Mediators vs Moderators: Definitions, Differences, Examples, Calculations

Getting your Trinity Audio player ready...

Contents

• Introduction
• Basic Definitions
• Visualizing the Difference Between Mediators and Moderators
• Statistical Approaches to Mediation
• Statistical Approaches to Moderation
• Requirements for a Variable to Be a Mediator or Moderator
• Confounders vs. Mediators vs. Moderators
• Combining Mediation and Moderation
• Practical Examples Across Disciplines
• Common Mistakes and Misconceptions
• Summary Table
• Frequently Asked Questions
• References

Introduction

In behavioral, social, health, and organizational research, it is rarely enough to know that an independent variable (IV) is related to a dependent variable (DV). Researchers also want to know how and why that relationship occurs, and under what conditions it gets stronger, weaker, or disappears altogether. Two statistical concepts answer these questions: mediators, which explain the mechanism or process underlying a relationship, and moderators, which explain the conditions or boundary limits of a relationship.

These two terms are among the most confused in quantitative research, partly because they sound similar, partly because both involve a “third variable,” and partly because both can be tested using regression-based methods. However, they answer fundamentally different questions, sit in different positions in a causal diagram, are tested with different statistical procedures, and require different interpretations.

This guide draws together the core ideas found across the classic methodological literature—including Baron and Kenny’s (1986) foundational paper, James and Brett’s (1984) work on mediators versus moderators, MacKinnon and colleagues’ work on mediation analysis, Hayes’ modern approach to conditional process analysis, and various applied teaching resources—to give a single, thorough explanation of mediation, moderation, the difference between them, how each is tested, and how they can be combined.

Basic Definitions

What Is a Mediator?

A mediator (or mediating variable, M) is a variable that lies between the independent variable and the dependent variable in a causal chain. It explains the process or mechanism by which the IV influences the DV. Symbolically, the relationship looks like:

X → M → Y

The independent variable (X) causes the mediator (M), and the mediator, in turn, causes the dependent variable (Y). The mediator is sometimes called an intervening variable because it intervenes between the cause and the effect.

Example:

stress (X) leads to illness (Y) because stress reduces immune function (M). Here, immune function is the mediator—it is the “how” that connects stress to illness. Without the mediator, we would only know that stress and illness are correlated; with the mediator, we understand the biological pathway.

Mediators answer the question: “By what process or mechanism does X affect Y?”

What Is a Moderator?

A moderator (Z or W) is a variable that affects the strength or direction of the relationship between the independent and dependent variables. Rather than sitting on the causal pathway between X and Y, a moderator interacts with X to produce different effects on Y depending on the level of the moderator. This is statistically represented as an interaction effect:

Y = b1(X) + b2(Z) + b3(X×Z) + error

If the interaction term (X×Z) is statistically significant, the moderator is said to moderate the relationship between X and Y. A moderator essentially answers: “For whom, under what circumstances, or in what context does the relationship between X and Y hold, strengthen, weaken, or reverse?”

Example:

Social support (X) reduces stress (Y), but this effect is stronger for people who are more introverted (Z) than for those who are more extroverted. Personality type does not explain how social support reduces stress. It simply changes how strongly that relationship operates depending on who you are.

The Core Distinction in One Sentence

The most widely repeated heuristic across the literature is this:

Mediators explain “how” or “why” an effect occurs (the mechanism); moderators explain “when,” “for whom,” or “under what conditions” an effect occurs or is stronger/weaker (the boundary conditions).

Visualizing the Difference Between Mediators and Moderators

Path Diagrams

Mediation model:

        a            b
   X ——-> M ——-> Y
    \                    ^
     \___________________/
              c (direct effect)

Path a: effect of X on M

Path b: effect of M on Y (controlling for X)

Path c: the total/direct effect of X on Y

Moderation model:

         Z
         |
         v
   X ——–> Y

Here Z is drawn as a variable that influences the arrow itself, the relationship between X and Y. Z is not a step on a path from X to Y. In regression terms, moderation is modeled by adding a multiplicative interaction term (X × Z) as a predictor of Y.

The Light Switch Analogy

A particularly useful analogy (popularized in introductory statistics teaching) imagines a light switch (X) and a light bulb (Y):

• The mediator is the electrical wiring connecting the switch to the bulb; it explains how flipping the switch causes the bulb to light up.
• The moderator is a dimmer dial. It doesn’t explain how electricity flows, but it changes how bright the bulb gets when the switch is flipped; i.e., it changes the strength of the X→Y relationship.

Different Statistical Signatures

Feature	Mediator	Moderator
Position in causal chain	Between X and Y (intervening)	Outside the X→Y pathway
Statistical representation	Indirect effect (a × b)	Interaction term (X × Z)
Relationship to X	Caused by X (and itself causes Y)	Independent of X (not caused by X)
Type of variable	Often continuous, can be a process/state variable	Often categorical (e.g., gender, group) but can be continuous
Question answered	Why/how does X affect Y?	When/for whom does X affect Y (more or less)?
Typical analysis	Path analysis, indirect effects, bootstrapping	Multiple regression with interaction term, simple slopes

Statistical Approaches to Mediation

The Baron and Kenny (1986) Causal Steps Approach

The most historically influential approach to testing mediation comes from Baron and Kenny’s 1986 paper “The Moderator-Mediator Variable Distinction in Social Psychological Research,” which set out a now-classic four-step (“causal steps”) procedure using three regression equations:

1. Equation 1: Regress the mediator (M) on the independent variable (X). This tests path a and must be significant—X must significantly predict M.
2. Equation 2: Regress the dependent variable (Y) on the independent variable (X). This tests path c—X must significantly predict Y (i.e., there must be an effect to mediate).
3. Equation 3: Regress Y on both X and M simultaneously. This tests path b (effect of M on Y controlling for X) and path c′ (the direct effect of X on Y controlling for M).

Mediation is supported when:

• X significantly predicts M (path a is significant)
• M significantly predicts Y when controlling for X (path b is significant)
• The effect of X on Y (path c′) is reduced when M is included in the model, compared to path c

If the direct effect c′ becomes zero (non-significant) after adding M, this is described as full (complete) mediation—the mediator fully accounts for the relationship between X and Y. If c′ is reduced but still significant, this is partial mediation—the mediator explains part, but not all, of the relationship.

Limitations of the Causal Steps Approach

While historically dominant, the causal steps approach has been criticized on several grounds, which the methodological literature (particularly work building on MacKinnon’s research and subsequent reviews) has emphasized:

• It does not directly test the indirect effect (a × b); it only infers mediation through a pattern of significance tests across multiple regressions.
• It has relatively low statistical power, especially in small samples, because it requires several conditions to all be statistically significant.
• The requirement that X must significantly predict Y (step 2) has been challenged. Some forms of mediation (especially inconsistent mediation, where direct and indirect effects have opposite signs) can occur even when the total effect (c) is non-significant.
• It does not provide a confidence interval or standard error for the indirect effect itself.

The Product-of-Coefficients (Sobel Test) Approach

An alternative and more direct approach is to estimate the indirect effect as the product of path a (X→M) and path b (M→Y, controlling for X), i.e., a × b. The Sobel test provides a formal significance test for this product term by estimating its standard error analytically and computing a z-statistic.

However, the Sobel test assumes that the sampling distribution of the product a × b is normal—an assumption that is frequently violated, because the product of two normally distributed variables is not itself normally distributed (it tends to be skewed). This means the Sobel test can be conservative (underpowered) in small to moderate samples.

Bootstrapping (the Modern Standard)

Because of the limitations above, bootstrapping has become the recommended method for testing indirect effects (this is the approach championed in modern mediation analysis, including Preacher and Hayes’ work and Hayes’ PROCESS macro for SPSS/SAS/R). Bootstrapping:

• Resamples the dataset (with replacement) thousands of times (e.g., 5,000–10,000 resamples).
• Recalculates the indirect effect (a × b) for each resample, building an empirical sampling distribution of a × b.
• Constructs a confidence interval (commonly a 95% bias-corrected confidence interval) directly from this empirical distribution.
• If the confidence interval does not contain zero, the indirect effect is considered statistically significant.

Bootstrapping does not require the assumption of a normally distributed indirect effect, has greater statistical power than the Sobel test, and does not require that the total effect (c) be significant—making it suitable for detecting mediation even in cases of suppression or inconsistent mediation.

Full vs. Partial Mediation, and a Modern Caveat

The traditional distinction is:

• Full (complete) mediation: the direct effect of X on Y becomes non-significant (essentially zero) after accounting for M—the entire effect operates through the mediator.
• Partial mediation: the direct effect of X on Y remains significant but is reduced after accounting for M—part of the effect operates through M, and part operates through other, unmeasured pathways.

A modern caveat raised in the methodological literature is that “full” versus “partial” mediation is somewhat sample- and model-dependent. The same underlying process could appear as “full” mediation in one study and “partial” in another simply due to sampling variability, the presence of other unmeasured mediators, or measurement error. As a result, many methodologists now recommend focusing on the size and significance of the indirect effect itself (and the proportion of the total effect it represents) rather than treating “full vs. partial” as a fixed property of the phenomenon.

Requirements and Assumptions for Mediation Analysis

Drawing on points emphasized across the literature (including discussions of confounding, mediation, and moderation aimed at applied researchers), mediation analysis carries several important assumptions:

• Temporal/causal ordering: X should precede M, which should precede Y—ideally established through longitudinal design, experimental manipulation, or strong theoretical justification, since cross-sectional data cannot establish causal order.
• No unmeasured confounding: there should be no third variable that causes both M and Y (or both X and M, or X and Y) that has not been accounted for, as this can produce spurious “mediation.”
• Reliable measurement: measurement error in M can bias estimates of both the a and b paths (typically attenuating the b path and inflating the apparent direct effect c′).
• Correct functional form: the relationships among X, M, and Y are typically assumed to be linear (though non-linear and moderated mediation models exist).

Statistical Approaches to Moderation

Moderated Multiple Regression

Moderation is most commonly tested using moderated multiple regression, in which the dependent variable is regressed on:

• The independent variable (X)
• The moderator (Z)
• The interaction term (X × Z), computed as the product of X and Z

Y = b0 + b1(X) + b2(Z) + b3(X×Z) + e

The key test is whether b3 (the coefficient for the interaction term) is statistically significant. If it is, this indicates that the relationship between X and Y depends on the level of Z—i.e., Z moderates the X–Y relationship.

Centering Variables

Before computing the interaction term, both X and Z are typically mean-centered (i.e., the mean is subtracted from each score so the variable has a mean of zero). This is recommended because:

• It reduces multicollinearity between the main effect terms and the interaction term, which can otherwise be severe when X and Z are correlated with the product term X×Z.
• It makes the main effect coefficients (b1 and b2) more interpretable—they now represent the effect of one variable at the mean of the other, rather than at the (often meaningless) value of zero.

Simple Slopes Analysis

When an interaction is significant, researchers typically follow up with a simple slopes analysis, which examines the relationship between X and Y at specific, theoretically meaningful values of the moderator—commonly at the mean of Z, and at one standard deviation above and below the mean (representing “low,” “average,” and “high” levels of the moderator). This produces separate regression lines for each level of the moderator, which are often visualized in an interaction plot showing how the slope of the X–Y line changes across levels of Z.

Types of Moderators

• Categorical moderators: e.g., gender, treatment group, ethnicity. The moderation analysis essentially tests whether the X–Y relationship differs across groups (similar in spirit to testing for an interaction in ANOVA).
• Continuous moderators: e.g., age, personality trait scores, severity of symptoms. The moderation analysis tests whether the slope of the X–Y relationship changes continuously as Z increases.

A moderator can enhance (strengthen), diminish (weaken/buffer), or reverse (cross-over interaction) the direction of the relationship between X and Y. It can also reveal that a relationship exists only at certain levels of the moderator and is absent (non-significant) at others.

ANOVA as a Special Case of Moderation

Several sources note that, conceptually, testing for an interaction effect in a factorial ANOVA is mathematically equivalent to testing for moderation: if the effect of one factor on the outcome differs depending on the level of a second factor, that second factor is acting as a moderator. This connects moderation analysis to a very familiar statistical framework for many researchers trained primarily in experimental design.

Requirements for a Variable to Be a Mediator or Moderator

Conditions for a Mediator (per Baron & Kenny and subsequent work)

For M to function as a mediator between X and Y:

1. X must be significantly associated with M.
2. M must be significantly associated with Y, controlling for X.
3. The relationship between X and Y must be attenuated (reduced toward zero) when M is included in the model.
4. Ideally, M should occur temporally after X and before Y.

Conditions for a Moderator

For Z to function as a moderator between X and Y:

1. Z should not be highly correlated with X (i.e., Z is not simply another cause of X, and is conceptually independent of the causal chain between X and Y)—a moderator is, in principle, uncorrelated with both X and Y in a “pure” moderation design, though this strict requirement is relaxed in much applied work.
2. The interaction term (X × Z) must significantly predict Y, over and above the main effects of X and Z.
3. Z is not caused by X (this is one of the key distinctions from a mediator, which is caused by X).

Distinguishing a Mediator from a Moderator: The Key Diagnostic Questions

Several of the sources converge on a small set of diagnostic questions a researcher can ask about a third variable to determine whether it functions as a mediator or moderator:

• Does the third variable come “after” X and “before” Y in a causal/temporal sequence? If yes, it may be a mediator. If it exists independently of, or “alongside,” X (e.g., a stable trait, demographic characteristic, or contextual factor), it is more likely a moderator.
• Is the third variable correlated with (caused by) X? Mediators are caused by X; moderators generally are not.
• Does including the third variable in the model reduce the X→Y relationship (suggesting it explains part of the mechanism), or does it change the strength/direction of the X→Y relationship depending on its own value (suggesting it’s a boundary condition)?
• What does your theory predict? Ultimately, whether a variable is treated as a mediator or moderator is a theoretical decision, not purely a statistical one. The same variable (e.g., coping style) could in principle be modeled as a mediator in one theoretical framework (it explains how a stressor leads to anxiety) or as a moderator in another (it explains for whom the stressor-anxiety link is stronger).

Confounders vs. Mediators vs. Moderators

A related but distinct concept is the confounding variable (or confounder), and distinguishing it from mediators and moderators is a recurring theme in methodological teaching resources.

• A confounder is a variable that is associated with both the independent variable and the dependent variable, and which provides an alternative explanation for an observed X–Y association. Unlike a mediator, a confounder is not part of the causal pathway from X to Y—it is a common cause of both. If a confounder is not controlled for, it can create a spurious (non-causal) association between X and Y, or mask a true association.
• A mediator, by contrast, is part of the causal pathway: X causes M, and M causes Y. Controlling for a mediator in the same way one controls for a confounder (i.e., simply adding it to a regression model) is a common mistake known as “controlling away” the effect of interest—because doing so removes part of the true causal effect of X on Y, not just a spurious association.
• A moderator is neither a confounder nor positioned on the causal path; it interacts with X to change the size of X’s effect on Y, but it does not by itself create an association between X and Y.

The practical implication is important: the same statistical action (adding a third variable to a regression model) can have very different meanings depending on the variable’s true role. Adding a confounder appropriately removes bias; “adding” (i.e., controlling for) a mediator removes part of the genuine effect you may be trying to estimate; and a moderator should be modeled via an interaction term, not simply added as another “control.”

Combining Mediation and Moderation

Real-world processes are often more complex than a single mediator or moderator can capture. Two important combined models appear repeatedly in the literature:

Moderated Mediation (Conditional Indirect Effects)

Moderated mediation occurs when the indirect effect of X on Y through a mediator M depends on the level of a fourth variable (the moderator). In other words, mediation itself is conditional—the mechanism linking X to Y operates more strongly (or only) for certain groups or under certain conditions.

Example:

Early childhood physical abuse (X) may lead to violent behavior in adulthood (Y) through deviant social information processing (M), but this indirect pathway (the mediated effect) may be stronger for males than for females, with gender (W) acting as the moderator of the a path, the b path, or both.

Moderated mediation models can take several forms depending on where the moderator enters the model:

• The moderator affects the a path (X→M): the strength of the mechanism’s “trigger” depends on the moderator.
• The moderator affects the b path (M→Y): the strength of the mechanism’s “consequence” depends on the moderator.
• The moderator affects both paths.

These models are often summarized using an index of moderated mediation, which tests whether the indirect effect differs significantly across levels of the moderator.

Mediated Moderation

Mediated moderation is, in a sense, the reverse: it occurs when an overall interaction effect (moderation) between X and Z on Y is itself explained by an intervening mediator. That is, the moderating effect of Z on the X–Y relationship operates through a mediating mechanism M.

While the conceptual distinction between moderated mediation and mediated moderation can seem subtle, the practical difference lies in the research question: moderated mediation asks “does the mechanism (mediation) depend on context?”, while mediated moderation asks “is the interaction effect itself explained by a mechanism?” In practice, moderated mediation is far more commonly tested and reported in applied research.

Conditional Process Analysis

The umbrella term often used in contemporary statistics (particularly associated with Hayes’ work and software tools such as the PROCESS macro) for models that combine mediators and moderators in any configuration is conditional process analysis (or conditional process modeling). This framework allows researchers to specify complex models with multiple mediators, multiple moderators, and moderators acting on different paths simultaneously, and to test them using bootstrapping methods for the relevant conditional indirect and direct effects.

Practical Examples Across Disciplines

To consolidate the concepts, here are worked examples spanning different fields, illustrating both mediation and moderation:

Example 1: Health Psychology (Mediation)

• IV: Job stress
• Mediator: Cortisol levels
• DV: Cardiovascular disease risk

Interpretation: Job stress increases cortisol levels, which in turn increases cardiovascular disease risk. Cortisol is the biological mechanism explaining how stress translates into physical health outcomes.

Example 2: Educational Psychology (Moderation)

• IV: Amount of feedback given to students
• Moderator: Students’ baseline self-efficacy
• DV: Academic performance

Interpretation: Feedback improves performance more strongly for students with low self-efficacy than for those with high self-efficacy (or vice versa)—self-efficacy moderates (changes the strength of) the feedback–performance relationship, but it doesn’t explain the mechanism by which feedback works.

Example 3: Social/Well-being Research (Mediation, drawing on quality-of-life research)

• IV: Income
• Mediator: Perceived control over one’s life / sense of autonomy
• DV: Life satisfaction

Interpretation: Higher income leads to a greater sense of autonomy and control, which in turn leads to higher life satisfaction. The psychological sense of control is the mechanism—the “why”—behind the income–life satisfaction link. This type of model is common in social indicators and quality-of-life research, where the goal is often to identify the psychological or social processes that translate objective circumstances (like income) into subjective well-being.

Example 4: Organizational Behavior (Moderated Mediation)

• IV: Transformational leadership
• Mediator: Employee trust in management
• DV: Employee performance
• Moderator: Organizational tenure

Interpretation: Transformational leadership increases trust, which increases performance (mediation), but this indirect effect is stronger for employees with longer tenure, who have had more opportunities to observe consistent leadership behavior (moderated mediation).

Common Mistakes and Misconceptions

Drawing together cautionary points raised across the literature:

Treating every “third variable” the same way

Confounders, mediators, and moderators require different statistical treatments (controlling for, modeling as an intervening variable, or modeling as an interaction, respectively). Misclassifying a variable can lead to incorrect conclusions—e.g., “controlling for” a true mediator can make a real effect disappear, leading a researcher to wrongly conclude there is no relationship between X and Y.

Assuming cross-sectional data can establish mediation

Because mediation implies a causal, temporal sequence (X → M → Y), data collected at a single time point cannot definitively establish that M mediates the X–Y relationship, even if the statistical pattern fits. Longitudinal or experimental designs provide much stronger evidence.

Relying solely on the causal steps (Baron & Kenny) approach

While historically important and still useful conceptually, the causal steps approach is now considered less rigorous than directly testing the indirect effect with bootstrapped confidence intervals.

Forgetting to center variables before computing interaction terms

This can produce misleading main-effect coefficients and inflate multicollinearity.

Confusing the direction of the question

A simple way to check: if you find yourself asking “why” or “through what process,” you are likely dealing with a potential mediator; if you find yourself asking “for whom” or “under what conditions,” you are likely dealing with a potential moderator.

Ignoring theory

Statistical tests cannot, by themselves, tell you whether a variable should be conceptualized as a mediator or a moderator. This decision must be grounded in substantive theory and the logic of the research question, with statistics used to test the theoretically derived model.

Summary Table

Question	Mediator	Moderator
What does it explain?	The mechanism/process (how/why X affects Y)	The boundary conditions (when/for whom X affects Y)
Causal position	Between X and Y (X → M → Y)	Outside the X–Y pathway; interacts with X
Caused by X?	Yes	Typically no
Statistical test	Indirect effect (a×b); bootstrapped CIs; Sobel test (older method)	Interaction term (X×Z) in regression; simple slopes
Common terms for the combined model	Moderated mediation, mediated moderation, conditional process analysis	Same
Example	Exercise → endorphins → improved mood	Social support → reduced stress, stronger for introverts
Key risk if misclassified	“Controlling away” a real effect	Missing a true interaction by treating Z as a simple control

Frequently Asked Questions

What is the simplest way to remember the difference between a mediator and a moderator?

Mediators explain “how” or “why” (the mechanism); moderators explain “when” or “for whom” (the conditions). A mediator sits between X and Y on the causal path; a moderator sits alongside and changes the strength of the X–Y relationship.

Can the same variable be both a mediator and a moderator?

Yes, but not in the same model role at the same time. Whether a variable is treated as a mediator or moderator depends on theory and the research question—e.g., coping style could mediate a stressor-anxiety link in one study or moderate it in another.

How is mediation tested statistically?

Historically via Baron and Kenny’s causal steps regression approach. The modern standard is to estimate the indirect effect (a × b) directly and test it using bootstrapped confidence intervals, since this has greater power and doesn’t assume normality.

How is moderation tested statistically?

Via moderated multiple regression, adding an interaction term (X × Z) to the model. If this term is significant, Z moderates the X–Y relationship. Variables are usually mean-centered first to reduce multicollinearity.

What’s the difference between full and partial mediation?

Full mediation means the direct effect of X on Y becomes non-significant once the mediator is included—the relationship operates entirely through M. Partial mediation means the direct effect is reduced but remains significant, meaning M explains part of the relationship.

How is a mediator different from a confounder?

A confounder is a common cause of both X and Y and sits outside the causal pathway—controlling for it removes bias. A mediator sits on the causal pathway (X → M → Y)—controlling for it removes part of the real effect you’re trying to estimate.

What is moderated mediation?

A model where the indirect effect of X on Y through M depends on the level of a fourth variable (moderator)—i.e., the mechanism itself is conditional on context (e.g., stronger for one group than another).

What is mediated moderation?

The reverse situation: an overall interaction effect (X × Z on Y) is itself explained by an intervening mechanism (M).

Can cross-sectional data establish mediation?

Not definitively. Mediation implies temporal/causal ordering (X before M before Y), which cross-sectional data cannot prove, even if the statistical pattern fits. Longitudinal or experimental designs provide stronger evidence.

Why is the Sobel test less favored now?

It assumes the indirect effect (a × b) is normally distributed, which is often untrue (the product of two normal variables is typically skewed), making it underpowered compared to bootstrapping.

References

1. Baron RM, Kenny DA. The moderator-mediator variable distinction in social psychological research: conceptual, strategic, and statistical considerations. J Pers Soc Psychol. 1986;51(6):1173-82.
2. MacKinnon DP, Fairchild AJ, Fritz MS. Mediation analysis. Annu Rev Psychol. 2007;58:593-614. Available from: https://pmc.ncbi.nlm.nih.gov/articles/PMC3366634/
3. Bhandari P. Mediator vs. moderator variables [Internet]. Scribbr; 2022 [updated 2023]. Available from: https://www.scribbr.com/methodology/mediator-vs-moderator/
4. McLeod S, Guy-Evans O. Mediating vs moderating variables [Internet]. Simply Psychology; 2025. Available from: https://www.simplypsychology.org/mediating-vs-moderating-variables.html
5. Statistics Solutions. Similarities and differences between mediation and moderation analyses [Internet]. Available from: https://www.statisticssolutions.com/similarities-and-differences-between-mediation-and-moderation-analyses/
6. Wu AD, Zumbo BD. Understanding and using mediators and moderators. Soc Indic Res. 2008;87:367-92.
7. York St John University, Study Skills Service. Confounding, mediator and moderator variables [Internet]. Available from: https://www.yorksj.ac.uk/media/content-assets/study-skills/maths-and-statistics/data-handling/Confounding_-mediator-and-moderator-variables.pdf
8. Baron RM, Kenny DA. The moderator-mediator variable distinction [Internet]. Society for Experimental Social Psychology. Available from: https://www.sesp.org/files/The%20Moderator-Baron.pdf
9. University of Southampton Library. Mediation vs moderation [Internet]. Available from: https://library.soton.ac.uk/mediation-vs-moderation
10. Lea M. Moderation and mediation explained [Internet]. Available from: https://martinlea.com/moderation-and-mediation-explained/