Mediation Analysis in Health Research: What Direct and Indirect Effects Really Require

What Is a Mediator Variable?

A mediator lies on a causal pathway from an exposure or treatment to an outcome. In the simplest structure:

The exposure affects the mediator, and changes in the mediator in turn affect the outcome. The part of the exposure effect operating through the mediator is described as an indirect or mediated effect; the part not operating through that mediator is described as a direct effect (Lash et al., 2021).

“Direct,” however, does not mean biologically direct or free of all mechanisms. It means not through the particular mediator being considered. Other unmeasured or unmodeled intermediate pathways may still contribute to what is labeled the direct effect (Lash et al., 2021).

Hernán and Robins similarly describe a mediator as a variable affected by treatment that affects the outcome. When the total causal effect is the estimand, conditioning on such a variable blocks the component of the treatment effect operating through it. Thus, adjustment for a mediator is not ordinary confounding control: it changes which causal effect is being targeted (Hernán & Robins, 2020).

Total Effect vs Direct Effect vs Indirect Effect

Mediation analysis begins by distinguishing causal questions that are easily blurred in ordinary regression.

Total effect

The total effect concerns the overall causal effect of changing the exposure on the outcome through all causal pathways.

If some of that effect operates through a mediator, the total effect includes that pathway rather than conditioning it away. This is why adjustment for a genuine mediator is inappropriate when the target estimand is the total causal effect (Hernán & Robins, 2020).

Direct effect

A direct effect concerns the part of the exposure effect that does not operate through the specified mediator.

But there is not just one universally applicable definition of “the direct effect.” Modern causal mediation analysis distinguishes, among other quantities, controlled direct effects and natural direct effects (Lash et al., 2021).

Indirect effect

An indirect effect concerns the portion of the exposure effect operating through changes in the mediator.

Under appropriate counterfactual definitions and assumptions, a total effect can be decomposed into a natural direct effect and a natural indirect effect. On an additive effect scale, Lash et al. describe this decomposition as:

The important point is that these quantities are causal estimands defined by hypothetical interventions and counterfactual outcomes. They are not automatically equivalent to whichever coefficients happen to emerge from two regression models.

Why Coefficient Attenuation Is Not Proof of Mediation

A traditional epidemiologic approach to mediation fits an outcome model without the mediator and then another outcome model containing the mediator. If the exposure coefficient becomes smaller after the mediator is included, the reduction is sometimes interpreted as an indirect effect and the remaining exposure coefficient as a direct effect.

Lash et al. explicitly describe this traditional “difference method,” while immediately emphasizing that the interpretation of those quantities as direct and indirect effects depends on additional assumptions. For continuous outcomes and mediators under ordinary least-squares linear models, the traditional difference and product-of-coefficients approaches coincide mathematically. That equivalence does not generally carry over to other settings; for example, the methods can diverge with logistic regression (Lash et al., 2021).

More fundamentally, coefficient attenuation does not establish the required causal structure.

The change could reflect confounding, model specification, interaction, scale, or other features of the data-generating and statistical models. If the mediator–outcome relationship is confounded, even the apparent direction of the mediator's association with the outcome can be misleading (Lash et al., 2021).

Causal Mediation Analysis Is a Counterfactual Problem

Modern mediation analysis becomes clearer when direct and indirect effects are defined using counterfactual outcomes.

Lash et al. introduce three key counterfactual quantities. In simplified notation:

Y_x

The outcome that would occur if exposure were set to x.

M_x

The mediator value that would occur if exposure were set to x.

Y_xm

The outcome that would occur if exposure were set to x and the mediator were set to m.

These counterfactuals allow direct and indirect effects to be defined in terms of interventions rather than merely associations among observed variables (Lash et al., 2021).

This framing exposes why causal mediation analysis is demanding: some mediation estimands require reasoning about combinations of counterfactual conditions that cannot simultaneously be observed for the same individual.

Controlled Direct Effects

A controlled direct effect asks what effect changing the exposure would have if the mediator were intervened upon and fixed at a particular value.

What would be the effect of changing the exposure if everyone’s mediator were held at m?

This estimand measures the exposure effect not operating through changes in that mediator under the specified intervention. Controlled direct effects can therefore be useful when the scientific question concerns what exposure effect would remain if a mediator could actually be fixed or intervened upon (Lash et al., 2021).

Natural Direct and Indirect Effects

A natural direct effect asks about changing the exposure while setting the mediator to the value it would naturally have taken under a reference exposure condition.

A natural indirect effect instead compares outcomes under mediator values that would naturally arise under different exposure conditions while holding the exposure at a specified level.

These definitions permit decomposition of the total effect into natural direct and indirect components under the required conditions (Lash et al., 2021).

But this conceptual appeal comes at a cost.

Hernán and Robins emphasize that a natural direct effect is a cross-world quantity: its definition combines a counterfactual outcome under one exposure condition with a mediator value that would arise under another exposure condition. They note that such an effect cannot be identified in a randomized experiment simply by randomizing treatment, and that estimates of natural direct effects depend on strong assumptions that cannot be empirically verified merely by having randomized the exposure (Hernán & Robins, 2020).

Lash et al. likewise discuss the controversial cross-world independence assumption involved in identifying natural effects and describe alternative interventional definitions that can sometimes be identified under weaker assumptions (Lash et al., 2021).

The Mediation Assumptions Are Stronger Than “No Confounding”

For the conventional causal interpretation of direct and indirect effects, Lash et al. identify four important confounding requirements.

Lash et al. summarize these as assumptions A1–A4: control exposure–outcome, mediator–outcome, and exposure–mediator confounding, plus the requirement that mediator–outcome confounders are not themselves exposure-induced (Lash et al., 2021).

Mediator–Outcome Confounding Is a Critical Threat

Suppose a third variable affects both the proposed mediator and the outcome.

Then an observed mediator–outcome association does not by itself tell us what would happen to the outcome if the mediator were changed. Part or all of the association could arise from their shared causes.

This issue persists even in randomized trials.

Randomizing the exposure can remove exposure–outcome and exposure–mediator confounding, but it does not automatically randomize the mediator. Mediator–outcome confounding may therefore remain, and causal direct and indirect effects can still be biased (Lash et al., 2021).

Exposure-Induced Mediator–Outcome Confounding Is Even More Difficult

An especially challenging structure occurs when exposure affects another variable that subsequently influences both the mediator and outcome.

Here L is simultaneously downstream of exposure and a confounder of the mediator–outcome relationship.

Simply adjusting for L does not solve the problem in the same way that adjusting for an ordinary baseline confounder would. Lash et al. note that standard regression approaches are no longer sufficient for natural-effect estimation in such settings; identification of direct and indirect effects becomes substantially more difficult, and specialized causal methods may be required (Lash et al., 2021).

With multiple mediators, the identification problem can become still harder. When another mediator precedes and affects the mediator of interest, direct and indirect effects through the selected mediator are not generally identified without further strong assumptions, even when all relevant variables have been measured (Lash et al., 2021).

Interaction Changes the Meaning of the Decomposition

Mediation and interaction are distinct concepts, but they can occur together.

The exposure may both:

1. change the mediator; and
2. modify how the mediator affects the outcome.

When exposure–mediator interaction is present, simple interpretations based on subtracting regression coefficients become inadequate. Causal mediation methods can accommodate exposure–mediator interaction, but the direct and indirect effect formulas then depend on that interaction as well as on the required confounding and model-specification assumptions (Lash et al., 2021).

Temporal Ordering Is Part of the Causal Argument

A mediator cannot plausibly carry the effect of an exposure that occurs after it.

Lash et al. emphasize that the assumptions required for causal mediation effectively contain this temporality requirement. Study design should therefore establish, as far as possible, that the exposure precedes the mediator and the mediator precedes the outcome (Lash et al., 2021).

Why cross-sectional mediation is particularly problematic

When exposure, mediator, and outcome are assessed at approximately the same time, statistical modeling cannot generally establish whether:

• Exposure → Mediator → Outcome
• Mediator → Exposure
• feedback or common causation generated the observed associations.

Lash et al. make the problem especially clear: without temporal ordering or additional causal knowledge, cross-sectional data may not allow researchers to distinguish mediation from confounding. Statistical techniques alone cannot determine which causal ordering generated the observed associations (Lash et al., 2021).

Even measuring exposure, mediator, and outcome sequentially does not eliminate every problem. Earlier values of the mediator or outcome may themselves confound later relationships. Where variables change over time, measurement of prior exposure, mediator, and outcome values can make the required confounding assumptions more plausible and help address reverse causation (Lash et al., 2021).

Before Calling a Variable a Mediator, Establish…

Use this as a practical causal mediation analysis checklist.

• A clearly defined causal question. Specify the exposure, mediator, outcome, population, exposure contrast, and whether the goal is a total effect, controlled direct effect, natural direct/indirect decomposition, or another mediation estimand.
• A defensible causal role for the proposed mediator. Explain why the exposure is believed to cause the mediator and why the mediator is believed to affect the outcome. Association alone does not establish either relationship.
• Temporal ordering. The exposure should precede the mediator and the mediator should precede the outcome in a way compatible with the proposed causal mechanism.
• Adequate control of exposure–outcome confounding. Identify and address common causes relevant to the overall exposure–outcome effect.
• Adequate control of exposure–mediator confounding. The apparent effect of exposure on mediator must not merely reflect uncontrolled common causes.
• Adequate control of mediator–outcome confounding. This remains necessary even when the exposure was randomized because the mediator itself usually was not.
• Assessment of exposure-induced mediator–outcome confounding. Determine whether any variable caused by exposure subsequently affects both mediator and outcome. If so, ordinary regression mediation methods may be inadequate.
• Consideration of exposure–mediator interaction. Do not assume that simple coefficient subtraction or product formulas remain appropriate when the effect of the mediator depends on exposure level.
• An estimand-specific identification argument. Natural direct and indirect effects require stronger assumptions than simply fitting mediator and outcome regressions.
• Appropriate statistical models. Correct specification of the relevant models remains necessary for regression-based estimators. A causal diagram or conceptual argument does not rescue badly specified statistical models.
• Consideration of unmeasured confounding. Because the required no-confounding assumptions are strong, sensitivity analysis can be important for evaluating how violations—particularly unmeasured mediator–outcome confounding—could alter conclusions (Lash et al., 2021).
• Interpretation matched to the evidence. If the design and assumptions do not support causal mediation, report an indirect association or statistical pattern rather than claiming that a biological or clinical mechanism has been established.

What Statistical Attenuation Can—and Cannot—Tell You

Suppose the exposure coefficient is 0.80 before adding a proposed mediator and 0.50 afterward.

What can you conclude?

You cannot conclude from that fact alone that:

• the exposure caused the mediator;
• the mediator caused the outcome;
• 0.30 represents a causal indirect effect;
• the remaining 0.50 is necessarily a causal direct effect;
• the mediator explains a biological mechanism;
• intervening on the mediator would remove the corresponding portion of the exposure effect.

Those conclusions require causal definitions and assumptions beyond coefficient attenuation.

When Natural Direct and Indirect Effects Become Difficult to Defend

Natural effects are attractive because they offer an apparently clean decomposition of a total effect into pathways through and outside the mediator. Their identification, however, can become difficult when:

• mediator–outcome confounding is unmeasured;
• a mediator–outcome confounder is itself affected by exposure;
• multiple mediators influence one another;
• exposure and mediator interact;
• temporal feedback exists;
• the exposure, mediator, or outcome varies over time;
• the required counterfactual “cross-world” assumptions are scientifically difficult to justify.

Hernán and Robins particularly emphasize the unverifiable nature of natural direct effects: because the estimand combines counterfactual conditions from different exposure worlds, even a randomized treatment experiment cannot directly identify it without additional assumptions (Hernán & Robins, 2020).

Lash et al. therefore distinguish natural effects from controlled direct effects and also discuss interventional direct and indirect effects, which modify the counterfactual intervention on the mediator and can sometimes be identified under weaker assumptions (Lash et al., 2021).

How to Report Mediation Claims More Carefully

When the causal requirements are defensible, report the estimand explicitly rather than simply saying that a variable “mediated” an effect.

Useful language distinguishes among:

• total causal effect;
• controlled direct effect;
• natural direct effect;
• natural indirect effect;
• assumptions used for identification;
• confounding control;
• exposure–mediator interaction;
• sensitivity to unmeasured mediator–outcome confounding.

When those requirements cannot be defended, narrower language is more appropriate:

“Adjustment for M attenuated the exposure coefficient.”

or:

“The observed associations were consistent with the proposed indirect pathway, but the analysis does not establish causal mediation.”

This preserves the statistical result without converting it into an unsupported mechanistic claim.

Bottom Line

Mediation analysis is not a procedure for proving mechanism by watching a regression coefficient shrink.

A causal mediation claim requires a causal structure in which the exposure affects the mediator and the mediator affects the outcome. Direct and indirect effects must be defined as causal estimands, typically using counterfactual interventions. Their identification requires assumptions concerning exposure–outcome, exposure–mediator, and mediator–outcome confounding, and natural effects impose additional difficulties when mediator–outcome confounders are themselves affected by exposure (Lash et al., 2021).

Temporal ordering is equally fundamental. Cross-sectional associations cannot generally distinguish a mediator from a confounder or establish the direction of causation.

The key question in mediation analysis is therefore not:

“Did the exposure coefficient get smaller after I added the mediator?”

It is:

Only after that question has been answered should regression coefficients be given causal labels.

FAQs

What is mediation analysis?

Mediation analysis examines whether and to what extent an exposure's effect on an outcome operates through a specified intermediate variable or mediator. For a causal interpretation, the mediator must occupy a causal pathway from exposure to outcome, and the relevant identification assumptions must hold (Lash et al., 2021).

Does a smaller exposure coefficient after adjustment prove mediation?

No. Coefficient attenuation is a statistical property of the fitted models. Interpreting the change as a causal indirect effect requires assumptions about temporal ordering, confounding, causal structure, interaction, and model specification. A smaller coefficient alone does not prove mediation or mechanism (Lash et al., 2021).

What is the difference between a total effect and a direct effect?

A total effect includes all causal pathways from exposure to outcome. A direct effect excludes pathways through a specified mediator according to the particular direct-effect definition being used. Conditioning on a mediator when the total effect is the target can therefore block part of the effect being estimated (Hernán & Robins, 2020).

What is the difference between a controlled and natural direct effect?

A controlled direct effect compares exposure conditions while fixing the mediator to a specified value. A natural direct effect compares exposure conditions while setting the mediator to the value it would naturally have taken under a reference exposure condition. Natural direct effects are useful for effect decomposition but involve stronger counterfactual assumptions (Lash et al., 2021; Hernán & Robins, 2020).

What confounding must be controlled in causal mediation analysis?

For conventional causal interpretations of natural direct and indirect effects, Lash et al. describe requirements to control exposure–outcome, mediator–outcome, and exposure–mediator confounding, together with an additional requirement that mediator–outcome confounders not themselves be affected by exposure (Lash et al., 2021).

Does randomizing the exposure solve mediation confounding?

Not completely. Randomization addresses exposure-related confounding, but the mediator itself is generally not randomized. Mediator–outcome confounding can therefore remain and bias causal direct and indirect effect estimates (Lash et al., 2021).

Can cross-sectional data establish causal mediation?

Generally, cross-sectional data provide weak support for causal mediation because temporal ordering, feedback, and reverse causation can be difficult or impossible to distinguish statistically. Substantive knowledge and study design are required to establish whether a variable is plausibly a mediator rather than a confounder or consequence (Lash et al., 2021).

Does finding an indirect effect prove a biological mechanism?

No. An estimated indirect effect does not by itself prove the underlying biological, behavioral, or clinical mechanism. Causal interpretation requires the relevant identification assumptions, and even a causal effect through a measured mediator should be described relative to that mediator rather than interpreted as proof of a complete mechanistic process.

References

Hernán, M. A., & Robins, J. M. (2020). Causal inference: What if. Chapman & Hall/CRC.

Lash, T. L., VanderWeele, T. J., Haneuse, S., & Rothman, K. J. (2021). Modern epidemiology (4th ed.). Wolters Kluwer.