Effect Modification and Interaction

Effect modification occurs when the magnitude of an exposure's effect on an outcome differs across levels of a third variable — for example, when a treatment helps younger people more than older ones. Unlike confounding and bias, effect modification is not an error to be removed but a real feature of the relationship that is worth describing. The closely related idea of interaction concerns the joint effect of two exposures acting together.

Definition

Effect modification (effect-measure modification) is present when the measure of an exposure's effect on an outcome varies across strata of a third variable; interaction refers more specifically to the joint causal effect of two exposures relative to their separate effects.

Scope

The entry covers the definition of effect modification, its dependence on the choice of effect measure (additive versus multiplicative scale), the distinction between effect modification and interaction, and why both differ from confounding. It is a methodological reference and does not provide clinical guidance.

Core questions

• Does the size of the exposure effect differ across levels of a third variable?
• On which scale — additive or multiplicative — is modification being assessed?
• Is the question about effect modification by a covariate or about interaction between two exposures?
• Is the apparent heterogeneity genuine, or an artefact of confounding or chance?

Key concepts

• Effect-measure modification
• Additive versus multiplicative scale
• Interaction (joint effects)
• Stratum-specific effects
• Qualitative versus quantitative modification
• Synergism and antagonism
• Distinction from confounding

Mechanisms

Effect modification is assessed by comparing the exposure's effect across strata of a modifier; crucially, whether modification is present can depend on the effect measure. An effect that is constant across strata on the risk-difference (additive) scale may vary on the risk-ratio (multiplicative) scale, and vice versa, so a claim of effect modification must specify the scale. Interaction, in the causal sense, concerns the joint effect of two exposures and whether their combined effect departs from what their separate effects would predict (for instance, additive interaction relevant to public-health impact). VanderWeele and Robins (2007) used directed acyclic graphs to classify distinct types of effect modification, and VanderWeele (2009) clarified that effect modification (heterogeneity of an effect across a covariate) and interaction (joint effects of two manipulable exposures) are related but not identical. None of these is confounding: confounding is a distortion to be controlled, whereas effect modification and interaction describe the structure of the effects themselves.

Clinical relevance

Recognising that an effect can differ across subgroups is important for interpreting whether evidence applies uniformly, and additive interaction informs how effects combine at the population level. The concept explains how effects are described and compared across groups; it is not a prescription for any individual's care.

Epidemiology

Effect modification and interaction are routinely examined through stratified analyses and product terms in regression models across observational and experimental studies. Because scale dependence and multiple-testing concerns complicate subgroup analysis, methodologic guidance emphasises pre-specification and cautious interpretation.

History

Epidemiology long distinguished real effect heterogeneity from confounding, and the scale-dependence of interaction (additive versus multiplicative) was emphasised in twentieth-century methodologic writing. In the 2000s, causal-inference work sharpened the concepts: VanderWeele and Robins (2007) gave a graphical classification of effect modification, and VanderWeele (2009) formalised the distinction between effect modification and interaction, linking both to potential-outcomes and causal-diagram frameworks.

Debates

Are effect modification and interaction the same thing?

The terms are often used interchangeably, but causal-inference accounts treat effect modification as heterogeneity of a single exposure's effect across a covariate, and interaction as the joint effect of two manipulable exposures; the distinction matters for interpretation and for which causal conclusions are warranted.

Which scale should interaction be measured on?

Because additive and multiplicative scales can disagree about whether modification exists, there is long-standing discussion over which is more relevant — additive measures are often argued to be more informative for public-health impact, while multiplicative measures arise naturally from common regression models.

Key figures

• Tyler VanderWeele
• James Robins
• Sander Greenland
• Kenneth Rothman

Seminal works

• vanderweele-2007
• vanderweele-2009

Frequently asked questions

How is effect modification different from confounding?

Confounding is a distortion of the overall association that should be removed; effect modification is genuine variation in the effect across subgroups and is a property of the relationship to be described, not eliminated.

Why does effect modification depend on the scale?

An effect can be uniform across strata on the additive (risk-difference) scale yet vary on the multiplicative (risk-ratio) scale, or the reverse, so whether modification is present depends on which effect measure is used.

Are interaction and effect modification interchangeable terms?

They are closely related and often used loosely as synonyms, but in causal-inference terms effect modification refers to heterogeneity of one exposure's effect across a covariate, whereas interaction refers to the joint effect of two exposures.