Correlation vs. causation
Open the Correlation or cause? room
This article opens a matching guided lab. Follow the on-screen prompts, make a prediction, and connect the model result back to this topic.
01 // The difference
Correlation means two variables change together in a pattern — as one goes up, the other tends to go up (or down) too. Causation means one variable's change directly produces the other's change. Every causal relationship shows up as a correlation, but plenty of correlations exist with no causal link at all — the phrase 'correlation does not imply causation' exists exactly because this mistake is so easy to make.
02 // A classic example
Correlation (observed)
Ice cream sales and drowning deaths- Both rise in the summer months
- Both fall in the winter months
- They move together — a real correlation
Actual cause
Hot weather (a confounding variable)- Heat makes people buy more ice cream
- Heat also drives more people to swim
- Ice cream doesn't cause drowning at all
03 // How experiments get closer to causation
- Simply observing two things happening together (an observational study) can only ever establish correlation, not causation.
- A controlled experiment — where one variable is deliberately changed while everything else is held constant — is what lets scientists claim causation, because it rules out other explanations.
- Even strong, repeated correlations are treated cautiously in science until a plausible causal mechanism is identified and tested directly.
04 // Three explanations for a correlation
When two variables change together, one may influence the other, a third variable may influence both, or the association may be coincidental or unstable. Ice-cream sales and swimming incidents may both rise in hot weather; temperature is a plausible third factor. The correlation alone does not show that buying ice cream causes swimming incidents or that one variable directly causes the other.
A scatterplot can reveal direction, form, and strength of an association, but it does not identify the mechanism by itself. Check axes, units, range, sample size, and outliers. A strong pattern can still be confounded, while a real causal effect can be hidden by noisy measurements or a narrow sample. Use causal language only when the design and evidence support it.
05 // What experiments add
A controlled experiment deliberately changes a factor and measures a response while holding other plausible influences steady. Random assignment can help balance unknown factors across groups. If the outcome differs consistently after the treatment, the design provides stronger evidence for a causal effect than a simple observation. Even then, researchers must consider compliance, measurement quality, sample size, and whether the result generalizes beyond the study conditions.
Some questions cannot be tested experimentally for ethical or practical reasons. Researchers use natural experiments, longitudinal observations, comparison groups, and statistical models to reduce alternative explanations. These methods can provide valuable evidence, but their assumptions should be stated. A careful conclusion distinguishes 'associated with' from 'caused by' and explains why the stronger wording is or is not justified.
06 // Check time order and mechanism
For a cause to produce an effect, the cause must occur before the effect. A cross-sectional snapshot may show an association without revealing which came first. Repeated measurements over time can help clarify sequence, but they still may not rule out a third factor. A plausible mechanism strengthens an explanation when it is supported by evidence, rather than added after the fact simply to make a correlation sound causal.
Look for dose-response patterns, repeated findings, and evidence from different methods. No single feature proves causation, but several consistent lines of evidence can strengthen confidence. Also consider reverse causality: perhaps the supposed outcome influences the supposed cause. Writing down competing causal diagrams before analysis can reveal assumptions that a headline or graph leaves unstated.
07 // Practice rewriting causal claims
When evidence is observational, replace 'X makes Y happen' with 'X was associated with Y in this sample' unless the study design supports a causal conclusion. Then add the population, time period, and measurement. This wording is not evasive; it accurately describes what the study measured and leaves room for further testing.
Ask what new evidence would distinguish the competing explanations. A randomized test may be possible for classroom materials but not for harmful exposures. In that case, carefully matched comparisons and multiple independent datasets can help. The strength of the claim should grow only as the evidence rules out plausible alternatives.
08 // A practical checklist for causal claims
When a study reports an association, ask whether the proposed cause happened before the outcome, whether other explanations were measured, and whether the relationship appears across repeated studies or settings. Check how participants or samples entered the comparison and whether the groups differed in ways besides the factor of interest. These questions help separate a plausible mechanism from a coincidence.
A well-controlled experiment can strengthen a causal claim, but it still has a defined population and range of conditions. Observational studies can also support causal reasoning when they use careful designs and converging evidence, though their assumptions must be explained. Use wording that matches the evidence: 'was associated with' is safer than 'caused' when alternative explanations remain.
09 // Check time order and alternative pathways
A proposed cause must occur before its effect, but time order alone is insufficient. A third factor may influence both variables, or the measured outcome may affect the predictor. Sketching a simple cause-and-effect diagram can make these possibilities visible and help identify what additional measurement or comparison would be useful.
Repeated observations, natural experiments, randomized studies, and mechanistic evidence each contribute different strengths. No single design answers every causal question. Report the association that was measured and describe the evidence supporting a causal interpretation, including assumptions and remaining alternatives, so readers can judge the strength of the claim.
10 // Questions people ask
What does 'correlation does not imply causation' mean?
It means that just because two variables change together in a pattern doesn't prove that one is causing the other to change. They might both be driven by a separate, hidden factor, or the relationship could even be coincidental.
How can you tell if a relationship is actually causal?
The strongest evidence comes from a controlled experiment, where researchers deliberately change one variable while holding everything else constant and observe the effect. If changing the variable reliably produces the same result, that supports causation far more than simply observing two things happening together.
What is a confounding variable's role in a false correlation?
A confounding variable is a hidden third factor that independently affects both of the variables that appear correlated, creating the illusion that one causes the other when really both are just responding to the same underlying cause.