Can correlation be evidence relevant to causation?

Yes, when interpreted within a defensible design and broader evidence.

Does randomization guarantee a flawless causal study?

No; adherence, missing outcomes, interference, and measurement can still matter.

Does a perfect correlation prove causation?

No. The constructed common-cause example has correlation one without the proposed direct causal relationship.

AI Glossary / Statistics and probability / Intermediate

Correlation does not imply causation

Definition

Correlation does not imply causation means that an observed statistical association, by itself, does not establish that changing one variable would change another. The association may arise from a causal relationship, a common cause, reverse causation, selection, chance, or other mechanisms. The principle distinguishes describing how variables vary together from identifying what would happen under an intervention.

What it means

If employees who use an AI tool produce more work, the tool may help, but users may also have greater experience or receive different assignments. The outcome could even influence tool use: high workloads might prompt people to adopt the tool. A correlation does not distinguish these explanations on its own.
The phrase does not mean correlations are useless or that causal conclusions are impossible. Associations can motivate hypotheses, support prediction, and contribute to causal evidence when combined with a suitable design and explicit assumptions. The missing step is justification for the intervention claim.

How it works

State the proposed cause, outcome, population, intervention, and time frame. Asking whether two variables are associated is different from asking what would happen if a specific action changed one of them. Some variables are markers rather than directly actionable causes.
List plausible explanations for the observed pattern. A common cause can influence both variables. Reverse causation can change the proposed direction. Selecting observations based on a shared consequence can create an association, while aggregation and common time trends can obscure the underlying relationships.
Use a study design appropriate to the causal question. A well-conducted randomized experiment can make treatment assignment independent of prior causes in expectation. Natural experiments, instrumental variables, longitudinal designs, and other observational approaches can also contribute, but each requires assumptions that must be defended.
Choose adjustment variables based on the causal structure rather than indiscriminately adding everything measured. Confounders, mediators, and colliders play different roles. Temporal order is necessary for many causal interpretations but is not sufficient: a variable occurring earlier can still be associated through another cause.
Assess alternative explanations, measurement quality, missingness, and sensitivity to assumptions. Explain the effect estimate and uncertainty within the design's scope. A small p-value, high correlation, or accurate prediction does not replace this reasoning.

Three practical examples

A business tests whether a new workflow improves outcomes instead of assuming an association among voluntary adopters is causal.
A product team checks whether an observed link between engagement and renewal reflects customer interest, product effects, or both.
An AI team distinguishes a model feature that predicts failure from an intervention that would prevent failure.

Think of it this way

Think of two clocks showing the same time. Their agreement does not show that moving one clock's hands will move the other's. Both may be responding to a shared time source. The question about coordinated observations is different from the question about changing the system.

Worked example

Consider a deliberately constructed system with an underlying workload level Z taking values one, two, three, and four. Let the number of active workstations be X = 10Z, and let the number of completed batches be Y = 5Z + 2. The paired observations are (10,7), (20,12), (30,17), and (40,22).
These points lie exactly on Y = 0.5X + 2, so Pearson correlation is one. Yet the specified structural model has Z determining both X and Y; there is no direct X term in the equation for Y.
At Z = 2, the system gives X = 20 and Y = 12. If an intervention changes X to thirty while holding Z and the Y-generating mechanism fixed, the model still gives Y = 12. Substituting thirty into the observational line would instead predict seventeen, answering a different question.
This is an artificial example showing that perfect correlation is compatible with no causal effect of X on Y under a stated mechanism. It is not a claim about real workstations or productivity. In a real system, determining the mechanism would require substantive evidence, not merely choosing equations that produce a preferred conclusion.

Illustrative case study

Fictional case: Juniper Services notices that customers who attend optional onboarding sessions renew more often. Managers initially propose treating the entire renewal difference as the effect of attending. Analysts point out that motivated customers may be more likely both to attend and to renew.
The team defines the intervention it can actually implement, such as offering additional onboarding support, and plans an evaluation suited to that question. It keeps attendance-based associations as descriptive evidence while separating them from the effect of the offer. The resulting decision considers participation, costs, and uncertainty rather than assuming that forcing attendance would reproduce the observed difference.

Common misconceptions and mistakes

Do not use the phrase to dismiss all evidence from observational research. Conversely, do not assume that a controlled regression, large sample, or temporal sequence proves causation. Predictive accuracy is not an intervention effect. Absence of a simple correlation also does not rule out causal effects: nonlinear relationships, opposing pathways, or variation across groups can conceal an overall association.

Business use

This principle protects decisions that attempt to change outcomes. Before acting on a dashboard relationship, identify the action and explain why changing it should affect the outcome. Use experiments when suitable and feasible, or make observational assumptions explicit. Keep predictive and causal uses separate in reports: a warning indicator can be valuable even when changing the indicator would not change the underlying problem.

Check your understanding

What produces the perfect association in the worked example?

Answer: Both variables depend on Z.

Why does intervening on X not change Y in that model?

Answer: Y's structural equation depends on Z, not X.

What additional work is needed in an application?

Answer: A credible causal design and justified assumptions about the relevant mechanisms.

Related terms

Confounding
Causal inference
Structural causal model
reverse causation
collider bias
intervention
randomized experiment
predictive association

Learn next

Draw several causal diagrams that could produce the same observed association. Then identify what experiment, additional information, or assumptions would distinguish their predictions about a specific intervention.

Sources

NIST/SEMATECH e-Handbook, Scatter Plot, causality and association discussion: https://www.itl.nist.gov/div898/handbook/eda/section3/eda33q.htm Miguel A. Hernan and James M. Robins, Causal Inference: What If, author book page: https://miguelhernan.org/whatifbook

Previous: Correlation coefficient Next: Correlation function