Tukuk-OSTwo numbers that rise together prove nothing, because the third variable is usually holding the umbrella. Ice cream sales and drowning deaths track each other perfectly across the year — summer causes both, and neither causes the other. This “confounder” is the commonest culprit: a hidden factor moves both curves, and the pair looks meaningful forever. A second trap is direction: do people with strong social ties stay healthier because ties protect them, or do healthy people find the energy to keep ties? Both stories fit the same correlation, and the answers lead to different interventions. The third is selection: happy people post more photos, but the posting did not create the happiness — the visible sample was never random.
The absurd ends make the point stick: an annual dataset (the famous Spurious Correlations archive) pairs Nicolas Cage film counts with pool-drowning deaths at r ≈ 0.66, and margarine consumption with divorce rates in the same state. No mechanism, pure coincidence — and with thousands of variables and a year of wiggles each, dozens of perfect-looking matches are not just possible, they are guaranteed. The lesson generalises: a chart’s persuasiveness is unrelated to its truth; the question that separates them is “what is the mechanism, and how was the comparison made?”
Causal evidence comes from designs, not correlations: a randomised trial (assign the exposure by lottery, so the only systematic difference between groups is the thing you changed), natural experiments (policy boundaries, staggered rollouts), or a mechanism tight enough to rule coincidence out in physics and chemistry. Where those are absent — most headline science — the honest sentence is “associated with”, and whenever you catch yourself writing or believing “leads to / increases / causes” from a pair of moving lines, you have the right reflex: what moved them both?