How does this differ from ordinary mediation?

Ordinary mediation need not allow the indirect effect to vary with another variable; moderated mediation explicitly studies that variation.

Is an interaction always evidence of moderated mediation?

No. Its location and relationship to the indirect pathway matter.

Can the moderator be categorical?

Yes. The conditional effects and their contrasts must be interpreted using the chosen coding and model specification.

AI Glossary / Statistics and probability / Advanced

Moderated mediation

Definition

Moderated mediation describes a model in which an indirect relationship between a predictor and an outcome through a mediator varies with the value of another variable, called a moderator. It combines the question of how an association operates with the question of when or for whom that indirect association differs. In a causal study, the corresponding interpretation requires assumptions that justify causal effects along the relevant paths. A fitted regression pattern alone does not establish a causal mechanism. The conditional indirect effect is the central quantity, rather than an interaction coefficient considered in isolation.

What it means

A mediator represents an intermediate variable in a proposed pathway from X to Y. A moderator W changes the strength or form of a relationship along that pathway. For example, the relationship between X and the mediator M may depend on W, while the relationship from M to Y remains constant in the specified model. The resulting indirect effect then depends on W. Other specifications allow moderation of the second path or several paths. These arrangements are not interchangeable, so the proposed mechanism and timing of variables should guide the equations.

How it works

Specify the mediator, outcome, predictor and moderator, including their measurement order and relevant confounders. Fit the equations appropriate to the hypothesized paths and interactions. Calculate the conditional indirect effect at meaningful, supported values of the moderator. In a simple linear first-stage model, M contains terms a1X and a3XW, while the outcome equation has mediator coefficient b. The conditional indirect effect is (a1 + a3W)b. Its rate of change with W is a3b, a linear index of moderated mediation. Estimate uncertainty using a method appropriate to the design, such as suitable bootstrap intervals, and assess the assumptions required for causal interpretation.

Three practical examples

Learning research: examine whether a training program's indirect relationship with performance through confidence differs by prior experience.
Service evaluation: study whether a proposed pathway from response speed through satisfaction to retention varies with request complexity.
Organizational analysis: assess whether an association between resources and output through coordination differs across working conditions. These are hypotheses to investigate, not causal conclusions supplied by the model label.

Think of it this way

Think of a route with two connected segments. A condition affecting the first segment changes the overall route even if the second segment stays the same. Moderated mediation asks whether the strength of an indirect pathway changes with that condition. The analogy does not prove that the route causes the observed destination; that requires a suitable study and defensible assumptions.

Worked example

Suppose a purely illustrative linear model has first-stage coefficients a1 = 0.5 and a3 = 0.25, and second-stage coefficient b = 0.8. The conditional indirect effect is (0.5 + 0.25W) times 0.8 = 0.4 + 0.2W. At W = 0 it is 0.4 outcome units per unit of X; at W = 2 it is 0.8. Their difference is 0.4, and the index of moderated mediation is 0.25 times 0.8 = 0.2 per unit of W. These original calculations describe the specified linear model. They provide no standard error, confidence interval or evidence of statistical significance. A nonzero estimated index is not automatically convincing evidence without uncertainty assessment. If both component paths depend on W, their product can be nonlinear in W and a single constant linear index may no longer summarize the relationship.

Illustrative case study

In a fictional case, Pine Learning evaluates a workplace course. Researchers hypothesize that the course relates to later task performance through increased confidence, with a different first-stage relationship for people with different prior experience. They measure variables at planned times, document confounding concerns and report conditional indirect estimates with uncertainty. The report avoids claiming that confidence has been proven to cause the performance difference. It also explains where the data provide little support for particular moderator values, limiting interpretation of those comparisons.

Common misconceptions and mistakes

A significant indirect effect at one moderator value and a nonsignificant effect at another do not by themselves demonstrate that the effects differ. The difference or relevant index requires direct assessment. Moderating the direct X-to-Y path alone does not necessarily imply moderated mediation. Centering variables can aid interpretation but does not establish a mechanism. Bootstrap intervals address sampling uncertainty under the analysis procedure; they do not remove confounding, measurement error or incorrect temporal ordering. A total association need not be statistically significant before an indirect pathway can be meaningfully investigated.

Business use

Businesses should use moderated mediation when a substantive pathway hypothesis matters to a decision and the study can support the needed interpretation. Define the analysis before exploring many mediator-moderator combinations, and avoid turning subgroup searches into confident stories. Report the equations, scales, moderator range, conditional estimates and uncertainty. Separate descriptive pathway associations from causal claims and consider sensitivity to unmeasured confounding. The analysis can sharpen questions about process differences, but operational changes should be based on the strength of the overall evidence.

Check your understanding

An indirect effect is significant in one group but not another. Has moderated mediation been established?

Answer: No. Directly assess the difference in indirect effects, or the relevant moderated-mediation parameter, with its uncertainty.

Related terms

Learn next

Review regression interactions and simple mediation before combining them. Practice deriving the conditional indirect effect from explicit equations, then examine the assumptions needed to interpret it causally.

Sources

Andrew F. Hayes, An Index and Test of Linear Moderated Mediation: https://doi.org/10.1080/00273171.2014.962683 Imai, Keele and Yamamoto, Identification, Inference, and Sensitivity Analysis for Causal Mediation Effects: https://imai.fas.harvard.edu/research/mediation.html

Previous: Model validation Next: Moderation (statistics)