Concepts · Evidence and Explanation

Correlation, Prediction, Causation and Mechanism

A variable can be associated with an outcome, help predict it, causally affect it, or operate through a particular mechanism. These are different scientific claims, and establishing one does not automatically establish the others.

By Yona Ole Lobulu ·

Concept6 min readFoundationalD2.3

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

What kind of relational, causal or mechanistic claim does the evidence actually support?

Definition

Association, prediction, causal effect and mechanism answer different questions. Evidence supporting one constrains what can be said, but it cannot simply be promoted into another kind of claim.

Four scientific claims can sound almost interchangeable while meaning very different things.

Suppose students who use a particular study practice tend to perform better on exams.

We might say that use of the practice is associated with performance, that it predicts later performance, that increasing it causes performance to improve, or that it improves performance through a particular mechanism.

Those are four different claims.

Association and prediction

An association means that variables are related in the observed data.

If students who use a study practice more often also tend to perform better, the two variables are associated.

That relationship may be scientifically important, but it does not reveal the causal structure behind it. The practice might influence performance. Higher-performing students might be more likely to use it. Other factors might contribute to both.

Correlation is one familiar way of describing association, but the broader question is simply whether variables vary together.

Prediction asks something else.

A variable or model has predictive value when information from it helps forecast an outcome.

If knowing how students study improves forecasts of later performance, the study behaviour is predictive.

A predictor can be useful without being a cause.

Prediction can exploit patterns that are stable and informative even when their causal meaning is uncertain.

So: association asks whether variables are related. Prediction asks whether information helps forecast an outcome.

Neither question, by itself, tells us what causes what.

Why association does not settle causation

Suppose stress and poor sleep are associated.

Several causal structures could produce that relationship.

Stress may contribute to poor sleep. Poor sleep may contribute to stress. Both may influence each other. Other causal differences may contribute to both.

This creates problems of directionality and confounding.

Directionality concerns which way a causal relation runs, if one exists.

Confounding occurs when other causal differences are mixed into a comparison so that the observed relationship does not cleanly represent the causal effect of interest.

If two groups differ not only in a candidate cause but also in other relevant ways, their outcome difference can be difficult to interpret causally.

The association alone does not settle which causal structure produced it.

Causation asks about alternative conditions

A causal claim asks whether an outcome would differ under relevant alternative conditions of a candidate cause.

The simplest version is: what difference would changing X make to Y?

That is not the same as asking whether X and Y tend to occur together.

Interventions can provide especially useful causal evidence because they change a candidate factor rather than merely observe its natural variation.

Suppose researchers alter use of a study practice under a design that supports the relevant causal comparison and later observe differences in performance.

That can support the conclusion that changing the practice made a difference to the outcome.

But intervention is not synonymous with proof. What it establishes depends on whether the design supports the causal inference being made.

Causation is therefore not a stronger correlation.

It is a different kind of claim.

Once causal relevance has been established, Causes, Conditions, Triggers and Constraints distinguishes the different roles a causally relevant factor may play within an explanation.

A causal effect is not a mechanism

Now suppose strong evidence shows that changing the study practice improves performance.

We have learned that the intervention makes a causal difference.

We may still not know how.

The effect might involve stronger retrieval, greater engagement with the material, altered attention, or several interacting processes.

Knowing that an intervention works is not the same as knowing how it works.

A causal effect concerns whether changing one factor changes another under the relevant conditions.

A mechanism concerns the process through which that effect is produced.

Evidence about effects and mechanisms can inform one another, but they are not interchangeable.

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.

That process need not be a simple linear chain. It may involve feedback, interacting components or multiple levels of organisation.

Nor is mechanism synonymous with neural mechanism.

The relevant level depends on the phenomenon and the explanatory question. A mechanism may involve behavioural processes, cognitive operations, bodily changes, interpersonal feedback, ecological conditions, neural processes, or relations among several of these.

A related distinction concerns mediators.

A mediator can identify an intermediate variable in a causal pathway, but identifying a mediator is not automatically the same as identifying the mechanism. A mechanism makes a broader claim about how the effect is produced.

Prediction is not explanation

Prediction can be scientifically valuable without revealing causal structure.

A model may accurately forecast relapse, performance or behaviour using previous observations and contextual information.

That does not by itself tell us which predictors are causes, what would happen if we intervened on them, or through what process the outcome is produced.

Predictive success alone does not establish causal or mechanistic explanation.

This does not place prediction below explanation.

Forecasting may be exactly the scientific goal.

The distinction matters only when predictive success is interpreted as though it had already answered a causal question.

Different questions require different evidence

These four categories should not be treated as a prestige ladder running from correlation up through prediction and causation to mechanism.

They differ in inferential commitment, not universal scientific value.

If the question is whether variables are related, association is the relevant target.

If the question is forecasting, prediction is the target.

If the question concerns what difference changing X would make, causal evidence is required.

If the question concerns how an effect is produced, mechanistic evidence becomes relevant.

The quality of the evidence depends on whether it answers the question being asked.

A study does not become weak because it stops at association when association is the question. A predictive model is not deficient because it does not identify a mechanism. A causal estimate can be highly informative even when the mechanism remains uncertain.

Four epistemic achievements

What varies with what? — Association.

What helps forecast what? — Prediction.

What difference would changing X make to Y? — Causation.

How is that effect produced? — Mechanism.

Each category licenses a different kind of conclusion.

Association can inform prediction. Prediction can generate causal hypotheses. Causal evidence can constrain candidate mechanisms. Mechanistic evidence can deepen causal understanding.

But:

Evidence supporting one kind of claim cannot simply be promoted into another.

Do not make a stronger claim than the evidence supports

Scientific language often drifts.

A variable is described as associated with an outcome, then as predicting it, then as causing it, then as explaining it.

Sometimes the evidence supports every transition.

Sometimes the language has moved further than the research.

Correlation, prediction, causation and mechanism answer different scientific questions. Knowing which one has actually been established is part of knowing what the evidence allows us to say.

Sources and research record7 sources, with findings, strengths and limitations as entered

References

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

  1. Hernán, M. A., Robins, J. M. (2020) Causal Inference: What If

    Book · Chapman & Hall/CRC, Boca Raton

    The standard treatment of association versus causation, defining causal effects through outcomes under alternative conditions and setting out confounding, directionality and the role of interventions.

    Read the source

  2. Shmueli, G. (2010) To Explain or to Predict?

    Methodological article · Statistical Science, 25(3) · 289–310

    The canonical statement that predictive modelling and explanatory modelling pursue different goals, and that a model can forecast an outcome accurately without describing the process producing it.

    doi:10.1214/10-STS330

  3. 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

  4. 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. Russo, F., Williamson, J. (2007) Interpreting causality in the health sciences

    Theoretical article · International Studies in the Philosophy of Science, 21(2) · 157–170

    Argues that difference-making evidence and mechanistic evidence are distinct and complementary, so establishing that an effect exists is a different achievement from establishing how it is produced.

    doi:10.1080/02698590701498084

  6. Illari, P. M. (2011) Mechanistic evidence: Disambiguating the Russo–Williamson thesis

    Theoretical article · International Studies in the Philosophy of Science, 25(2) · 139–157

    Clarifies what mechanistic evidence contributes to a causal claim, and why evidence of an effect and evidence of a mechanism answer different questions.

    doi:10.1080/02698595.2011.574856

  7. 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

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. Association means variables are related in the observed data

    Established

    What this does not assert: An observed relationship describes the data; it does not identify the causal structure that produced it.

    1. Hernán, M. A., Robins, J. M. (2020) Causal Inference: What If

      Book · Chapman & Hall/CRC, Boca Raton

      The standard treatment of association versus causation, defining causal effects through outcomes under alternative conditions and setting out confounding, directionality and the role of interventions.

      Read the source

  2. 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

  3. Mechanisms need not be linear chains and need not be neural

    High confidence

    What this does not assert: The relevant level of organisation depends on the phenomenon and the explanatory question being asked.

    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

  4. Identifying a mediator is not automatically identifying the mechanism

    High confidence

    What this does not assert: A mediator marks an intermediate variable in a pathway; a mechanism makes a broader claim about how the effect is produced. D2.5 owns the deeper distinction.

    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

  5. Predictive success alone does not establish causal or mechanistic explanation

    Established

    What this does not assert: A model can forecast accurately without identifying which predictors are causes or how the outcome is produced.

    1. Shmueli, G. (2010) To Explain or to Predict?

      Methodological article · Statistical Science, 25(3) · 289–310

      The canonical statement that predictive modelling and explanatory modelling pursue different goals, and that a model can forecast an outcome accurately without describing the process producing it.

      doi:10.1214/10-STS330

  6. Different research questions legitimately require different forms of evidence

    Canonical inference

    What this does not assert: The categories differ in inferential commitment, not in universal scientific worth; a study is not weak for stopping at association when association is the question.

    1. Shmueli, G. (2010) To Explain or to Predict?

      Methodological article · Statistical Science, 25(3) · 289–310

      The canonical statement that predictive modelling and explanatory modelling pursue different goals, and that a model can forecast an outcome accurately without describing the process producing it.

      doi:10.1214/10-STS330

  7. Association, prediction, causation and mechanism answer four different questions

    Canonical synthesis

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

  8. Evidence supporting one kind of claim cannot simply be promoted into another

    Canonical synthesis

    What this does not assert: Moving from association to prediction, causation or mechanism requires additional justification each time.

  9. Correlation is one way of describing association

    Established

    What this does not assert: The broader question is whether variables vary together, of which correlation is a particular formal expression.

    1. Hernán, M. A., Robins, J. M. (2020) Causal Inference: What If

      Book · Chapman & Hall/CRC, Boca Raton

      The standard treatment of association versus causation, defining causal effects through outcomes under alternative conditions and setting out confounding, directionality and the role of interventions.

      Read the source

  10. Prediction concerns whether information helps forecast an outcome

    Established

    What this does not assert: Predictive modelling pursues forecasting accuracy, which is a different goal from explanatory modelling.

    1. Shmueli, G. (2010) To Explain or to Predict?

      Methodological article · Statistical Science, 25(3) · 289–310

      The canonical statement that predictive modelling and explanatory modelling pursue different goals, and that a model can forecast an outcome accurately without describing the process producing it.

      doi:10.1214/10-STS330

  11. A predictor can be useful without being a cause

    Established

    What this does not assert: Predictive value can rest on stable patterns whose causal meaning remains uncertain.

    1. Shmueli, G. (2010) To Explain or to Predict?

      Methodological article · Statistical Science, 25(3) · 289–310

      The canonical statement that predictive modelling and explanatory modelling pursue different goals, and that a model can forecast an outcome accurately without describing the process producing it.

      doi:10.1214/10-STS330

  12. Association alone does not determine directionality or rule out confounding

    Established

    What this does not assert: Several causal structures can produce the same observed relationship, including reverse causation and shared causes.

    1. Hernán, M. A., Robins, J. M. (2020) Causal Inference: What If

      Book · Chapman & Hall/CRC, Boca Raton

      The standard treatment of association versus causation, defining causal effects through outcomes under alternative conditions and setting out confounding, directionality and the role of interventions.

      Read the source

  13. A causal claim concerns whether an outcome would differ under relevant alternative conditions

    Established

    What this does not assert: Counterfactual formulations differ across frameworks, but the contrast with mere co-variation is standard.

    1. Hernán, M. A., Robins, J. M. (2020) Causal Inference: What If

      Book · Chapman & Hall/CRC, Boca Raton

      The standard treatment of association versus causation, defining causal effects through outcomes under alternative conditions and setting out confounding, directionality and the role of interventions.

      Read the source

  14. Interventions can provide especially useful causal evidence

    Established

    What this does not assert: What an intervention establishes depends on whether the design supports the causal comparison being made; intervention is not synonymous with proof.

    1. Hernán, M. A., Robins, J. M. (2020) Causal Inference: What If

      Book · Chapman & Hall/CRC, Boca Raton

      The standard treatment of association versus causation, defining causal effects through outcomes under alternative conditions and setting out confounding, directionality and the role of interventions.

      Read the source

  15. Causation is not simply a stronger correlation

    Established

    What this does not assert: It is a different kind of claim, not a higher grade of the same claim.

    1. Hernán, M. A., Robins, J. M. (2020) Causal Inference: What If

      Book · Chapman & Hall/CRC, Boca Raton

      The standard treatment of association versus causation, defining causal effects through outcomes under alternative conditions and setting out confounding, directionality and the role of interventions.

      Read the source

  16. A causal effect and a causal mechanism are different achievements

    Established

    What this does not assert: Difference-making evidence and mechanistic evidence are related but distinct forms of support.

    1. Russo, F., Williamson, J. (2007) Interpreting causality in the health sciences

      Theoretical article · International Studies in the Philosophy of Science, 21(2) · 157–170

      Argues that difference-making evidence and mechanistic evidence are distinct and complementary, so establishing that an effect exists is a different achievement from establishing how it is produced.

      doi:10.1080/02698590701498084

    2. Illari, P. M. (2011) Mechanistic evidence: Disambiguating the Russo–Williamson thesis

      Theoretical article · International Studies in the Philosophy of Science, 25(2) · 139–157

      Clarifies what mechanistic evidence contributes to a causal claim, and why evidence of an effect and evidence of a mechanism answer different questions.

      doi:10.1080/02698595.2011.574856

Sources

  1. Hernán, M. A., Robins, J. M. (2020) Causal Inference: What If

    Book · Chapman & Hall/CRC, Boca Raton

    The standard treatment of association versus causation, defining causal effects through outcomes under alternative conditions and setting out confounding, directionality and the role of interventions.

    Read the source

  2. Shmueli, G. (2010) To Explain or to Predict?

    Methodological article · Statistical Science, 25(3) · 289–310

    The canonical statement that predictive modelling and explanatory modelling pursue different goals, and that a model can forecast an outcome accurately without describing the process producing it.

    doi:10.1214/10-STS330

  3. 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

  4. 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. Russo, F., Williamson, J. (2007) Interpreting causality in the health sciences

    Theoretical article · International Studies in the Philosophy of Science, 21(2) · 157–170

    Argues that difference-making evidence and mechanistic evidence are distinct and complementary, so establishing that an effect exists is a different achievement from establishing how it is produced.

    doi:10.1080/02698590701498084

  6. Illari, P. M. (2011) Mechanistic evidence: Disambiguating the Russo–Williamson thesis

    Theoretical article · International Studies in the Philosophy of Science, 25(2) · 139–157

    Clarifies what mechanistic evidence contributes to a causal claim, and why evidence of an effect and evidence of a mechanism answer different questions.

    doi:10.1080/02698595.2011.574856

  7. 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

What this opens up

What becomes readable once you have this.

Where to go from here

Next published piece

Observation, Explanation and Intervention

Observing a pattern, explaining why it occurs and demonstrating what happens when something is deliberately changed are different scientific achievements. Evidence for one does not automatically establish the others.

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

Back to the Library →