Separate Mediators From Confounders Before Adjusting a Model

Start with the effect being estimated

The same variable can play different roles in different questions. A mediator lies on a causal pathway from exposure to outcome in the proposed model. A confounder helps explain a noncausal path between them. These roles require substantive and temporal reasoning; they are not determined by whether adjustment changes a coefficient.

Schisterman, Cole and Platt’s paper on overadjustment discusses how adjustment can obstruct estimation of a total effect. A model containing every available measurement can therefore be less appropriate than a deliberately chosen model, even when all measurements are precise.

Build two timelines

First map when the underlying processes occur: exposure initiation, intermediate response and outcome development. Then map when each process is actually measured. A measurement collected before the recorded exposure date may still reflect an earlier exposure history, while a late measurement may represent a pre-existing condition.

Write a short rationale for each covariate’s proposed role. If the aim is a total effect, identify which variables may transmit part of that effect. If a direct-effect question is intended, specify which direct effect and the additional assumptions it requires. Simply adding a mediator to an ordinary regression does not automatically identify a meaningful causal direct effect.

Compare a hypothetical pathway

Suppose an intervention A changes an intermediate behavior M, and M changes outcome Y. Baseline factor C influences both intervention receipt and the outcome. In the stipulated graph, C is a common cause relevant to confounding control, whereas M carries part of the effect under study.

An analysis adjusting for C addresses the illustrated backdoor pathway while preserving the A→M→Y pathway. An analysis additionally adjusting for M asks a different question and can introduce further bias if mediator–outcome common causes are not properly handled. A smaller adjusted coefficient is not sufficient evidence that the difference measures mediation.

Keep the models and interpretations aligned

Label models by their scientific purpose rather than calling the largest model fully adjusted. Record whether time-varying confounders may themselves be affected by earlier treatment, because that situation can require methods beyond standard regression adjustment.

In the report, state the target effect and explain why intermediate variables were excluded or analyzed separately. Describe uncertainty about causal ordering and measurement timing. If presenting exploratory mediation results, distinguish them from the primary total-effect estimate and avoid treating a percentage change in a coefficient as a universally valid mediated proportion. The reader should be able to connect the covariate set to the intended causal question.

Source verification: 6 November 2026. Examples are hypothetical.