mediation analysis

A statistical method that splits an association between two things into the part running through a third thing and the part that does not, and it rests on assumptions that cannot be checked.

Tönnies and colleagues, writing in Deutsches Ärzteblatt International, describe the goal of mediation analysis as expressing an overall exposure effect as a combination of an indirect and a direct effect. The indirect effect is the part that runs through the mediator, the go-between factor under study; the direct effect is everything left over. Their example is obesity and type 2 diabetes, with insulin resistance as the candidate mediator: the question is not only whether obesity raises diabetes risk but how much of that risk arrives by way of insulin resistance. The share is reported as the proportion mediated, defined as the ratio of the indirect effect over the total effect, and it is usually printed as a percentage.

The method asks for more than an ordinary effect estimate does. Angriman and colleagues, in JACC: Advances, list three separate no-unmeasured-confounding assumptions a causal reading requires: none between exposure and outcome, none between exposure and mediator, and none between mediator and outcome, plus no mediator-outcome confounder that is itself affected by the exposure. Tönnies and colleagues make the consequence explicit: because three sets of confounders must be adjusted for rather than one, mediation analysis is more prone to bias than a study that estimates the total effect alone, and in observational data it is usually not possible to verify whether the assumption of no unknown confounders holds at all.

Timing matters as much as adjustment. The same authors note that a valid analysis wants a clear chronological order, exposure first, then mediator, then outcome, ideally measured at three separate points, because otherwise it may not be guaranteed that the mediator follows the exposure rather than the other way round. So a sentence reporting that one factor explained some percentage of an association is making a claim about mechanism on top of the claim about association, and it is the more fragile of the two. Read it as a decomposition the authors have modelled, not as a quantity anyone observed.

Sources