Interpreting Interventional Effects as Mechanism Evidence

From an Effect Estimate to a Mechanism Claim

An interventional indirect effect answers a focused causal question: if exposure were held at a specified level, how would the expected outcome change if the mediator were instead drawn from one exposure-specific distribution rather than another? This can provide evidence that changing a mediator distribution matters for the outcome under the stated intervention. It does not, by itself, establish that a particular biological or social mechanism has been discovered.

The distinction matters because “the mediator distribution affects the outcome” and “this is the mechanism through which exposure operates” are not equivalent claims. The first concerns a defined contrast between interventions. The second often implies a detailed account of the process: which component is causally active, how it operates, whether the measured mediator represents it, and whether competing pathways have been excluded. An effect estimate alone does not answer all of those questions.

The stochastic draw \(G_{a^*}\) is not necessarily the mediator value that a particular person would naturally have under exposure \(a^*\). It represents a distributional intervention: conditional on baseline covariates, the mediator is generated according to the distribution associated with \(a^*\). Thus the effect describes what happens when that distribution is substituted, not what happens when every individual is assigned a fixed mediator value.

The contrast is indexed by the exposure level held fixed. For example, \(\mu(1,1)-\mu(1,0)\) compares mediator distributions while setting exposure to 1 in both terms. The outcome scale also matters: a contrast in mean outcome, risk difference, or another chosen scale has a distinct interpretation and units. Interpret the reported effect on the scale actually used.

What the Contrast Supports

If the contrast is credibly identified, its direction and magnitude support a causal statement about the particular mediator-distribution intervention. A positive contrast in a continuous outcome measured in points means that, at the fixed exposure level and under the specified intervention, drawing the mediator from the first distribution rather than the second increases the expected outcome by that many points. For a binary outcome on the risk-difference scale, the contrast is a difference in risks.

This is meaningful mechanism-related evidence when the mediator is substantively connected to a proposed process and the distributional shift corresponds to a plausible intervention. But the estimate does not uniquely identify a microscopic or social process. A measured mediator may be a marker of an unmeasured active process; several mediators may move together; and different interventions can produce the same mediator value while affecting the outcome through other routes. The estimated contrast pertains to the intervention as defined, not automatically to every way of changing the mediator.

Identification Comes Before Mechanism Language

An estimated contrast is not automatically a causal effect. For the simple identification formula below, assume consistency and a well-defined exposure and mediator-distribution intervention; conditional exchangeability of exposure given \(C\); no unmeasured mediator–outcome confounding after conditioning on \(A\) and \(C\); and positivity for the exposure and mediator values needed by the intervention. Also assume there is no exposure-induced mediator–outcome confounder omitted from the formula. In particular, the mediator distribution and outcome regression must be supported by the observed data in the relevant covariate strata. These conditions are substantive requirements, not consequences of fitting a model.

Under these assumptions, the identifying functional is

with integrals replacing sums for continuous variables. In words, average the outcome regression at exposure \(a\) over the mediator distribution under \(a^*\), then average over the baseline covariate distribution. This formula makes the target explicit: the mediator distribution changes, exposure in the outcome regression remains fixed, and baseline covariates are averaged over the target population.

The key calculation follows from averaging over the stochastic mediator draw. Conditional on \(C=c\), the expected outcome under exposure \(a\) and a draw from the \(a^*\) mediator distribution is

Consistency and the exchangeability assumptions allow the potential-outcome mean in this expression to be represented by the observed conditional mean \(E[Y\mid A=a,M=m,C=c]\). Averaging over \(P(C=c)\) gives the formula above. The justification also shows why an unmeasured common cause of \(M\) and \(Y\), or lack of support for relevant mediator values, can undermine a causal interpretation.

Reading a Null or Imprecise Estimate

A small estimated interventional indirect effect does not automatically prove that the mediator is irrelevant. The specified exposure contrast may produce only a small distributional change; the outcome may respond weakly over the mediator values shifted; measurement error may obscure the relevant feature; or uncertainty may be substantial. Conversely, a large estimate does not establish a mechanism if confounding, model misspecification, or unsupported extrapolation is plausible.

Building a Defensible Mechanism Argument

Mechanism claims become more credible when the estimated distributional intervention is connected to evidence beyond the contrast itself. Researchers should specify what intervention on the mediator distribution represents, show that the relevant mediator values are supported, and make clear which outcome scale and population the estimate concerns. They should also assess whether the mediator was measured adequately and whether plausible confounding or alternative pathways could explain the result.

Convergent evidence may include direct experimental manipulation of the proposed process, measurements that distinguish the proposed mediator from correlated markers, temporal evidence consistent with the proposed sequence, and replication under interventions that alter the process in different ways. Such evidence does not eliminate the need for careful causal assumptions; it strengthens the argument that the statistical intervention corresponds to the substantive mechanism being claimed.

A useful reporting discipline is to state the causal result first and the mechanistic interpretation second. For example: “At exposure level 1, substituting the mediator distribution estimated under exposure 1 for that estimated under exposure 0 increased expected outcome by 1.60 points, under the stated identification assumptions.” Then state whether the mediator is believed to represent a proposed mechanism and what evidence supports that interpretation. This wording keeps the statistical estimand separate from the broader scientific conclusion.

Common Mistake

Wrong reasoning: “The interventional indirect effect is nonzero, so the measured mediator is the mechanism through which exposure affects the outcome.”

Correction: A nonzero contrast supports an effect of the specified mediator-distribution intervention on the outcome, if identification assumptions are credible. It does not establish that the measured mediator is the unique active process, that all of its observed association is causal, or that other pathways are absent. State the intervention and assumptions, report the effect on its actual scale, and treat a specific mechanism as a further interpretation requiring additional evidence.