What Is a Mediating Variable, and What Does It Mean to Say Something “Explains” a Relationship?

What Does It Really Mean When Researchers Say a Variable “Explains” a Relationship?

Research articles often contain statements such as “self-efficacy mediated the relationship between feedback and achievement” or “engagement explains why social support is associated with persistence.” These statements sound straightforward: X is related to Y, another variable M is introduced, and M apparently tells us why.

But the word explains carries considerable weight.

Does mediation mean that M is the reason X affects Y? Does the relationship between X and Y need to disappear once M is included? Is a statistically significant indirect effect enough to demonstrate a mechanism? Can researchers make the same claim when X, M, and Y were measured simultaneously?

A mediating variable is best understood as part of a proposed process linking an exposure or predictor to an outcome. Statistical mediation analysis can estimate quantities corresponding to that pathway. Whether those quantities establish a causal mechanism, however, depends on the design, temporal ordering, causal assumptions, measurement quality, and alternative explanations.

A Mediator Represents a Proposed Pathway Between X and Y

Mediation Is About a Pathway, Not Merely a Third Variable

The simplest mediation model is X → M → Y

In a simple mediation model, the researcher proposes a sequence:

X → M → Y

X is the exposure, intervention, predictor, or antecedent of interest. M is the proposed mediator. Y is the outcome.

Suppose a university introduces a formative-feedback intervention. Researchers theorize that the intervention increases students' academic self-efficacy and that increased self-efficacy subsequently improves persistence.

The proposed mechanism is:

Feedback intervention → Self-efficacy → Persistence

Self-efficacy is not merely another variable correlated with persistence. Its theoretical role is more specific: the intervention is proposed to change self-efficacy, which then contributes to the outcome.

What are the a, b, c, and c′ paths?

Traditional notation for a simple mediation model commonly uses several paths.

• a path: the relationship or effect of X on M;
• b path: the relationship or effect of M on Y conditional on X;
• c path: the total relationship or effect of X on Y;
• c′ path: the direct relationship or effect of X on Y not operating through the specified mediator, under the assumptions of the model.

The indirect effect through M is commonly represented in a simple linear model as the product of the a and b paths.

In simple linear settings without complications such as exposure–mediator interaction, the total effect can often be decomposed into direct and indirect components:

Total effect = Direct effect + Indirect effect

Modern causal mediation analysis allows more general definitions of direct and indirect effects and makes explicit that effect definition, identification, and statistical estimation are separate problems. The familiar path equations are therefore useful starting points, but they should not be mistaken for universal causal identities applicable without assumptions.

What does “indirect effect” actually mean?

The indirect effect represents the portion of the X–Y relationship or causal effect attributed, under the specified model, to the pathway through M.

Conceptually, the researcher is asking something like:

How much of the effect of X on Y operates through changes in M?

That question is inherently more demanding than asking whether X and M are correlated or whether M predicts Y.

To interpret the pathway causally, researchers need a defensible causal structure. They must consider whether important confounding exists for the exposure–outcome, exposure–mediator, or mediator–outcome relationships and whether the assumed temporal sequence is plausible.

This is why distinguishing a mediator from other third variables matters. A confounder, mediator, and moderator may all appear in the same dataset while requiring completely different treatment.

What does it mean to say the mediator “explains” the relationship?

In everyday research language, saying that M “explains” the relationship between X and Y usually means that the proposed pathway through M accounts for some portion of how X becomes related to Y.

That is useful shorthand, but it can easily become stronger than the evidence warrants.

There are at least two distinct meanings researchers may have in mind:

The second claim is much stronger.

A statistically estimated indirect effect does not automatically demonstrate that the proposed mediator is the real-world mechanism. Several causal structures can sometimes generate similar observed covariance patterns. Unmeasured common causes, reverse ordering, measurement problems, or alternative mediators may produce an apparently convincing statistical mediation pattern.

Thus, “M statistically mediated the association” and “M is the mechanism through which X causes Y” should not be treated as interchangeable conclusions.

Mediation is inherently directional

The arrows in X → M → Y are not decorative. They express a proposed sequence.

If self-efficacy mediates an intervention's effect on persistence, the theory implies that the intervention changes self-efficacy before the relevant change in persistence occurs.

This directionality is why mediation claims become difficult when all variables are measured at the same time. Cross-sectional covariance can show that X, M, and Y fit a particular statistical model, but it often cannot establish whether X preceded M, whether M preceded Y, or whether some reciprocal process generated the pattern.

If the theoretical direction itself is uncertain, researchers should first consider what to do when the direction of a relationship is unclear rather than forcing a mediation sequence because software can estimate it.

The mediator does not need to make the X–Y relationship disappear

An enduring misconception is that mediation exists only if the relationship between X and Y becomes nonsignificant after M is introduced.

That is not a general requirement.

Historically influential causal-steps procedures encouraged researchers to examine a sequence of significance tests involving X, M, and Y. Contemporary mediation analysis generally places greater emphasis on directly estimating the indirect effect and its uncertainty.

A meaningful indirect effect can exist even when the total X–Y effect is small or not statistically distinguishable from zero. Multiple pathways can operate in different directions, and direct and indirect components may offset one another.

“Full mediation” and “partial mediation” can be misleading shorthand

Researchers have often described mediation as full when the estimated direct X–Y relationship becomes nonsignificant after accounting for M and partial when a residual direct relationship remains.

These labels can encourage overinterpretation.

A nonsignificant direct effect does not demonstrate that the mediator captures the entire causal process. The estimate may be imprecise, other mediators may remain unmeasured, direct and indirect pathways may cancel one another, or the study may simply lack power to detect the remaining direct effect.

It is generally more informative to report and interpret the estimated indirect effect, direct effect, total effect where appropriate, their uncertainty, and the assumptions required for causal interpretation.

A mediator and a confounder differ by causal position

Suppose socioeconomic resources are associated with access to educational technology and with academic performance. If those resources precede both exposure and outcome and distort the exposure–outcome comparison, they may function as a confounder.

Now suppose technology access increases opportunities for deliberate practice, and practice subsequently improves academic performance. Practice may be a mediator.

Both variables may correlate with the exposure and outcome. Yet adjusting for them has different implications.

Appropriately addressing a confounder may be necessary to estimate a causal effect. Adjusting for a mediator while estimating a total effect can remove part of the very pathway through which the exposure operates.

This is why a variable should not be classified according to whether its inclusion makes a coefficient smaller. Researchers should instead determine whether the variable is a confounder or mediator in the specific causal question.

Mediation analysis requires assumptions about confounding too

Researchers sometimes imagine mediation analysis as a way to “go beyond” confounding once the exposure–outcome relationship has been adjusted. In reality, mediation introduces additional causal relationships that may themselves be confounded.

A causal mediation analysis may require assumptions concerning confounding of:

• the exposure–outcome relationship;
• the exposure–mediator relationship;
• the mediator–outcome relationship.

The mediator–outcome relationship is particularly challenging because the mediator is usually not randomized. Even if X is randomized, individuals are not generally randomized to their naturally occurring mediator values.

Randomization of X therefore strengthens causal inference about the exposure but does not automatically remove all assumptions required for a causal interpretation of mediation.

Randomized X does not automatically make the mediator causal

Suppose students are randomly assigned to an instructional intervention or control condition. The intervention later changes motivation, and motivation is associated with achievement.

Random assignment provides strong protection against baseline confounding of the intervention itself. But motivation was not necessarily randomized.

Students who develop greater motivation may also differ in unmeasured ways that affect achievement. If those differences confound the mediator–outcome relationship, the estimated indirect pathway may still be biased.

This does not make mediation analysis useless. It means that claims about mechanisms require assumptions beyond those needed merely to establish that the randomized intervention changed the outcome.

Cross-sectional mediation deserves especially cautious interpretation

Many mediation studies use a single survey in which X, M, and Y are measured at approximately the same time. Statistical software can estimate an indirect effect from such data, but temporal interpretation is difficult.

Suppose a survey finds:

Academic stress → Self-efficacy → Academic engagement

If all three constructs were measured simultaneously, several alternatives remain plausible. Stress may affect self-efficacy, but self-efficacy may also affect perceived stress. Engagement may influence self-efficacy. Earlier experiences may shape all three. Reciprocal effects may also occur.

A cross-sectional mediation model therefore provides limited evidence that the proposed sequence unfolded over time.

When theory suggests that variables may influence each other reciprocally, a unidirectional mediation model may be an oversimplification.

A longitudinal design can help, but timing must match the process

Measuring X at Time 1, M at Time 2, and Y at Time 3 can provide more credible temporal information than measuring all three simultaneously. Yet merely having three waves does not guarantee valid causal mediation.

The measurement intervals should correspond reasonably to the process being studied. Some mediators may change within minutes; others may evolve over semesters or years.

Researchers should also consider stability in the constructs, prior levels of the mediator and outcome, attrition, time-varying confounding, and whether meaningful changes could have occurred between measurement occasions.

The design should represent the hypothesized mechanism rather than merely satisfy a visual pattern of Time 1 → Time 2 → Time 3.

There can be more than one mediator

Real processes frequently operate through multiple pathways.

An educational intervention might improve achievement because it increases practice, strengthens self-efficacy, improves feedback quality, or changes several processes simultaneously.

Researchers may therefore consider multiple mediators, including parallel mediators or sequential mediators. Such models can be theoretically informative, but every added pathway introduces additional assumptions and interpretation challenges.

A complex model should not be treated as automatically superior. Sometimes adding more variables merely gives the diagram more arrows to admire, a temptation familiar to anyone who has survived enough structural equation modeling presentations.

The broader principle is that researchers should justify which variables actually belong in the study rather than adding every plausible mediator available in the questionnaire.

A moderator answers a different question

Mediation asks how an effect or relationship operates. Moderation asks whether that effect or relationship differs across conditions.

Suppose an intervention increases self-efficacy, which subsequently improves achievement. Self-efficacy is a mediator.

If the intervention is more effective among students with low prior knowledge than among those with high prior knowledge, prior knowledge is a moderator.

The distinction between a mediator and a moderator therefore rests on the theoretical question rather than on which variable appears third in the dataset.

Mediation can itself be conditional

The process represented by mediation does not always operate equally for everyone.

Imagine that an intervention improves self-efficacy, which improves persistence, but the self-efficacy pathway is much stronger for first-year students than for graduating students.

The indirect effect now depends on another variable. This combines mediation and moderation and is often described as moderated mediation or studied within conditional process analysis.

The existence of such models reinforces an important point: mediation and moderation are not competing labels for “third variables.” They represent different questions that can even coexist within a single theoretical model.

“Explained percentage” should be interpreted carefully

Researchers sometimes report the proportion or percentage of an effect that is mediated. Such summaries can appear intuitively attractive: perhaps 30% of an intervention effect is said to operate through motivation.

However, proportions mediated can be unstable or difficult to interpret in some situations, particularly when direct and indirect effects have different signs, the total effect is close to zero, nonlinear models are used, or exposure–mediator interactions are present.

A percentage can therefore create an impression of mechanistic precision that the study may not actually support.

When reporting mediation, researchers should generally prioritize clearly defined direct and indirect effect estimates, uncertainty intervals, the causal assumptions underlying those estimates, and substantive interpretation.

A statistically significant indirect effect is evidence, not a mechanistic verdict

Suppose an indirect-effect confidence interval excludes zero. That is evidence that the data are inconsistent with a zero indirect effect under the specified model and inferential procedure.

It does not demonstrate that:

• the proposed mediator is the only mechanism;
• the causal direction is correct;
• all relevant confounding has been eliminated;
• the mediator was measured without consequential error;
• changing the mediator experimentally would necessarily change the outcome by the estimated amount.

Statistical mediation is therefore one part of a mechanistic argument rather than a substitute for that argument.

From “It Works” to “How Might It Work?”

Suppose the estimated total effect of the intervention is 5 points on the writing assessment, and the estimated indirect effect through self-efficacy is 2 points.

A cautious interpretation might be that the results are consistent with part of the intervention's effect operating through self-efficacy, under the assumptions of the mediation model.

A much stronger statement such as “self-efficacy is the reason the intervention works” would claim more. Other mechanisms could also contribute to the effect, and the causal mediator–outcome relationship may require assumptions not guaranteed by randomization of the intervention.

The word “explains” should therefore be calibrated to the evidence.

Common Misinterpretations of Mediation and “Explanation”

Use Mediation When You Have a Process to Explain, Not Merely an Extra Variable

Before conducting mediation analysis, write the proposed process in ordinary language. If you cannot explain why X should change M and why changes in M should subsequently affect Y, the problem may not yet be a mediation question.

The language used in the manuscript should match this distinction. “The indirect effect through M was estimated as…” is narrower than “M explains why X causes Y.” The latter may be appropriate in a strong causal design with well-supported assumptions, but it should not be the automatic translation of a mediation output table.

Similarly, do not add candidate mediators simply because they have been used before. Whether a construct belongs in the model should follow from the research question and a defensible account of why that relationship belongs in the conceptual framework.

Before Claiming That a Variable Mediates or Explains a Relationship, Check This

Frequently Asked Questions About Mediating Variables

Mediation Can Clarify a Pathway, but “Explanation” Requires More Than a Significant Indirect Effect

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.