Effect Modification (Interaction)

Effect modification (closely related to statistical interaction) means the effect of an exposure genuinely varies depending on another variable — the modifier. For instance, a treatment might cut risk substantially in younger people but little in older people. Unlike confounding, effect modification is not a bias to be removed; it is a real feature of the data worth reporting, because it tells you for whom an exposure matters.

It matters in population science and health equity because average effects can mask important heterogeneity across age, sex, race/ethnicity, or socioeconomic groups. Identifying modifiers helps target interventions to those who benefit most and reveals where a one-size-fits-all program may fail.

A key subtlety is that effect modification is scale-dependent: an exposure can show interaction on the additive (risk-difference) scale but not the multiplicative (ratio) scale, or vice versa. Always state the scale. Another pitfall is confusing effect modification with confounding — modifiers are reported via stratified or interaction analyses, whereas confounders are adjusted away. Subgroup analyses also demand caution, since testing many subgroups inflates false positives.

For example, if a screening program reduces late-stage cancer risk by 5 percentage points in well-insured patients but by only 1 point in uninsured patients, insurance status modifies the program's effect.