or by statistical adjustment. The site's earlier piece explains how confounding distorts a finding.
Effect modification is a different kind of thing. It is not a distortion of the truth but part of it. The same tool, a stratified analysis (splitting the data into groups, or strata, and estimating the effect within each), can reveal either one, but the pattern differs:
- Confounding: the estimates within each stratum are similar to each other, but they differ from the crude, unstratified estimate. Adjusting brings the overall estimate back in line.
- Effect modification: the estimates within the strata differ from each other. Averaging them into one adjusted number would hide a real difference.
The ENCePP guide makes the practical consequence explicit: factors that change the effect across groups "should be assessed as potential effect-modifying factors rather than confounding factors," and "treating these factors as confounders can induce bias."
Why does the answer depend on the scale?
This is the point most often missed. ENCePP states that "The presence of effect modification depends on which measure is used in the study (absolute or relative)." An exposure can show effect modification on one scale and none on the other.
A simple hypothetical example shows why. The numbers below are invented for illustration only and do not come from any study:
| Group | Risk if unexposed | Risk if exposed | Risk ratio | Risk difference |
|---|---|---|---|---|
| Group A | 2 percent | 4 percent | 2.0 | 2 percentage points |
| Group B | 10 percent | 20 percent | 2.0 | 10 percentage points |
On the multiplicative scale (risk ratios), the exposure doubles risk in both groups, so there is no effect modification. On the additive scale (risk differences), the exposure adds 2 percentage points of risk in Group A and 10 in Group B, so there is strong effect modification. Both statements are correct. They answer different questions. The distinction is the same one drawn in the explainer on relative versus absolute risk.
Which scale matters? Mirjam Knol and Tyler VanderWeele, in recommendations published in the International Journal of Epidemiology, argue that "For public health purposes, the additive scale indicates whether the effect of a risk factor would be greater in one subpopulation," which makes it useful for targeting groups and allocating resources (Knol and VanderWeele, 2012). They also found that additive measures were "especially likely to be omitted" from published analyses.
Not every methodologist frames it the same way. In a commentary in the same journal, David Clayton argued that the case for the additive scale's special public health relevance is overstated, and cautioned that a significant interaction term in a regression model, in general, "has few implications for mechanism" (Clayton, 2012). He also noted that interaction parameters are estimated much less precisely than main effects. The practical lesson from both sides is the same: say which scale you used, and do not read biology into a statistical term on its own.
Is effect modification the same as interaction?
The two words are often used interchangeably, but Knol and VanderWeele separate them:
- Effect modification asks how the effect of one exposure varies across strata of another factor. Only one set of confounders needs to be controlled: those for the primary exposure.
- Interaction asks about the joint effect of two exposures together. Because both are treated as causes, confounders of each exposure must be controlled.
ENCePP adds a related distinction: "Contrary to effect modification, interaction is a symmetric concept." In an interaction analysis, A modifying B and B modifying A are the same question. In effect modification, one variable is the exposure of interest and the other is the context.
Gene-environment research is a common setting for both ideas, since a genotype can be studied as a modifier of an environmental exposure. The site's piece on gene-environment interaction and detoxification genes discusses that kind of study in more depth.
How should a study report effect modification?
Knol and VanderWeele propose a four-part template for an exposure A, an outcome D and a modifier X:
- Report risk ratios, odds ratios or risk differences, with confidence intervals, for each combination of A and X, using a single reference category, ideally the one with the lowest risk.
- Report the effect of A within each stratum of X.
- Report measures of effect modification on both the additive and the multiplicative scale, with confidence intervals and p-values.
- List the confounders adjusted for in the relationship between A and D.
They also recommend reporting risk ratios rather than odds ratios when possible, because when the outcome is common "using ORs may exaggerate interaction measures."
Their worked example, drawn from an earlier cohort analysis by Knol and colleagues, looked at antidepressant use and diabetes, with a chronic disease score (0 versus 1 or more) as the potential modifier. The risk ratios were adjusted for age, sex and benzodiazepine use. Using people with a score of 0 and no antidepressant use as the reference group:
- Score 0, antidepressant use: RR 1.16 (95% confidence interval 0.95 to 1.42)
- Score 1 or more, no antidepressant use: RR 1.55 (1.30 to 1.84)
- Score 1 or more, antidepressant use: RR 2.07 (1.73 to 2.47)
Within strata, the risk ratio for antidepressant use was 1.16 (0.95 to 1.42) among people with a score of 0 and 1.34 (1.13 to 1.58) among those with a score of 1 or more.
The additive measure is the relative excess risk due to interaction (RERI), calculated here as 2.07 minus 1.55 minus 1.16 plus 1, which equals 0.36 (95% CI minus 0.003 to 0.73, P = 0.052). The multiplicative measure, the ratio of risk ratios, was 1.15 (0.89 to 1.49, P = 0.282).
Read together, these numbers show why both scales are reported. The additive estimate points toward a larger association in the group with a higher chronic disease score, and the authors read it as an indication of positive additive effect modification, though its confidence interval just crosses zero. The multiplicative estimate is closer to no modification and more uncertain. A reader shown only one of the two would come away with a different impression. This example is presented to illustrate reporting methods, not as guidance about any medication.
What should a careful reader watch for?
Claims of effect modification often come from subgroup analyses, and these deserve caution. Splitting a study into many groups shrinks the sample in each, and, as Clayton notes, interaction parameters are estimated much less precisely than main effects. Testing many subgroups also raises the chance that one will look different by chance alone.
A few questions help when reading a paper that reports a difference between groups:
- Which scale was used? A finding on the risk ratio scale says nothing by itself about the risk difference scale, and the reverse.
- Were both measures reported? Knol and VanderWeele's template asks for both, with confidence intervals.
- What is the reference group? Estimates for every stratum should share one reference category.
- Was the subgroup planned? A modifier specified before the analysis, with a plausible reason behind it, is more credible than one found after the data were sliced many ways.
- Is the language proportionate? A statistical difference between strata is a description of the data, not proof of a biological mechanism.
A measured close
Effect modification is one of the more useful ideas in epidemiology because it moves the question from "is this exposure associated with harm?" to "for whom, and on what scale?" It is also easy to misread: it is not confounding, its presence depends on the measure chosen, and a single subgroup result rarely settles anything.
The best protection is transparent reporting, of the kind Knol and VanderWeele lay out, so that readers can see both the additive and the multiplicative picture for themselves. Readers interested in the underlying research can find Dr. Hessabi's peer-reviewed publications.