Mediation Analysis: How a Study Tries to Show the Pathway, Not Just the Effect

Mediation analysis tries to open the black box between a cause and an effect by asking how much of the effect travels through a specific intermediate step, the mediator, and how much reaches the outcome by other routes. It splits a total effect into an indirect effect through the mediator and a direct effect that does not. The split is only believable when several confounding assumptions hold, and the one about the mediator and the outcome is the hardest to defend, because the mediator was almost never randomized.

The question mediation asks

Sometimes knowing that a treatment works is not enough, and you want to know why. Does a blood pressure drug prevent strokes because it lowers pressure, or partly through some other action? Does an education program improve health because it changes behavior, or through income? Mediation analysis is the formal attempt to answer that kind of how question by singling out an intermediate variable, the mediator, that sits on the path between the exposure and the outcome.

The appeal is obvious. If most of an effect runs through one modifiable step, that step becomes a target. But the appeal is also the trap, because a convincing story about mechanism can outrun the evidence that actually supports it.

Direct and indirect effects

The core idea is a decomposition. The total effect of an exposure on an outcome is split into two pieces. The indirect effect is the part that flows through the mediator, the amount the outcome changes because the exposure changed the mediator, which in turn changed the outcome. The direct effect is everything else, the part of the effect that does not go through that particular mediator.

Modern causal mediation defines these as natural direct and natural indirect effects, which lets the method handle realistic situations that older approaches could not, such as a binary outcome or an interaction where the exposure and the mediator work partly together. The older product-of-coefficients approach, familiar from many textbooks, is a special case that assumes a linear model and no such interaction.

The assumptions that make or break it

A mediation estimate has a causal meaning only if a set of no-unmeasured-confounding conditions holds. In plain terms, four things must be true. There must be no uncontrolled confounding of the exposure and the outcome, none of the mediator and the outcome, and none of the exposure and the mediator. And there must be no confounder of the mediator and the outcome that is itself caused by the exposure.

That last condition sounds technical but it is the practical heart of the problem. Even in a randomized trial, only the exposure was randomized, not the mediator. So a hidden variable that influences both the mediator and the outcome can make an ordinary association look like a pathway. Randomizing the treatment does not rescue you here, which is why mediation from a clean trial can still mislead about mechanism.

Where mediation misleads

The commonest failure is unmeasured mediator-outcome confounding, described above. Close behind are reverse timing, where the outcome actually preceded or drove the supposed mediator, and measurement error in the mediator, which tends to shrink the estimated indirect effect and inflate the direct one. Ignoring an exposure-mediator interaction can also distort both pieces.

There is also a presentation problem. The proportion mediated, a tidy percentage that says how much of the effect runs through the mediator, is unstable. It can swing widely with small changes in the model, it behaves strangely when direct and indirect effects point in opposite directions, and it is easy to quote out of context. Treat a lone proportion-mediated figure as a headline, not as a conclusion.

What good reporting looks like

There is now a reporting guideline built specifically for this, called AGReMA, for trials and observational studies that use mediation analysis. It asks authors to state the causal model and the assumed diagram up front, to define the mediator and the timing clearly, to say which effects were estimated and under what assumptions, to address exposure-mediator interaction, and to report a sensitivity analysis for how much unmeasured confounding would overturn the result.

As a reader, you can use that same list as a lens. If a paper claims a mechanism but never states its confounding assumptions, never mentions timing, and offers no sensitivity analysis, the pathway is being asserted more than shown.

Reading it in practice

When you meet a mediation claim, first draw the arrows the authors are proposing: exposure to mediator to outcome. Then ask what could cause both the mediator and the outcome that was not accounted for, because that is where the estimate is most likely to be borrowed from confounding. Ask whether the mediator was measured before the outcome, and whether the analysis allowed the exposure and mediator to interact.

Held to those questions, mediation analysis is a legitimate and useful tool for generating and refining mechanistic hypotheses. What it rarely does on its own is prove that a mechanism is real, and the more decisively a paper claims to have nailed the pathway, the more carefully its assumptions deserve to be read.