Statistics Reference
Correlation vs Causation
A correlation coefficient is a measurement; a causal claim is a theory about how the world works. Confusing the two is the most consequential mistake in applied statistics — it sells useless supplements, misdirects policy, and derails product decisions. This guide lays out the four explanations behind any correlation, the classic traps with real examples, and what evidence actually supports “X causes Y.”
Any Correlation Has Four Possible Explanations
| Explanation | Pattern | Example |
|---|---|---|
| Causation | X → Y | Smoking and lung cancer |
| Reverse causation | Y → X | “Hospitals correlate with deaths” — the sick go to hospitals, not the reverse |
| Confounding | Z → X and Z → Y | Ice cream sales and drownings, both driven by summer heat |
| Coincidence | chance | Any two rising time series correlate; search enough pairs and absurd matches appear |
The data alone look identical in all four cases — a correlation calculator reports the same r either way. Telling the stories apart requires information the coefficient does not contain: time order, experimental control, and knowledge of plausible third variables.
A Worked Correlation, Read Two Ways
Five students report weekly study hours and exam scores: (2, 65), (4, 70), (6, 75), (8, 85), (10, 90). The correlation works out to r = 0.9912 — nearly perfect. (Enter the pairs into the correlation calculator to reproduce it.)
The measurement
“In this sample, study hours and scores move together almost perfectly linearly.” This claim is fully justified by the data.
The overreach
“Studying two more hours will raise a student's score by ~6 points.” Not established: conscientious students may both study more and attend more classes, sleep better, or start stronger — all confounders.
The causal version predicts what happens under intervention — forcing a change in study hours — which this observational snapshot never tested. An experiment that randomly assigns extra study sessions would test exactly that.
The Classic Traps, With Diagnoses
- Shoe size predicts reading ability (in children). Confounder: age. Within a single age group, the correlation collapses.
- Coffee drinkers had higher lung cancer rates in early studies. Confounder: smoking, which historically accompanied coffee. Adjusting for smoking removed most of the association.
- Players who complain more lose more — likely reverse causation: losing causes complaining.
- Spurious time-series pairs: two unrelated quantities that both trend upward over years (population, prices, adoption curves) correlate strongly by construction. Detrend first, or the correlation measures the calendar, not a relationship.
- Selection effects: among hospitalized patients, two diseases can correlate negatively even if independent in the population, because having either one is what got patients admitted (Berkson's paradox).
What Actually Supports a Causal Claim
- Randomized experiments. Random assignment severs every confounder link at once — the reason A/B tests and clinical trials are decisive where observational data argues. The remaining role of chance is quantified by the p-value and the effect estimated with a confidence interval.
- Time order. Causes precede effects; longitudinal data can rule out reverse causation where snapshots cannot.
- Dose-response. More exposure → more effect, visible as a monotone trend in a regression, strengthens (but does not prove) a causal reading.
- Mechanism. A known pathway from X to Y makes causation plausible; its absence demands stronger data.
- Adjustment with humility. Controlling for measured confounders helps only against the confounders you measured — the unmeasured ones remain, which is why observational conclusions stay provisional.
Try the Correlation Calculator
Measure the strength and direction of a linear relationship — then interpret it with the four explanations in mind.
Frequently Asked Questions
What does 'correlation does not imply causation' actually mean?
A correlation says two variables move together; causation says changing one would change the other. The slogan warns that moving-together has at least four explanations: X causes Y, Y causes X, a third variable drives both, or the pattern is a coincidence of the sample. The correlation coefficient alone cannot distinguish between them - that requires knowledge of how the data was generated.
What is a confounder, with a concrete example?
A confounder is a third variable that influences both correlated variables, creating an association without any direct link. Ice cream sales correlate with drowning deaths - not because dessert is dangerous, but because hot weather increases both swimming and ice cream consumption. Controlling for temperature makes the association largely disappear.
Can a strong correlation (like r = 0.9) still be non-causal?
Yes - strength says nothing about mechanism. Confounded relationships can produce correlations arbitrarily close to 1: children's shoe sizes and reading scores correlate strongly because age drives both. Conversely, real causal effects can show weak correlations when the effect is small, noisy, or diluted across contexts. Strength measures signal-to-noise, not causality.
How can causation be established?
The gold standard is a randomized controlled experiment: randomly assigning the treatment breaks every link between the treatment and lurking confounders, so a resulting difference can only come from the treatment (or chance, which the p-value quantifies). When experiments are impossible, causal inference leans on natural experiments, longitudinal designs, dose-response patterns, mechanism knowledge, and statistical adjustment - each with assumptions that must be argued, not just computed.
What is reverse causation?
Mistaking the direction of a real causal link. Depression correlates with low exercise: does inactivity worsen mood, or does low mood reduce activity? Both are plausible, and a cross-sectional correlation cannot tell them apart. Time ordering helps - a cause must precede its effect - which is why longitudinal data beats snapshots for causal questions.
Does zero correlation prove there is no causal relationship?
No. Pearson's r measures only linear association, so a U-shaped causal relationship (like fertilizer dose and crop yield, which rises then falls) can produce r near 0. Effects can also cancel across subgroups, or be masked by a suppressor variable. Always plot the data before concluding 'no relationship' - the scatter plot catches what the coefficient misses.