Definition
Causation is the relationship in which a change in one variable produces, directly or through an intervening mechanism, a change in another. The causal arrow is asymmetric — a cause influences an effect, but the effect does not run back the other way. This asymmetry, together with a plausible mechanism, is what distinguishes causation from mere statistical association.
In statistical practice, causation is the goal that most analyses aspire to even when they can only deliver weaker evidence. Correlations, regressions, and contingency tables can all describe how variables move together, but they cannot by themselves prove that one drives the other. Establishing causation requires either a controlled experiment that randomly assigns the suspected cause, or carefully designed observational study that rules out confounders, reverse causation, and chance.
Why it matters
How it works
A claim of causation rests on three components. The first is statistical association: the two variables must in fact move together more than chance would predict. The second is temporal ordering: the cause must precede the effect. The third — and the most demanding — is the elimination of alternative explanations, including reverse causation (the effect actually drives the cause), confounding (a third variable drives both), and chance (the apparent pattern is sampling noise).
Experimental design is the gold standard because random assignment to treatment and control conditions ensures that, on average, the only systematic difference between the groups is the treatment itself. Any subsequent difference in outcomes can therefore be attributed to the treatment with quantified confidence. When experiments are not possible — for ethical, practical, or historical reasons — observational designs lean on natural variation, instrumental variables, regression discontinuities, and difference-in-differences strategies to approximate the experimental ideal. None of these match a true randomised trial, but each can support credible causal conclusions when the underlying assumptions are defensible.