Correlation, Causation, and How Public Debate Confuses Them

Everyone has heard that correlation is not causation. The more useful knowledge is why a correlation arises when causation is absent — because each mechanism suggests its own test.

#Three ways a correlation lies

• Confounding. A third factor drives both variables. Ice-cream sales and drowning deaths rise together because summer causes both. In observational health data, socioeconomic status confounds almost everything.
• Reverse causality. The arrow points the other way. Do people become unhealthy because they stop exercising — or stop exercising because they became unhealthy? Cross-sectional snapshots cannot say.
• Selection. The comparison groups formed themselves. People who choose a programme differ from those who do not, in ways usually unmeasured — the core problem randomisation exists to solve, as we explain in our RCT primer.

#What supports a causal claim

The strongest support is design: randomised assignment, or natural experiments in which exposure varied for reasons plausibly unrelated to the outcome. Where only observational data exist, researchers triangulate — checking dose-response patterns, temporal order, robustness to adjustment, and consistency across settings. The epidemiologist Austin Bradford Hill’s famous 1965 “viewpoints” (strength, consistency, temporality, plausibility, and others) remain a useful checklist, so long as they are treated as considerations for judgement rather than a scoring rubric.

#Reading claims in the wild

Practical heuristics: be most sceptical when the claimed cause is something people select into (diets, apps, neighbourhoods); check whether the study design could distinguish direction; and notice whether effect sizes survive adjustment or shrink toward zero as controls improve — a classic signature of confounding. And symmetrically: correlation plus strong design is how nearly everything we reliably know about causes was learned. The slogan should raise the bar for evidence, not lower it to nihilism.