Effect modification and interaction
The effect of an exposure on an outcome can differ across levels of a third variable. Reporting such a difference turns on keeping effect modification apart from statistical interaction and confounding, and a moderator apart from a mediator.
What effect modification is
Effect modification is present when a third variable modifies the association between an exposure and a disease, so that the effect of the exposure on an outcome differs because of the presence of the third factor. The classical case is study participants from different baseline sociodemographic or clinical strata who experience distinct outcomes in response to the same exposure, whether a risk factor or a treatment. Effect modification refers to disease biology and mechanisms, whereas interaction and confounding have mainly statistical meanings.
Moderators and mediators
Factors that affect the direction or strength of an association between an exposure, the independent variable, and an outcome, the dependent variable, are termed moderators. Gender, for example, is said to moderate the association between diabetes and major depressive disorder: women with type 1 or type 2 diabetes are more likely to experience major depressive disorder than men with diabetes.
Mediators, unlike moderators, may explain a disease or treatment mechanism. Mediators identify why and how treatments have effects, whereas moderators identify on whom and under what circumstances treatments have different effects.
Statistical interaction
Interaction may occur when the observed joint effects of risk factors are markedly lower (biased toward the null) or markedly higher (biased away from the null) than expected from their independent, additive effects. Operationally, interaction can be defined as joint effects of risk factors that exceed or fall below what their individual components lead one to expect, where the joint effects are expressed as a sum in additive models such as linear regression and as a product in multiplicative models such as logistic regression. The whole exceeds the sum or product of the parts in positive interaction, or synergy, and falls short of it in negative interaction, or antagonism.
Published cases of effect modification
Family history of allergy modifies the association between caesarean birth and childhood atopic sensitisation: children with a family history of allergy who are born by caesarean section are at greater risk of atopy than children born by caesarean section who have no such history. In a study of these factors, after adjustment for potential confounders, the odds ratio against vaginal birth was significantly raised for children with a family history of allergy (2.62; 95% CI 1.38 to 5.00) and not for those without (1.16; 95% CI 0.64 to 2.11), although the p value for effect modification was 0.06.
In another published study, the association between relative telomere length and progression of chronic kidney disease was significant in smokers and in participants with diabetes mellitus, and not significant in nonsmokers without diabetes, whose hazard ratio overlapped 1.0. After adjustment for gender, glomerular filtration rate, proteinuria and baseline age, each 0.1-unit decrease in relative telomere length was significantly associated with an increased relative risk of progression in smokers (hazard ratio 1.44; 95% CI 1.16 to 1.81; p = 0.001) and in patients with diabetes (hazard ratio 1.16; 95% CI 1.01 to 1.34; p = 0.03).
What studying effect modification serves
Studying effect modification makes it possible to:
- adjust or otherwise augment the precision of measures of association, such as hazard ratios adjusted for effect modifiers
- identify special populations who may derive especially marked benefits from different interventions
- help to formulate hypotheses about the etiologic pathways from exposure to disease
Reporting subgroup analyses
In 2008 the editors and statistical consultants of the New England Journal of Medicine published guidelines for reporting subgroup analyses of clinical trial data in that journal, and noted that many of the issues discussed also apply to observational studies.