Intervening Variable – Definition, Examples, and How to Test It

An intervening variable is a variable that explains how or why an independent variable influences a dependent variable. It sits in the proposed causal pathway: the independent variable affects the intervening variable, which then affects the outcome. In contemporary research, the term is usually used interchangeably with mediator or mediating variable.

Intervening variables help researchers move beyond identifying whether two variables are related. They allow researchers to investigate the process or mechanism through which that relationship may occur.

This article explains what an intervening variable is, how it differs from moderators and confounders, how direct and indirect effects are calculated, and how mediation hypotheses can be evaluated and reported.

Key Takeaways

• An intervening variable explains the mechanism connecting an independent variable to an outcome.
• The basic model is X → M → Y, where M is the intervening or mediating variable.
• The indirect effect is commonly represented by the product ab.
• A mediator is different from a moderator, confounder, and ordinary control variable.
• A statistical indirect effect does not, by itself, prove a causal mechanism.
• Modern analysis focuses on the size and uncertainty of the indirect effect rather than relying only on traditional causal steps.

What Is an Intervening Variable?

An intervening variable is a variable positioned between a proposed cause and an outcome. It explains the process through which the independent variable may influence the dependent variable.

The basic relationship can be represented as:

Independent variable → Intervening variable → Dependent variable

Or, using common statistical notation:

X → M → Y

Where:

• X is the independent, predictor, exposure, or treatment variable.
• M is the intervening or mediating variable.
• Y is the dependent or outcome variable.

For example, suppose a researcher finds that employee training is associated with better job performance. Training may not improve performance solely through a direct effect. It may first improve employees’ skills, which subsequently improve their performance.

The proposed model would be:

Employee training → Skill acquisition → Job performance

In this example, skill acquisition is the intervening variable.

Simple definition

An intervening variable is the variable that carries or helps explain the effect of one variable on another.

Why is it called “intervening”?

The word intervening means occurring between two things. The variable is described as intervening because it appears conceptually and temporally between the proposed cause and the outcome.

However, its position should be justified by theory, previous evidence, and study design. A variable does not become an intervening variable merely because it is entered between two other variables in a statistical model.

How Does an Intervening Variable Work?

An intervening variable divides the relationship between X and Y into separate pathways.

In a simple linear mediation model:

1. X predicts or influences M.
2. M predicts or influences Y, while X is included in the outcome model.
3. Part of the relationship between X and Y operates through M.

Path a

Path a represents the effect of the independent variable on the intervening variable:

X → M

For example:

Training → Skill acquisition

Path b

Path b represents the relationship between the intervening variable and the outcome while controlling for the independent variable:

M → Y

For example:

Skill acquisition → Job performance

Direct effect: c′

The direct effect, written as c′, is the effect of X on Y after the intervening variable has been included in the model.

X → Y, controlling for M

Indirect effect: ab

The indirect effect represents the portion of the effect that operates through the intervening variable:

Indirect effect = a × b

It is commonly written as:

ab

Total effect: c

The total effect represents the overall relationship between X and Y before decomposing it into direct and indirect pathways.

In a simple linear model without an interaction between X and M:

c = c′ + ab

Where:

• c = total effect
• c′ = direct effect
• ab = indirect effect

This simple decomposition should not be applied automatically to nonlinear models, models with treatment–mediator interactions, or other complex causal structures. Such models may require different definitions and estimation procedures.

Regression equations

A simple linear mediation model may be written as:

M = i₁ + aX + e₁

Y = i₂ + c′X + bM + e₂

Where:

• i₁ and i₂ are intercepts.
• a, b, and c′ are path coefficients.
• e₁ and e₂ are residual errors.

Worked numerical example

Consider a hypothetical study examining whether subject understanding mediates the relationship between tutoring time and examination scores.

Suppose the estimated unstandardized coefficients are:

• Path a = 0.80
• Path b = 2.00
• Direct effect c′ = 1.10

The indirect effect is:

ab = 0.80 × 2.00 = 1.60

The estimated total effect is:

c = c′ + ab

c = 1.10 + 1.60 = 2.70

Under this hypothetical model, each additional unit of tutoring time is associated with an estimated 2.70-unit increase in examination score. Approximately 1.60 units operate through improved subject understanding, while 1.10 units represent the estimated direct pathway and any other pathways not included in the model.

These estimates would still require confidence intervals and an evaluation of model assumptions before conclusions were drawn.

Is an Intervening Variable the Same as a Mediating Variable?

In most contemporary applied research, intervening variable, mediator, and mediating variable are used to describe the same general role: a variable that transmits or explains part of the relationship between an independent variable and an outcome.

However, the terms have not always been completely interchangeable.

Historical meaning

In earlier behavioural science and philosophy-of-science discussions, an intervening variable sometimes referred to an abstract summary of observable relationships. MacCorquodale and Meehl (1948), for example, distinguished intervening variables from hypothetical constructs.

This historical distinction explains why some older sources describe intervening variables as theoretical or unobservable.

Contemporary meaning

In modern mediation research, a mediator may be:

• Directly observed.
• Measured using a questionnaire or test.
• Represented as a latent variable.
• Continuous, binary, ordinal, or categorical.
• Measured at one or several time points.
• Experimentally manipulated in some research designs.

Therefore, it is inaccurate to state that an intervening variable must always be unobservable or impossible to measure.

Practical terminology recommendation

For student research, the following wording is usually clear:

“The study examines whether M mediates the relationship between X and Y. M is therefore the proposed mediating or intervening variable.”

Use the terminology preferred by the discipline, supervisor, target journal, or methodological framework, but define it clearly when it first appears.

Intervening Variable vs. Moderator, Confounder, and Control Variable

These variables play different conceptual and statistical roles.

Variable	Main question answered	Typical causal position	Example
Intervening or mediating variable	How or why does X affect Y?	X → M → Y	Training improves skills, which improve performance
Moderator	When, for whom, or under what conditions does X affect Y?	Changes the size or direction of an effect	Training is more effective for inexperienced employees
Confounder	Is another variable producing or distorting the observed relationship?	Common cause of two or more variables	Previous experience affects both training participation and performance
Control variable or covariate	What variable is being statistically adjusted for?	Depends on its actual causal role	Age is entered as an adjustment variable
Independent variable	What is the proposed predictor, exposure, or cause?	Beginning of the model	Training programme
Dependent variable	What outcome is being explained or predicted?	End of the model	Job performance

Intervening variable vs. moderator

A mediator explains how or why a relationship occurs.

A moderator explains when, for whom, or under what conditions the relationship becomes stronger, weaker, positive, negative, or absent.

Compare these models:

Mediation

Training → Skill acquisition → Performance

Moderation

Training × Previous experience → Performance

In the second model, previous experience changes the effect of training on performance.

A variable can act as a mediator in one theoretical model and a moderator in another. Its role depends on the research question, causal assumptions, timing, and statistical specification.

Intervening variable vs. confounding variable

A mediator is part of the proposed effect pathway. A confounder is a cause of variables whose relationship the researcher is trying to estimate.

For example:

Socioeconomic background → Access to tutoring

Socioeconomic background → Examination performance

Socioeconomic background may confound the association between tutoring and examination performance because it influences both.

By contrast:

Tutoring → Subject understanding → Examination performance

Subject understanding is a proposed mediator because it occurs after tutoring and transmits part of its effect.

Confusing mediators with confounders can produce serious analytical errors. Adjusting for a mediator while estimating the total effect can block part of the effect the researcher intends to measure. Failing to adjust appropriately for mediator–outcome confounders can also bias the indirect-effect estimate.

Intervening variable vs. control variable

“Control variable” describes how a variable is used in a model, not what causal role it plays.

A variable entered as a control could actually be:

• A confounder.
• A precision-improving covariate.
• A mediator.
• A collider.
• A competing predictor.
• An inappropriate post-treatment variable.

Researchers should therefore select adjustment variables using theory and a causal diagram rather than automatically controlling for every available characteristic.

Examples of Intervening Variables

The following are hypothetical models intended to demonstrate variable roles. They do not establish that the stated causal relationships are true in every context.

Field	Independent variable	Intervening variable	Dependent variable	Proposed explanation
Education	Tutoring time	Subject understanding	Examination score	Tutoring improves understanding, which improves performance
Psychology	Stress-management programme	Perceived coping ability	Psychological distress	The programme strengthens coping, reducing distress
Workplace research	Employee training	Skill acquisition	Job performance	Training develops skills that improve performance
Public health	Health campaign exposure	Knowledge of preventive behaviour	Preventive action	Campaigns may change behaviour by increasing knowledge
Technology adoption	Interface redesign	Perceived ease of use	Continued system use	A simpler interface improves ease of use, encouraging continued use
Marketing	Advertising exposure	Brand trust	Purchase intention	Advertising may influence intention by building trust
Sociology	Socioeconomic status	Access to healthcare	Health outcome	Resources may affect health partly through access to care
Environmental psychology	Access to green space	Perceived restoration	Well-being	Green space may improve well-being through restorative experiences

Education example

A researcher finds that classroom participation is positively associated with academic achievement.

A possible intervening model is:

Classroom participation → Understanding of course material → Academic achievement

Understanding of the material is the mediator. The hypothesis is that participation improves understanding, which then contributes to achievement.

Psychology example

Suppose social support is related to lower stress.

A proposed model might be:

Social support → Perceived coping resources → Stress

The model suggests that support reduces stress partly because it strengthens a person’s perceived ability to cope.

Business example

A company introduces a professional-development programme and later observes greater employee productivity.

A possible mechanism is:

Professional development → Job-related competence → Productivity

The competence measure is an intervening variable only if the theory, measurement timing, and design support that ordering.

Technology example

Researchers evaluate whether a redesigned mobile application increases continued use.

A potential model is:

Interface redesign → Perceived ease of use → Continued use

The indirect pathway represents the change in continued use that operates through perceived ease of use.

Main Types of Mediation Models

Intervening-variable models can take several forms.

1. Simple mediation

A simple mediation model contains one independent variable, one mediator, and one outcome:

X → M → Y

This is the most suitable model for introducing mediation concepts.

2. Parallel multiple mediation

A parallel mediation model contains two or more mediators operating alongside one another.

For example:

Training → Skills → Performance

Training → Motivation → Performance

Skills and motivation are estimated as separate pathways, without assuming that one causes the other.

3. Serial or sequential mediation

A serial model proposes a sequence of intervening variables:

X → M₁ → M₂ → Y

For example:

Training → Skills → Self-efficacy → Performance

This model requires especially strong theoretical and temporal justification because the order of the mediators matters.

4. Moderated mediation

Moderated mediation occurs when the size of an indirect effect depends on another variable.

For example, the indirect effect of training through skills may be stronger for employees who receive supervisory support.

This is also called a conditional indirect effect.

5. Latent-variable mediation

A mediator such as motivation, trust, anxiety, or satisfaction may be represented as a latent variable measured through several indicators.

Structural equation modelling can estimate the measurement model and mediation pathways together, although valid interpretation still depends on study design and causal assumptions.

6. Longitudinal mediation

Longitudinal mediation measures variables at multiple time points, such as:

• Training at Time 1.
• Skills at Time 2.
• Performance at Time 3.

This provides stronger evidence about temporal ordering than measuring all three variables simultaneously, although longitudinal data do not automatically eliminate confounding or reverse causation.

7. Multilevel mediation

Multilevel mediation is used when observations are nested, such as students within classrooms, employees within organisations, or repeated observations within individuals.

The mediation pathway may operate:

• Within groups.
• Between groups.
• Across different levels.

Ordinary single-level regression may produce misleading standard errors or effects when the nested structure is ignored.

How to Identify an Intervening Variable

An intervening variable should be identified through theory and research design before statistical testing.

Step 1: Define the proposed cause and outcome

Clearly state:

• What is expected to produce change?
• What outcome is expected to change?

For example:

• X: feedback quality
• Y: student performance

Step 2: Ask how or why the effect might occur

Identify the process that could connect the two variables.

For example:

• Better feedback may improve students’ understanding of their errors.
• Improved understanding may lead to better performance.

The proposed intervening variable is therefore understanding of errors.

Step 3: Establish temporal order

The proposed sequence should make theoretical and temporal sense:

1. X occurs first.
2. M changes afterward.
3. Y is observed after the mediator.

When all variables are measured simultaneously, this order is assumed rather than demonstrated.

Step 4: Draw a conceptual or causal diagram

A basic diagram is:

X → M → Y

Researchers should add plausible confounders and alternative paths rather than drawing only the preferred relationship.

A directed acyclic graph can help distinguish mediators, confounders, colliders, and variables that should or should not be adjusted for.

Step 5: Define and measure the mediator carefully

State:

• The conceptual definition.
• The operational definition.
• The instrument or data source.
• The measurement time.
• Reliability and validity evidence.
• Whether the mediator is observed or latent.

Poor measurement can weaken path estimates and distort the estimated indirect effect.

Step 6: Identify competing explanations

Consider whether:

• Y might influence M.
• A third variable affects both M and Y.
• X changes another variable that confounds the M–Y relationship.
• The proposed mediator is merely correlated with the true mechanism.

Step 7: Prespecify the mediation hypothesis

Where possible, define the model, variables, covariates, effect measure, analysis method, and primary indirect pathway before inspecting the final results.

Prespecification reduces the risk of choosing a mediator only because it produced a desirable result.

How to Formulate Mediation Hypotheses

A mediation study can include separate path hypotheses, but the central hypothesis should concern the indirect effect.

Example research question

Does academic self-efficacy mediate the relationship between formative feedback and academic performance?

Example conceptual model

Formative feedback → Academic self-efficacy → Academic performance

Example hypotheses

H1: Formative feedback is positively associated with academic self-efficacy.

H2: Academic self-efficacy is positively associated with academic performance after accounting for formative feedback.

H3: Formative feedback has a positive indirect effect on academic performance through academic self-efficacy.

H3 is the principal mediation hypothesis because it concerns the product of the two component paths.

Researchers should not treat significance of every individual relationship as a universal prerequisite for evaluating the indirect effect. Nor is a statistically significant total effect always required before an indirect effect is examined (Hayes, 2009; MacKinnon et al., 2002).

How to Test an Intervening Variable

Mediation analysis evaluates whether the data are consistent with a proposed indirect pathway. The following procedure applies most directly to a simple observed-variable model.

Step 1: Specify the theoretical model

Before running software, define:

• Independent variable X.
• Mediator M.
• Outcome Y.
• Relevant confounders.
• Expected direction of each path.
• Timing of measurements.
• Whether interactions are expected.
• Whether the analysis is exploratory or confirmatory.

Step 2: Evaluate the research design

Determine whether the design supports the proposed order.

Consider:

• Was X randomized?
• Was the mediator measured after X?
• Was the outcome measured after the mediator?
• Could the outcome influence the mediator?
• Were important confounders measured?
• Are participants nested or repeatedly measured?

Step 3: Inspect measurement quality and data

Before estimating mediation, assess:

• Missing data.
• Outliers and influential observations.
• Reliability of multi-item measures.
• Scale coding and direction.
• Distribution of variables.
• Clustering or repeated observations.
• Whether the sample provides adequate precision.

Power for indirect effects depends on the sizes of paths a and b, their uncertainty, model complexity, measurement reliability, and the chosen estimator. A single universal sample-size rule is therefore inappropriate. Simulation or mediation-specific power analysis is preferable for planned studies.

Step 4: Estimate the mediator model

Estimate path a:

M = i₁ + aX + covariates + e₁

This model evaluates the association or effect of X on M.

Step 5: Estimate the outcome model

Estimate paths b and c′:

Y = i₂ + c′X + bM + covariates + e₂

This model estimates:

• Path b: mediator–outcome relationship conditional on X and included covariates.
• Path c′: direct effect of X on Y after including M.

Step 6: Calculate the indirect effect

For the basic linear model:

Indirect effect = ab

The sign of the effect depends on the signs of a and b.

• Positive a and positive b produce a positive indirect effect.
• Negative a and negative b also produce a positive product.
• Paths with opposite signs produce a negative indirect effect.

When direct and indirect effects have opposing signs, the model may be described as inconsistent mediation or competitive mediation. In this situation, the total effect can be small even when an indirect pathway is meaningful.

Step 7: Calculate an interval estimate

The sampling distribution of ab is often asymmetric, particularly in smaller samples. For that reason, researchers commonly use:

• Bootstrap confidence intervals.
• Monte Carlo confidence intervals.
• Distribution-of-the-product methods.
• Model-based intervals appropriate to the selected causal framework.

The traditional Sobel test assumes an approximately normal sampling distribution and should not normally be the sole basis for inference.

An indirect effect is commonly considered statistically distinguishable from zero when its chosen confidence interval does not contain zero. However, the effect size, precision, design quality, and practical importance should also be interpreted.

Step 8: Examine assumptions and sensitivity

Assess whether conclusions change under:

• Alternative covariate sets.
• Alternative variable codings.
• Plausible reverse models.
• Robust or cluster-adjusted standard errors.
• Different missing-data treatments.
• Sensitivity analysis for unmeasured mediator–outcome confounding.

A statistically significant model with implausible assumptions is not strong evidence of a causal mechanism.

Step 9: Report estimates rather than only labels

Report:

• Path a.
• Path b.
• Direct effect.
• Indirect effect.
• Total effect where relevant.
• Standard errors or confidence intervals.
• Estimator and software.
• Number of bootstrap samples, when used.
• Covariates and their rationale.
• Measurement timing.
• Missing-data treatment.
• Assumptions and limitations.

Avoid reporting only that “full” or “partial” mediation occurred.

Full and Partial Mediation

Traditional explanations commonly divide findings into full and partial mediation.

Full mediation

Full mediation is usually said to occur when:

• The indirect effect is present.
• The direct effect is no longer statistically significant after including the mediator.

However, a nonsignificant direct effect does not prove that the true direct effect is exactly zero. It may reflect low precision, limited statistical power, or measurement error.

Partial mediation

Partial mediation is usually said to occur when:

• An indirect effect is present.
• A direct effect remains after the mediator is included.

The term indicates that the proposed mediator explains only part of the relationship.

Better reporting practice

Rather than treating full and partial mediation as rigid categories, report:

• The estimated indirect effect.
• The estimated direct effect.
• Their confidence intervals.
• The direction and magnitude of each pathway.
• The design limitations affecting causal interpretation.

This provides more information than a binary label.

Assumptions Required for Causal Interpretation

An estimated indirect effect is not automatically a causal indirect effect.

Causal interpretation generally requires several assumptions.

1. Correct temporal ordering

The cause should precede the mediator, and the mediator should precede the outcome.

2. No important unmeasured exposure–outcome confounding

There should not be an omitted variable that causes both X and Y, unless the design or analysis accounts for it.

3. No important unmeasured exposure–mediator confounding

There should not be an omitted common cause of X and M.

4. No important unmeasured mediator–outcome confounding

There should not be an omitted common cause of M and Y after appropriate conditioning.

This assumption remains important even when X is randomized, because the mediator is usually not randomized.

5. Appropriate handling of exposure-induced confounding

A variable caused by X may subsequently affect both M and Y. Standard mediation procedures may not correctly handle this exposure-induced mediator–outcome confounder.

6. Correct model specification

The functional form, interactions, links, distributions, and clustering structure should be appropriate.

7. Adequate measurement

Measurement error in the mediator can bias path estimates. This is especially important when complex constructs are measured with short or unreliable instruments.

8. Positivity and meaningful variation

Relevant combinations of exposure, mediator, and confounder values must be represented sufficiently in the data for the intended effects to be estimated.

9. Consistency and interference considerations

In causal frameworks, the treatment and mediator levels should be meaningfully defined. One participant’s exposure should not improperly alter another participant’s outcome unless interference is explicitly modelled.

Because some of these assumptions cannot be fully tested from the observed data, researchers should state them transparently and use sensitivity analyses where appropriate (Imai et al., 2010).

Can Cross-Sectional Data Establish Mediation?

Cross-sectional data can be used to estimate a statistical indirect-effect model, but they usually cannot establish the temporal sequence required for a strong causal mediation claim.

If X, M, and Y are measured at the same time, several models may fit the observed associations:

• X → M → Y
• X → Y → M
• M → X → Y
• A common cause influencing all three variables

Statistical model fit alone cannot determine which sequence occurred in time.

Researchers using cross-sectional data should therefore write:

“The findings were consistent with the proposed indirect-effect model.”

They should generally avoid writing:

“The study proved that M caused the effect of X on Y.”

Stronger research designs

Evidence is strengthened when:

• X is experimentally manipulated.
• Baseline values are measured.
• The mediator is assessed after X.
• The outcome is assessed after the mediator.
• Relevant confounders are measured.
• The mediator is itself manipulated where ethically and practically possible.
• Repeated measurements are analysed with an appropriate longitudinal model.

Even with longitudinal data, correct timing and confounding control remain necessary.

How Intervening Variables Are Used in Modern Research

Intervening-variable analysis is used to investigate mechanisms in many fields.

Programme and intervention evaluation

Researchers may ask not only whether a programme works, but how it produces an outcome.

For example:

Training intervention → Behavioural skill → Target behaviour

Identifying a credible mechanism can help improve future programmes.

Psychology and behavioural science

Mediators are frequently proposed to explain changes in cognition, emotion, motivation, attitudes, and behaviour.

Examples include self-efficacy, perceived control, coping, trust, knowledge, and motivation.

Education research

Researchers may study whether feedback, teaching methods, technology, or classroom environments influence achievement through engagement, understanding, motivation, or self-regulated learning.

Public health and clinical research

Mediation analysis can examine behavioural, biological, social, and treatment-related pathways. Because clinical interpretations may affect decisions, causal assumptions and transparent reporting are especially important.

Organisational and management research

Potential mediators include employee engagement, job satisfaction, organisational commitment, competence, trust, and role clarity.

Digital and information-systems research

Researchers may examine whether system quality affects adoption through usefulness, trust, satisfaction, perceived risk, or ease of use.

Current methodological developments

Modern research extends beyond a single linear mediator. Current applications include:

• Causal mediation using potential-outcomes frameworks.
• Sensitivity analysis.
• Multiple and sequential mediators.
• Moderated mediation.
• Longitudinal and multilevel models.
• Latent-variable mediation.
• Generalized models for binary or categorical variables.
• Bayesian mediation models.
• High-dimensional mediation involving many potential pathways.

These developments expand what can be investigated, but they also require more explicit assumptions and careful reporting.

Digital Research Tools for Mediation Analysis

Several tools can estimate intervening-variable models.

PROCESS for SPSS, SAS, and R

PROCESS is widely used for observed-variable mediation, moderation, and conditional-process analysis.

It is useful for:

• Simple mediation.
• Parallel multiple mediation.
• Moderation.
• Moderated mediation.
• Bootstrap confidence intervals.

Researchers should select a model based on theory rather than choosing a PROCESS template only because it is available.

R and lavaan

The lavaan package supports:

• Path analysis.
• Structural equation modelling.
• Latent mediators.
• Multiple mediators.
• Indirect, direct, and total effects.
• Bootstrap standard errors and confidence intervals.

It is particularly useful when measurement models and structural paths must be estimated together.

JASP and jamovi

JASP and jamovi provide graphical interfaces that can make introductory mediation analysis more accessible. Users must still understand variable roles, assumptions, estimators, and output interpretation.

Other statistical environments

Mediation can also be implemented in:

• Stata.
• SAS.
• Mplus.
• Python.
• Dedicated R packages for causal, multilevel, longitudinal, or Bayesian mediation.

Software output does not determine whether the causal model is credible.

How Artificial Intelligence Can Assist

Artificial intelligence can support parts of the research workflow, but it should not replace methodological judgment.

AI may help researchers:

• Generate alternative literature-search terms.
• Explain unfamiliar software output in plain language.
• Draft or debug analysis code.
• Create simulated examples for learning.
• Check whether a report covers major reporting items.
• Improve the readability of a methods section.
• Compare terminology used across disciplines.

AI cannot reliably:

• Decide the true causal order from a dataset alone.
• Identify every unmeasured confounder.
• Make cross-sectional data longitudinal.
• Repair poor measurement.
• Determine whether a mediator has been validly operationalised.
• Guarantee that generated code uses the correct model.
• Replace substantive expertise or statistical review.

Researchers should verify AI-generated code against official documentation, test it on simulated or known data, examine all output, and avoid uploading confidential participant information to unapproved systems.

Advantages of Studying Intervening Variables

Explains mechanisms

Mediation analysis can move a study from “X is related to Y” toward a more informative explanation of how that relationship may operate.

Improves theory

A theory is more useful when it identifies processes rather than merely listing correlated variables.

Supports intervention development

If an intervention works through a particular mechanism, future programmes may be designed to strengthen that mechanism.

Distinguishes multiple pathways

A total effect can be separated into direct and indirect components, subject to the assumptions of the model.

Generates new research questions

An intervening model may reveal where additional experiments, measurements, qualitative research, or longitudinal studies are needed.

Limitations of Intervening-Variable Analysis

Causal conclusions require strong assumptions

A significant indirect effect does not prove that the mediator caused the outcome.

Temporal order may be unclear

Simultaneous measurement makes reverse causation difficult to exclude.

Unmeasured confounding can bias results

The mediator–outcome pathway is particularly vulnerable because mediators are seldom randomized.

Measurement error matters

An unreliable mediator measure can weaken or distort estimated relationships.

Results are model-dependent

Different covariates, interactions, link functions, time points, or causal definitions can produce different estimates.

Indirect effects may have low precision

The product of two uncertain paths may require substantial data for precise estimation.

Multiple mediators may compete or overlap

Correlated mediators can make individual pathways difficult to interpret.

The mediator may be a proxy

A measured variable may represent another unmeasured process rather than being the mechanism itself.

Common Mistakes

1. Treating correlation as proof of mediation

Three correlated variables do not establish the sequence X → M → Y.

2. Selecting the mediator after seeing the results

A post hoc mediator may describe the current dataset without providing strong confirmatory evidence.

3. Requiring a significant total effect

An indirect effect can exist even when positive and negative pathways cancel one another or the total-effect estimate is imprecise.

4. Testing only whether the X coefficient decreases

A smaller direct-effect coefficient after adding M is not a sufficient test of the indirect effect.

5. Using the Sobel test as the only analysis

The product ab often has an asymmetric sampling distribution. More appropriate interval methods are generally available.

6. Confusing a mediator with a confounder

A mediator lies on the pathway; a confounder helps cause variables whose relationship is being estimated.

7. Controlling for every available variable

Automatic adjustment can introduce bias, block part of the target effect, or condition on a collider.

8. Ignoring measurement timing

A mediator measured after the outcome cannot plausibly explain an earlier outcome without a more complex longitudinal theory.

9. Declaring “full mediation” from a nonsignificant direct effect

Nonsignificance is not proof that the direct effect is zero.

10. Reporting only p-values

Effect estimates, confidence intervals, measurement information, design details, and assumptions are also required.

11. Ignoring interactions

If the effect of the mediator depends on X or another variable, the indirect effect may vary across conditions.

12. Assuming software validates the conceptual model

Statistical software estimates the specified model; it does not establish that the model represents the true mechanism.

How to Report an Intervening-Variable Analysis

A transparent report should cover the research rationale, study design, variable timing, measurement, model, assumptions, estimates, uncertainty, and limitations.

Recommended reporting checklist

Report:

1. The theoretical reason for proposing the mediator.
2. A diagram of the hypothesised causal structure.
3. Definitions and measurement procedures for X, M, and Y.
4. Measurement timing.
5. Study design and sampling procedure.
6. Covariates and the reason each was included.
7. Statistical models and estimators.
8. Direct, indirect, and total effects where relevant.
9. Confidence intervals and the method used to calculate them.
10. Missing-data treatment.
11. Clustering, weighting, or repeated-measure procedures.
12. Model diagnostics and sensitivity analyses.
13. Limitations affecting causal interpretation.

The AGReMA statement provides more detailed guidance for reporting mediation analyses in randomized and observational research (Lee et al., 2021).

Example reporting template

A mediation analysis examined whether [mediator] explained part of the relationship between [independent variable] and [outcome]. The model was specified before/after data analysis on the basis of [theory or evidence]. Path a, from [X] to [M], was [estimate, confidence interval]. Path b, from [M] to [Y] conditional on [X and covariates], was [estimate, confidence interval]. The estimated indirect effect was [ab estimate], with a [percentage]% [bootstrap/other] confidence interval of [lower, upper]. The direct effect was [estimate and interval]. These findings were [consistent/not consistent] with the hypothesised indirect pathway. Causal interpretation is limited by [design, timing, measurement, or confounding limitations].

Example cautious interpretation

The estimated indirect effect was positive, and its bootstrap confidence interval excluded zero. The findings were therefore consistent with the hypothesis that academic self-efficacy partly explained the association between formative feedback and academic performance. Because all variables were observationally measured, unmeasured confounding and reverse causation cannot be excluded.

This wording distinguishes statistical evidence from a definitive causal claim.

Conclusion

An intervening variable explains the process through which an independent variable may influence an outcome. In contemporary research, it is usually called a mediator and is represented by the pathway X → M → Y.

A credible mediation study requires more than inserting a third variable into a regression. Researchers must justify the causal order, measure the mediator appropriately, estimate the indirect effect and its uncertainty, address plausible confounding, and report limitations transparently. When these requirements are met, intervening-variable analysis can provide valuable insight into the mechanisms underlying observed relationships.

References

• Baron, R. M., & Kenny, D. A. (1986). The moderator–mediator variable distinction in social psychological research: Conceptual, strategic, and statistical considerations. Journal of Personality and Social Psychology, 51(6), 1173–1182. https://doi.org/10.1037/0022-3514.51.6.1173
• Hayes, A. F. (2009). Beyond Baron and Kenny: Statistical mediation analysis in the new millennium. Communication Monographs, 76(4), 408–420. https://doi.org/10.1080/03637750903310360
• Hayes, A. F., & Rockwood, N. J. (2017). Regression-based statistical mediation and moderation analysis in clinical research: Observations, recommendations, and implementation. Behaviour Research and Therapy, 98, 39–57. https://doi.org/10.1016/j.brat.2016.11.001
• Imai, K., Keele, L., & Tingley, D. (2010). A general approach to causal mediation analysis. Psychological Methods, 15(4), 309–334. https://doi.org/10.1037/a0020761
• Lee, H., Cashin, A. G., Lamb, S. E., Hopewell, S., Vansteelandt, S., VanderWeele, T. J., MacKinnon, D. P., Mansell, G., Collins, G. S., Golub, R. M., McAuley, J. H., Localio, A. R., van Amelsvoort, L., Guallar, E., Rijnhart, J., Goldsmith, K., Fairchild, A. J., Lewis, C. C., Kamper, S. J., . . . Henschke, N. (2021). A guideline for reporting mediation analyses of randomized trials and observational studies: The AGReMA statement. JAMA, 326(11), 1045–1056. https://doi.org/10.1001/jama.2021.14075
• MacCorquodale, K., & Meehl, P. E. (1948). On a distinction between hypothetical constructs and intervening variables. Psychological Review, 55(2), 95–107. https://doi.org/10.1037/h0056029
• MacKinnon, D. P., Lockwood, C. M., Hoffman, J. M., West, S. G., & Sheets, V. (2002). A comparison of methods to test mediation and other intervening variable effects. Psychological Methods, 7(1), 83–104. https://doi.org/10.1037/1082-989X.7.1.83
• MacKinnon, D. P., Lockwood, C. M., & Williams, J. (2004). Confidence limits for the indirect effect: Distribution of the product and resampling methods. Multivariate Behavioral Research, 39(1), 99–128. https://doi.org/10.1207/S15327906MBR3901_4
• Preacher, K. J., & Hayes, A. F. (2008). Asymptotic and resampling strategies for assessing and comparing indirect effects in multiple mediator models. Behavior Research Methods, 40(3), 879–891. https://doi.org/10.3758/BRM.40.3.879