Is causality the same as correlation?

No. Correlation describes association; causality concerns how outcomes depend on causes under a specified interpretation.

Can a predictor be noncausal?

Yes, including indicators or consequences of the outcome process.

Is there one universal statistical test for causality?

No. The appropriate evidence depends on the question, design and assumptions.

AI Glossary / Causal inference and experimentation / Intermediate

Causality

Definition

Causality concerns relationships in which changes in one part of a system can produce changes in another. In machine learning and statistics, causal frameworks make those relationships explicit so that questions about interventions and counterfactuals can be distinguished from questions about prediction or association.

What it means

A model may predict an outcome accurately without explaining what would happen if an input were deliberately changed. For example, a status indicator can reveal a process condition without controlling that condition. Causality asks about the underlying relationship that makes an action capable of changing an outcome.
There are several formal notions. Structural causal models describe mechanisms and interventions; potential-outcome frameworks compare outcomes under alternatives. Granger causality concerns additional predictive information in time series and does not, by itself, establish the same intervention meaning.
The term therefore needs context. Clear writing should state which causal question and framework are being used rather than assuming every use of “causal” makes the same claim.

How it works

Begin with a system, variables and a question about change. A structural causal model specifies how each variable is generated from its direct causes and background influences. An intervention replaces a specified mechanism, while a counterfactual question considers an alternative action for a particular observed situation under additional assumptions.
Other frameworks formalize the comparison differently, but all require clarity about alternatives and evidence. Experiments can help separate effects from confounding. Observational analysis relies on assumptions linking data to interventions. Temporal order is relevant, yet occurring earlier is not sufficient: a common cause can make one variable precede and predict another without the first causing the second.

Three practical examples

A product team distinguishes a feature that predicts customer retention from a feature change that would actually affect retention.
An engineer separates a machine's warning signal from the mechanism generating the underlying problem.
A time-series analyst reports predictive Granger relationships without presenting them automatically as effects of manipulating one series.

Think of it this way

Think of the difference between reading a thermometer and adjusting a thermostat. Both are related to temperature, but observing a reading and changing a control are different operations. The analogy is deliberately limited: a real thermostat system has feedback, delays and other causes, so its behavior requires a more detailed model than one simple arrow.

Worked example

Consider a toy structural model: outdoor temperature U determines both cold-drink sales X and cooling-system demand Y. Let X=2U and Y=3U, ignoring noise for illustration. Observationally, Y=1.5X, so X predicts Y perfectly.
Now imagine an intervention that changes recorded drink sales through an unrelated promotion while leaving U and the cooling system unchanged. Under the stated model, changing X does not change Y because Y's mechanism depends on U, not X.
The perfect predictive relationship therefore does not imply that increasing drink sales will increase cooling demand by the fitted coefficient. This is a hypothetical model, and actual promotions could have additional effects in a real venue. The example isolates the logical distinction between association generated by a common cause and an intervention effect.

Illustrative case study

Fictional case: Marigold Analytics develops a model predicting equipment downtime from maintenance tickets. The ticket count is highly predictive, and a manager proposes reducing tickets to improve uptime.
The analysts explain that tickets may be responses to underlying faults rather than causes of them. They map the reporting and fault processes, then define interventions that could change equipment condition instead of merely changing the recorded indicator. The team also separates forecasting use from evaluation of maintenance actions. This example illustrates causal reasoning in problem definition; it does not claim that every ticket is purely an effect or that reporting practices never influence outcomes.

Common misconceptions and mistakes

Correlation does not establish causality, but an absence of simple correlation does not rule out every causal relationship either. Effects may vary, cancel or depend on context. Earlier timing alone is insufficient evidence.
Granger causality is a predictive notion with its own assumptions, not automatic proof of a manipulable mechanism. Causal language modeling uses causal in yet another sense: restricting access to future tokens. Do not treat a causal diagram as self-validating or assume that a plausible mechanism eliminates the need for careful empirical evaluation.

Business use

Understanding causality helps businesses choose actions rather than optimize indicators that merely accompany desired outcomes. It is relevant whenever a model is used to recommend what to change, explain a result or evaluate a program.
Separate forecasting questions from intervention questions in project briefs. Record which variables are controllable, what changing them means and which other mechanisms might also change. Ask whether the available evidence supports the proposed action in the intended population and setting. A useful causal analysis may reveal that more measurement or an experiment is needed, even when a predictive model already performs well.

Check your understanding

Two variables are perfectly correlated because both are determined by a third variable. Does changing one necessarily change the other?

Answer: No. Their association may reflect the shared cause. The effect of an intervention depends on the system's causal mechanisms, not only on the observed correlation.

Related terms

Learn next

Learn association, intervention and counterfactual questions as distinct levels of inquiry. Rewrite one predictive claim as a concrete intervention question and identify the additional assumptions or evidence it requires. Then compare structural and potential-outcome descriptions of the same simple example.

Sources

Aalto Dictionary of Machine Learning, causality entry: https://aaltodictionaryofml.github.io/ADictML_Main.pdf Judea Pearl, Causal inference in statistics: An overview: https://ftp.cs.ucla.edu/pub/stat_ser/r350.pdf Supplied dictionary entry read and primary-paper excerpts consulted. Examples are original; full methodological and editorial review remain pending.

Previous: Causal positivity Next: Causation