Concepts · Evidence and Explanation

Mediators, Moderators and Mechanisms

Mediation asks what lies on an intermediate pathway between a candidate cause and an outcome. Moderation asks when, where or for whom an effect differs. Mechanism asks how the effect is produced.

By Yona Ole Lobulu ·

Concept6 min readFoundationalD2.5

Topic
Evidence and Explanation
Read first
One piece should be read before this one
Reading time
About 6 minutes of reading
Difficulty
Advanced reading: this page assumes a fair amount of earlier reading

The question

Through what pathway, under what conditions, and by what process does an effect occur?

Definition

Mediation concerns intermediate pathways, moderation concerns conditional variation, and mechanism concerns the process producing an effect. None should be upgraded into another without evidence appropriate to that inference.

Suppose an intervention changes how students study, and their exam performance later improves.

Researchers might ask three different questions.

Did the intervention increase active retrieval, which in turn contributed to better performance?

Did the intervention work differently depending on students' baseline knowledge or study conditions?

What process actually produced the improvement?

These are questions about mediation, moderation and mechanism.

They concern the same effect, but they ask different things.

Mediation asks about an intermediate pathway

A mediator is an intermediate variable through which part of the relation or causal effect between a candidate cause and an outcome may operate.

The simplest schematic is X → M → Y, where X is the candidate cause or intervention, M is the mediator, and Y is the outcome.

In the study example, the intervention might increase active retrieval, which is then related to improved exam performance.

The mediation question is: through what intermediate variable or pathway might part of the effect operate?

Mediation can narrow the search for how an effect occurs by identifying candidate intermediate pathways.

But the arrows in a mediation model do not establish their own causal interpretation.

A mediation analysis may reveal an indirect statistical relationship consistent with a proposed pathway. Whether that pathway can be interpreted causally depends on the design, measurement and assumptions supporting the inference.

Even if the intervention itself is experimentally manipulated, the mediator often is not. Other factors may influence both the mediator and the outcome.

So a statistical indirect effect is not automatically a causal indirect effect.

Or more generally:

A statistical pathway is not automatically a causal process.

This does not make statistical mediation uninformative. It means the strength of the claim should match the strength of the evidence.

A mediator is not automatically the mechanism

Suppose there is good evidence that the intervention increases active retrieval and that this increase contributes causally to improved performance.

We have learned something important about the pathway through which the intervention works.

We still may not have identified the complete mechanism.

The broader process could involve changes in attention, memory, feedback, practice and other interacting components.

A mediator can be part of a mechanism without being the mechanism.

A mediator may capture one component of a process, one pathway among several, or an imperfect proxy for a broader process.

The simple structure X → M → Y should therefore not be mistaken for a complete map of causal reality.

The same caution applies to time.

If X occurs before M and M before Y, that ordering may support a proposed causal pathway. But temporal sequence alone does not establish that X caused M, that M caused Y, or that M constitutes the mechanism.

A sequence in time is not yet an explanation of process.

Moderation asks when or for whom an effect differs

Now suppose the same study intervention produces larger improvements for students with higher baseline knowledge, or works differently under different study conditions.

The question has changed: when, where or for whom does the effect differ?

This is moderation.

A moderator is a variable or condition across which the size, direction or presence of an effect differs.

Moderation therefore concerns conditional effects and can reveal effect heterogeneity: an effect need not be uniform across all people, contexts or conditions.

That variation can matter scientifically and practically even before its cause is understood.

A moderator does not automatically explain the difference

Suppose the intervention works better for students with higher baseline knowledge.

That tells us something important: the effect differs according to baseline knowledge.

It does not yet tell us why.

Baseline knowledge might itself play a causal role. But it could also mark differences in previous experience, learning strategies, opportunity, motivation or other relevant conditions.

A moderator can identify where an effect differs without identifying what causes that difference.

Moderation can therefore reveal a pattern that calls for further explanation without providing that explanation by itself.

A moderator can be causal.

But demonstrating moderation does not automatically establish that causal role.

Mechanism asks how the effect is produced

A mechanism is an organised process or set of causal relations through which an effect or phenomenon is produced.

Its question is: by what process is the effect produced?

This is broader than identifying one intermediate variable or one condition under which the effect changes.

A mechanism may contain several interacting components or pathways. It need not reduce to one measured mediator, and it need not exist at one particular level of explanation. Relevant mechanisms can be behavioural, cognitive, biological, interpersonal or ecological, or span several levels at once.

Mechanistic evidence must therefore bear on the process producing the effect, not merely on the existence of an intermediate association.

The three concepts can coexist

Distinguishing mediation, moderation and mechanism does not mean only one can apply at a time.

A mediator may participate in a mechanism. A mediated pathway may itself vary across people or conditions. A mechanism may operate differently depending on context.

Their distinction concerns the question each answers, not a competition among three explanations.

These are different explanatory achievements, not a ranking of scientific value.

Three explanatory questions

Through what intermediate variable or pathway might part of the effect operate? — Mediation.

When, where or for whom does the effect differ? — Moderation.

By what process is the effect produced? — Mechanism.

The distinction can be stated compactly: mediation concerns intermediate pathways, moderation concerns conditional variation, and mechanism concerns the process producing an effect.

Evidence about one can inform the others.

But none should simply be upgraded into another kind of explanation without evidence appropriate to that inference.

Ask which question the evidence actually answered

Research language can make these distinctions disappear.

A study identifies an indirect effect and calls it a mechanism.

Another finds that an effect differs across groups and says it has explained who responds.

Another observes the expected temporal sequence and treats the causal process as established.

In each case, the useful question is not how sophisticated the analysis sounds.

It is what the evidence actually established.

If M mediated the effect, did the analysis identify an indirect statistical relationship, or is its causal interpretation justified?

If Z moderated the effect, did the study identify where the effect differs, or establish why it differs?

If M is described as the mechanism, what evidence actually bears on the process producing the effect?

Mediators, moderators and mechanisms can all deepen our understanding of human change. Their value depends partly on keeping their contributions distinct.

Do not confuse an intermediate pathway, a condition of variation and a process of production.

Sources and research record8 sources, with findings, strengths and limitations as entered

References

8 sources this piece rests on, as entered in the Library.

  1. Baron, R. M., Kenny, D. A. (1986) The moderator–mediator variable distinction in social psychological research: Conceptual, strategic, and statistical considerations

    Methodological article · Journal of Personality and Social Psychology, 51(6) · 1173–1182

    The reference statement of mediation as an intermediate variable in a pathway, used here only to mark that identifying a mediator is not the same as identifying a mechanism.

    doi:10.1037/0022-3514.51.6.1173

  2. MacKinnon, D. P. (2008) Introduction to Statistical Mediation Analysis

    Book · Lawrence Erlbaum Associates, New York

    The standard treatment of mediation and indirect effects, setting out what an estimated indirect effect does and does not establish about process.

    Read the source

  3. Imai, K., Keele, L., Tingley, D. (2010) A general approach to causal mediation analysis

    Methodological article · Psychological Methods, 15(4) · 309–334

    Formalises the identification assumptions required to read an indirect effect causally, making explicit the gap between statistical and causal mediation.

    doi:10.1037/a0020761

  4. VanderWeele, T. J. (2015) Explanation in Causal Inference: Methods for Mediation and Interaction

    Methodological text · Oxford University Press, New York

    Shows that a total intervention effect and the pathway producing it are separate estimands, so mechanism requires evidence beyond the demonstration that the intervention changed the outcome.

    Read the source

  5. Fairchild, A. J., MacKinnon, D. P. (2009) A general model for testing mediation and moderation effects

    Methodological article · Prevention Science, 10(2) · 87–99

    Sets out mediation and moderation terminology together, clarifying that they answer different questions about an effect.

    doi:10.1007/s11121-008-0109-6

  6. Bansak, K. (2021) Estimating causal moderation effects with randomized treatments and non-randomized moderators

    Methodological article · Journal of the Royal Statistical Society: Series A, 184(1) · 65–86

    Distinguishes descriptive effect heterogeneity across an unrandomised moderator from a causal moderation claim about that moderator.

    doi:10.1111/rssa.12614

  7. Craver, C. F., Darden, L. (2013) In Search of Mechanisms: Discoveries across the Life Sciences

    Book · University of Chicago Press, Chicago

    Develops mechanistic explanation across multiple levels of organisation, supporting the point that mechanisms need not be linear chains and need not be located at a neural level.

    Read the source

  8. Machamer, P., Darden, L., Craver, C. F. (2000) Thinking about mechanisms

    Theoretical article · Philosophy of Science, 67(1) · 1–25

    One influential account of mechanisms as organised entities and activities productive of a phenomenon, used here as a representative tradition rather than the sole philosophical definition of mechanism.

    doi:10.1086/392759

Further reading

Behind this page

The claims this concept makes, the evidence behind them, and the limits it accepts.

Evidence status

High confidence

Strongly supported, though resting on synthesis or principle rather than a single decisive body of evidence.

Claims

  1. A mediator is an intermediate variable through which part of an effect may operate

    Established

    What this does not assert: The mediation model specifies a candidate pathway; it does not by itself certify that the pathway is causal.

    1. Baron, R. M., Kenny, D. A. (1986) The moderator–mediator variable distinction in social psychological research: Conceptual, strategic, and statistical considerations

      Methodological article · Journal of Personality and Social Psychology, 51(6) · 1173–1182

      The reference statement of mediation as an intermediate variable in a pathway, used here only to mark that identifying a mediator is not the same as identifying a mechanism.

      doi:10.1037/0022-3514.51.6.1173

    2. MacKinnon, D. P. (2008) Introduction to Statistical Mediation Analysis

      Book · Lawrence Erlbaum Associates, New York

      The standard treatment of mediation and indirect effects, setting out what an estimated indirect effect does and does not establish about process.

      Read the source

  2. A moderator can nonetheless be causal

    High confidence

    What this does not assert: Causal moderation is possible and sometimes demonstrable; it is simply not implied by the observation of moderation.

    1. Bansak, K. (2021) Estimating causal moderation effects with randomized treatments and non-randomized moderators

      Methodological article · Journal of the Royal Statistical Society: Series A, 184(1) · 65–86

      Distinguishes descriptive effect heterogeneity across an unrandomised moderator from a causal moderation claim about that moderator.

      doi:10.1111/rssa.12614

  3. A mechanism is an organised process or set of causal relations through which an effect is produced

    High confidence

    What this does not assert: This reflects one influential tradition in philosophy of science rather than a single agreed formal definition.

    1. Machamer, P., Darden, L., Craver, C. F. (2000) Thinking about mechanisms

      Theoretical article · Philosophy of Science, 67(1) · 1–25

      One influential account of mechanisms as organised entities and activities productive of a phenomenon, used here as a representative tradition rather than the sole philosophical definition of mechanism.

      doi:10.1086/392759

    2. Craver, C. F., Darden, L. (2013) In Search of Mechanisms: Discoveries across the Life Sciences

      Book · University of Chicago Press, Chicago

      Develops mechanistic explanation across multiple levels of organisation, supporting the point that mechanisms need not be linear chains and need not be located at a neural level.

      Read the source

  4. Mechanisms need not reduce to one measured mediator or one level of explanation

    High confidence

    What this does not assert: Relevant mechanisms may be behavioural, cognitive, biological, interpersonal or ecological, or span several levels.

    1. Craver, C. F., Darden, L. (2013) In Search of Mechanisms: Discoveries across the Life Sciences

      Book · University of Chicago Press, Chicago

      Develops mechanistic explanation across multiple levels of organisation, supporting the point that mechanisms need not be linear chains and need not be located at a neural level.

      Read the source

  5. Mediation, moderation and mechanism can coexist in the same phenomenon

    Canonical inference

    What this does not assert: A mediated pathway may vary across conditions, and a mechanism may operate differently by context.

    1. Fairchild, A. J., MacKinnon, D. P. (2009) A general model for testing mediation and moderation effects

      Methodological article · Prevention Science, 10(2) · 87–99

      Sets out mediation and moderation terminology together, clarifying that they answer different questions about an effect.

      doi:10.1007/s11121-008-0109-6

  6. Mediation, moderation and mechanism are different explanatory achievements, not a ranking of scientific value

    Canonical inference

    What this does not assert: They differ in the question answered and the inference licensed, not in worth.

  7. Mediation, moderation and mechanism answer three different explanatory questions

    Canonical synthesis

    What this does not assert: This three-question teaching architecture is The Shifting Point's explanatory synthesis, not a claim that research must proceed through three ordered stages.

  8. Do not confuse an intermediate pathway, a condition of variation and a process of production

    Canonical synthesis

    What this does not assert: Upgrading one into another requires evidence appropriate to that inference each time.

  9. A statistical indirect effect is not automatically a causal indirect effect

    Established

    What this does not assert: Causal interpretation requires identification assumptions that observational mediators rarely satisfy on their own.

    1. Imai, K., Keele, L., Tingley, D. (2010) A general approach to causal mediation analysis

      Methodological article · Psychological Methods, 15(4) · 309–334

      Formalises the identification assumptions required to read an indirect effect causally, making explicit the gap between statistical and causal mediation.

      doi:10.1037/a0020761

    2. VanderWeele, T. J. (2015) Explanation in Causal Inference: Methods for Mediation and Interaction

      Methodological text · Oxford University Press, New York

      Shows that a total intervention effect and the pathway producing it are separate estimands, so mechanism requires evidence beyond the demonstration that the intervention changed the outcome.

      Read the source

  10. Randomising the treatment does not randomise the mediator

    Established

    What this does not assert: Common causes of mediator and outcome can remain even in an experiment on the intervention itself.

    1. Imai, K., Keele, L., Tingley, D. (2010) A general approach to causal mediation analysis

      Methodological article · Psychological Methods, 15(4) · 309–334

      Formalises the identification assumptions required to read an indirect effect causally, making explicit the gap between statistical and causal mediation.

      doi:10.1037/a0020761

    2. VanderWeele, T. J. (2015) Explanation in Causal Inference: Methods for Mediation and Interaction

      Methodological text · Oxford University Press, New York

      Shows that a total intervention effect and the pathway producing it are separate estimands, so mechanism requires evidence beyond the demonstration that the intervention changed the outcome.

      Read the source

  11. Mediation analysis can still narrow the search for how an effect occurs

    High confidence

    What this does not assert: Its informative value depends on matching the strength of the claim to the strength of the design.

    1. MacKinnon, D. P. (2008) Introduction to Statistical Mediation Analysis

      Book · Lawrence Erlbaum Associates, New York

      The standard treatment of mediation and indirect effects, setting out what an estimated indirect effect does and does not establish about process.

      Read the source

  12. A mediator can be part of a mechanism without being the mechanism

    High confidence

    What this does not assert: A measured mediator may be one component, one pathway among several, or an imperfect proxy for a broader process.

    1. VanderWeele, T. J. (2015) Explanation in Causal Inference: Methods for Mediation and Interaction

      Methodological text · Oxford University Press, New York

      Shows that a total intervention effect and the pathway producing it are separate estimands, so mechanism requires evidence beyond the demonstration that the intervention changed the outcome.

      Read the source

    2. Craver, C. F., Darden, L. (2013) In Search of Mechanisms: Discoveries across the Life Sciences

      Book · University of Chicago Press, Chicago

      Develops mechanistic explanation across multiple levels of organisation, supporting the point that mechanisms need not be linear chains and need not be located at a neural level.

      Read the source

  13. Temporal sequence alone does not establish a causal process

    Established

    What this does not assert: Ordering can support a proposed pathway while leaving directionality, confounding and process unresolved.

    1. VanderWeele, T. J. (2015) Explanation in Causal Inference: Methods for Mediation and Interaction

      Methodological text · Oxford University Press, New York

      Shows that a total intervention effect and the pathway producing it are separate estimands, so mechanism requires evidence beyond the demonstration that the intervention changed the outcome.

      Read the source

  14. A moderator is a variable or condition across which the size, direction or presence of an effect differs

    Established

    What this does not assert: Moderation is a statement about conditional effects, not about the internal pathway of an effect.

    1. Fairchild, A. J., MacKinnon, D. P. (2009) A general model for testing mediation and moderation effects

      Methodological article · Prevention Science, 10(2) · 87–99

      Sets out mediation and moderation terminology together, clarifying that they answer different questions about an effect.

      doi:10.1007/s11121-008-0109-6

  15. Moderation can reveal effect heterogeneity

    Established

    What this does not assert: Variation across people, contexts or conditions can be scientifically and practically important before its cause is known.

    1. Bansak, K. (2021) Estimating causal moderation effects with randomized treatments and non-randomized moderators

      Methodological article · Journal of the Royal Statistical Society: Series A, 184(1) · 65–86

      Distinguishes descriptive effect heterogeneity across an unrandomised moderator from a causal moderation claim about that moderator.

      doi:10.1111/rssa.12614

  16. Observed moderation does not automatically establish that the moderator causes the difference

    Established

    What this does not assert: A non-randomised moderator can mark other differences in experience, opportunity or context.

    1. Bansak, K. (2021) Estimating causal moderation effects with randomized treatments and non-randomized moderators

      Methodological article · Journal of the Royal Statistical Society: Series A, 184(1) · 65–86

      Distinguishes descriptive effect heterogeneity across an unrandomised moderator from a causal moderation claim about that moderator.

      doi:10.1111/rssa.12614

Sources

  1. Baron, R. M., Kenny, D. A. (1986) The moderator–mediator variable distinction in social psychological research: Conceptual, strategic, and statistical considerations

    Methodological article · Journal of Personality and Social Psychology, 51(6) · 1173–1182

    The reference statement of mediation as an intermediate variable in a pathway, used here only to mark that identifying a mediator is not the same as identifying a mechanism.

    doi:10.1037/0022-3514.51.6.1173

  2. MacKinnon, D. P. (2008) Introduction to Statistical Mediation Analysis

    Book · Lawrence Erlbaum Associates, New York

    The standard treatment of mediation and indirect effects, setting out what an estimated indirect effect does and does not establish about process.

    Read the source

  3. Imai, K., Keele, L., Tingley, D. (2010) A general approach to causal mediation analysis

    Methodological article · Psychological Methods, 15(4) · 309–334

    Formalises the identification assumptions required to read an indirect effect causally, making explicit the gap between statistical and causal mediation.

    doi:10.1037/a0020761

  4. VanderWeele, T. J. (2015) Explanation in Causal Inference: Methods for Mediation and Interaction

    Methodological text · Oxford University Press, New York

    Shows that a total intervention effect and the pathway producing it are separate estimands, so mechanism requires evidence beyond the demonstration that the intervention changed the outcome.

    Read the source

  5. Fairchild, A. J., MacKinnon, D. P. (2009) A general model for testing mediation and moderation effects

    Methodological article · Prevention Science, 10(2) · 87–99

    Sets out mediation and moderation terminology together, clarifying that they answer different questions about an effect.

    doi:10.1007/s11121-008-0109-6

  6. Bansak, K. (2021) Estimating causal moderation effects with randomized treatments and non-randomized moderators

    Methodological article · Journal of the Royal Statistical Society: Series A, 184(1) · 65–86

    Distinguishes descriptive effect heterogeneity across an unrandomised moderator from a causal moderation claim about that moderator.

    doi:10.1111/rssa.12614

  7. Craver, C. F., Darden, L. (2013) In Search of Mechanisms: Discoveries across the Life Sciences

    Book · University of Chicago Press, Chicago

    Develops mechanistic explanation across multiple levels of organisation, supporting the point that mechanisms need not be linear chains and need not be located at a neural level.

    Read the source

  8. Machamer, P., Darden, L., Craver, C. F. (2000) Thinking about mechanisms

    Theoretical article · Philosophy of Science, 67(1) · 1–25

    One influential account of mechanisms as organised entities and activities productive of a phenomenon, used here as a representative tradition rather than the sole philosophical definition of mechanism.

    doi:10.1086/392759

Where to go from here

Next published piece

Group Averages and Individual Lives

Group-level evidence can change what is reasonable to expect about a person without determining that person’s outcome, mechanism or trajectory.

Continue through the Library →See where this sits in the graph →

Back to the Library →