"Correlation is not causation" is one of those phrases everyone can recite and almost everyone still trips over in practice. The pull is strong: two things move together, our pattern-hungry minds supply a story about why one drives the other, and a decision gets made on the strength of that story. But a correlation is only a statement that two quantities tend to move together. It says nothing, on its own, about why — and mistaking it for a causal claim is behind a startling amount of bad forecasting, wasted spend, and irreproducible science. Getting this right is a genuine thinking skill, and it is learnable.
What each claim actually asserts
Start by pinning the two ideas apart. Correlation is descriptive: as X goes up, Y tends to go up (or down). It is symmetric and makes no promises. Causation is a claim about intervention: if you reach in and change X, Y will change as a result. That difference — passively observing versus actively intervening — is the whole game. The reason it matters is that decisions are interventions. "Should we do more X to get more Y?" is a causal question, and a correlation cannot answer it, no matter how strong.
The intervention test
To tell whether a relationship is causal, ask: if I intervened and changed the cause, holding nothing else fixed by hand, would the effect change? Causation licenses that prediction; correlation alone does not.
The confounder: the usual culprit
Most spurious correlations come from a confounder — a third variable that influences both of the things you are looking at. The classic illustration: ice-cream sales and drowning deaths rise and fall together across the year. Ice cream does not cause drowning; hot weather causes both — more ice cream eaten and more people swimming. The correlation between ice cream and drowning is real and strong, and completely non-causal, because a lurking common cause drives both.
Once you know to look for confounders, you see candidates everywhere: the "expensive hospitals have higher death rates" story (sicker patients go to specialist hospitals), the "students who own more books score higher" story (family circumstances drive both). The discipline is to ask, before believing any correlation is causal, "what else could be driving both of these at once?"
Simpson's paradox: when the trend flips
There is a stranger trap still. Simpson's paradox is when a trend that appears in aggregated data reverses once you split the data into meaningful groups. A treatment can look worse overall yet be better for every subgroup, because the groups differ in size and baseline risk in ways that distort the pooled numbers. The unsettling lesson is that an aggregate correlation is not just incomplete — it can point in the opposite direction from the truth. Which grouping is the right one to analyse is itself a causal question, not a statistical one you can resolve by staring harder at the same table.
More data does not fix the wrong question
It is tempting to think a bigger dataset settles causal questions. It does not. A confounded correlation computed on a billion rows is still confounded — just measured more precisely. Escaping the trap requires a different kind of evidence, not a larger quantity of the same kind.
How causation actually gets established
If observation alone cannot prove cause, what can? Two routes.
The gold standard is the randomised controlled experiment. By randomly assigning who receives an intervention, you break the link between the treatment and any confounder — on average the groups differ only in the treatment — so a difference in outcome can be attributed to it. This is why A/B tests, when run properly, are so valued: randomisation is what turns a correlation into a causal estimate.
When you cannot experiment — you cannot randomly assign people to smoke, or economies to a policy — you turn to causal inference: making your assumptions about what causes what explicit (often as a diagram of causal relationships), then reasoning carefully about which correlations those assumptions would and would not explain. It is more demanding than running a regression and reading off a coefficient, because it forces you to state a model of the world rather than letting the data pretend to speak for itself.
| Correlation | Causation | |
|---|---|---|
| Claim | X and Y move together | Changing X changes Y |
| Evidence needed | Observational data | Experiment, or explicit causal reasoning |
| Answers "should we...?" | No | Yes |
| Broken by | Confounders, Simpson's paradox | (Randomisation removes confounders) |
Practical takeaway
Train yourself to add a beat of friction between "these move together" and "so one causes the other." When you see a correlation that implies an action, run three quick checks: could a confounder drive both, could the trend reverse inside subgroups, and is there any experiment or explicit causal argument behind the claim or just observed co-movement? In machine learning the same discipline applies — a model can exploit a correlation to predict beautifully while learning nothing causal, which is exactly why a great predictor can give terrible advice about what to change. The phrase is a cliché because it is true; the skill is remembering it precisely when a tidy story is tempting you not to.
Sources & Further Reading
- 01Simpson's Paradox — Stanford Encyclopedia of PhilosophyA rigorous treatment of how aggregate trends can reverse across groups.
- 02Judea Pearl — causal inference (home page) — Judea Pearl, UCLAFoundational work on causal reasoning; see also his book The Book of Why.
- 03Spurious Correlations — Tyler VigenA memorable gallery of strong correlations between plainly unrelated things.
Editorial note — A conceptual explainer of the correlation/causation distinction, confounding, Simpson's paradox, and how causation is established. The ice-cream and hospital examples are standard teaching illustrations, not measured datasets; no statistics are quoted.
