Correlation Does Not Imply Causation

Correlation does not imply causation is the principle that a statistical association (correlation) between two variables does not, on its own, establish that one variable causes changes in the other. Possible explanations for an observed correlation besides direct causation include reverse causation (Y causes X, not the reverse), confounding (a third variable causes both X and Y), and coincidence or spurious correlation. Establishing causation typically requires additional evidence, such as a randomized controlled experiment, a plausible causal mechanism, or specialized causal inference methods, e.g. instrumental variables, regression discontinuity.

• Confounding is often the most subtle of the three alternative explanations -- a hidden third variable driving both X and Y can produce a strong, entirely real correlation between X and Y even though neither directly causes the other.
• Randomization is what gives randomized controlled experiments their special claim to causal evidence: by breaking any systematic link between the treatment assignment and pre-existing confounders, a properly randomized experiment rules out confounding as an explanation for an observed association.
• When randomized experiments aren't feasible or ethical, observational causal inference methods, instrumental variables, regression discontinuity designs, difference-in-differences, are specifically built to approximate the logic of randomization and support causal claims from non-experimental data, though each relies on its own set of assumptions that must be justified, not merely assumed.