Correlation and Causation

In brief

That two things move together is a fact about the data. That one of them causes the other is a claim about the world, and it is a much stronger claim. A correlation is easy to measure: gather the numbers, compute how tightly they track. Causation is hard to establish, because the same pattern can be produced in at least three different ways. The first is genuine cause, X producing Y. The second is a hidden third variable, some Z that drives both X and Y so that they rise and fall together while neither touches the other. The third is reverse causation, Y producing X while you assumed the arrow ran the other way. Nothing in the correlation itself tells you which of the three you are looking at. The randomized experiment exists precisely to rule out the second and third, and that is why it, and almost nothing short of it, licenses the word "cause."

The full treatment

The problem it answers

Human beings are relentless causal reasoners. We see two events keep company and the mind supplies a story linking them, usually before we have noticed we did it. This machinery is useful; it is also cheap, and it fires on patterns that mean nothing. The discipline of separating correlation from causation is the correction we bolt onto that fast, story-hungry intuition. It answers a single question that sits beneath every empirical claim in medicine, economics, and psychology: when we observe that people who do X tend to have outcome Y, have we learned that X brings about Y, or only that X and Y keep showing up together?

The third-variable problem

The most common trap is the lurking third variable, the confounder. Ice cream sales and drownings rise together across the year, but banning ice cream would not save a single swimmer; summer heat drives both. Countries with more Nobel laureates consume more chocolate, an association popularized in a 2012 note in the New England Journal of Medicine, but the plausible common cause is national wealth, which buys both research universities and confectionery. Tyler Vigen's Spurious Correlations project catalogs hundreds of these, some tracking almost perfectly for a decade while being related by nothing but chance and shared trends. The lesson is not that correlations lie. It is that a correlation is consistent with a confounded world and a causal world equally, and the numbers alone cannot referee between them.

Reverse causation

The arrow can also run backward. Studies find that people who exercise are less depressed, and the natural reading is that exercise lifts mood. But depression saps the energy and motivation to exercise, so at least part of the association is depression suppressing exercise, not the reverse. Firefighters at a blaze correlate with the size of the fire; the fire summons the firefighters. Deciding the direction of the arrow is a separate problem from detecting that the two move together, and observation rarely settles it on its own.

Why the experiment is the tool that licenses cause

Ronald Fisher gave the definitive answer in The Design of Experiments (1935): randomization. If you take a group and assign each member to treatment or control by a coin flip, then on average the two groups are alike in every respect, measured or unmeasured, known or unknown, except the one thing you control. Any confounder you never thought of is, in expectation, balanced across the groups by the randomizing itself. Reverse causation is blocked because you fixed the cause before the outcome could occur. What remains, if the outcome differs, is the treatment. This is why the randomized controlled trial is called the gold standard: not because it is precise, but because it manufactures the one situation in which a difference in outcomes can only be laid at the door of the difference you imposed. See internal-and-external-validity for what the trial buys and what it costs.

The case that changed medicine

For decades, large observational studies found that postmenopausal women taking hormone replacement therapy had markedly less coronary heart disease, and the treatment was prescribed for heart protection to millions. Then the Women's Health Initiative ran the experiment: a randomized trial of estrogen plus progestin in 16,608 women. In 2002 it was halted early because the treatment group had more heart disease, strokes, and blood clots, not fewer. The observational finding had been confounded: women who chose hormone therapy were, on the whole, wealthier, healthier, and more attentive to their bodies, and it was those traits, not the hormones, that had been protecting their hearts. A correlation that looked overwhelming, replicated across many studies, was overturned the moment the confounders were randomized away.

Lineage

The maxim that correlation is not causation is old; the Scottish philosopher David Hume (1711 to 1776) had already argued that we never observe causation directly, only the constant conjunction of events. What made the distinction operational was twentieth-century statistics. Fisher formalized randomization in the 1920s and 1930s at Rothamsted, an agricultural station, where he was assigning fertilizers to plots. Austin Bradford Hill, faced with observational data linking smoking to lung cancer that could not ethically be tested by experiment, proposed in 1965 nine viewpoints, strength, consistency, a dose-response gradient, temporality, and others, for judging when an association is strong enough to be read as causal without a trial. In our own time Judea Pearl has built a formal calculus of causation, arguing in The Book of Why (2018) that with the right assumptions drawn explicitly as a diagram, causal claims can sometimes be recovered from observational data, reclaiming ground that a strict reading had ceded.

The strongest case for it

The distinction is one of the highest-yield ideas in all of reasoning because the error it guards against is both universal and expensive. Mistaking correlation for causation has sold worthless remedies, cancelled useful ones, and justified policies that did nothing or harm. It is the engine behind confirmation bias in its evidential form: we find a correlation that fits our prior, read it as proof, and stop looking. Holding the line, refusing to say "causes" until the alternatives are ruled out, is not pedantry; it is the difference between the hormone-therapy debacle and its correction. And unlike many epistemic warnings, this one comes with a constructive remedy: the experiment. It tells you not only when you do not know but exactly what to do to find out.

The strongest case against it

The slogan can harden into a reflex that is itself an error. Two lines of attack matter.

First, from within science, the reflex "correlation is not causation" is sometimes used to dismiss any observational finding, which would forbid conclusions we are right to draw. We cannot run a randomized trial that assigns people to smoke, yet the causal link between smoking and lung cancer is beyond serious doubt. Fisher himself, one of the great statisticians of the century, resisted that conclusion and argued the correlation might reflect a genetic predisposition that caused both the smoking and the cancer, a textbook third-variable objection that turned out to be wrong. Bradford Hill's viewpoints, and Pearl's later formalism, exist because a blanket refusal to infer cause from observation is not scientific caution but paralysis; often the experiment is impossible and the observational evidence is nonetheless decisive.

Second, the experiment is not a clean triumph either. Randomized trials are expensive, slow, often unethical, and run on narrow, willing samples that may not resemble the world, the external-validity problem. Worse, the statistical machinery meant to read their results is fragile. The 2016 statement of the American Statistical Association, drafted by Ronald Wasserstein and Nicole Lazar, warned that a small p-value does not measure the size or the importance of an effect and was never a license to declare a cause. The wider reckoning of the-replication-crisis showed that a great many published "significant" experimental effects do not reproduce. Establishing causation, in short, is harder than the tidy story of randomization suggests, and the tool that licenses the word "cause" can still be misused to bless a phantom.

Where it stands now

The distinction is settled doctrine and taught in every introductory course in statistics, medicine, and the social sciences. The frontier has moved from the slogan to the machinery. A field of causal inference now takes seriously the project Hill and Pearl began: extracting defensible causal claims from observational data when experiment is out of reach, using instrumental variables, natural experiments, regression discontinuities, and explicit causal diagrams. The randomized trial remains the standard against which those methods measure themselves. What has not changed, and will not, is the first move: seeing that two things travel together, and refusing to say why until you have ruled out the confounder and the backward arrow.

Test yourself

Bring to mind a claim you currently believe on the strength of a pattern: a food that "gives you" energy, a habit that "makes" successful people successful, a policy that "caused" an economy to grow. Ask the two questions the experiment is built to answer. Could a third thing be producing both sides of the pattern? Could the arrow run the other way? If you cannot rule both out, you have a correlation and a story, which is not yet a cause.