Correlation vs Causation

Correlation measures that two variables move together; causation means one variable's change actually produces a change in the other. The distinction between the two is one of the most frequently repeated warnings in all of statistics precisely because a correlation coefficient, by its mathematical definition, cannot tell the difference between them — a correlation of 0.8 between two variables says exactly the same thing numerically whether one genuinely causes the other, the causal arrow runs the opposite direction, both are driven by some third unmeasured factor, or the relationship is purely coincidental, and no amount of staring at the number alone can distinguish which of these is actually true.

Confounding variables are the most common real-world explanation for a correlation that isn't causal, where a third factor independently influences both variables being compared, creating an association between them with no direct causal link at all: ice cream sales and drowning deaths both rise during summer, correlated with each other not because ice cream causes drowning but because warm weather — the confounder — independently drives both more swimming (and therefore more drowning risk) and more ice cream purchases. This specific example has become one of the standard teaching illustrations in statistics education precisely because the correlation is real and easily verified in actual data, while the suggested causal story is obviously absurd — making it a clear, low-stakes way to build the intuition needed before encountering a confounded correlation in a context where the wrong causal story isn't nearly so obviously absurd.

Reverse causation is a second, equally common explanation that's often overlooked specifically because the assumed causal direction tends to feel intuitively obvious: a correlation between a company's advertising spend and its revenue might be interpreted as advertising driving revenue growth, but the causal arrow could just as plausibly run the other direction — companies with higher revenue having more budget available to spend on advertising — and distinguishing between these two genuinely different stories, with genuinely different business implications, requires more than the correlation itself, since the same observed number is equally consistent with both. Spurious correlation — a statistically real but entirely coincidental association with no underlying causal mechanism connecting the two variables at all — is the third major category, and it becomes increasingly likely to appear by pure chance as more variable pairs are compared (a phenomenon formally related to the multiple comparisons problem), which is why correlations discovered by searching broadly across many variable pairs deserve more skepticism than a correlation predicted in advance from a specific, testable hypothesis.

Establishing genuine causation reliably generally requires either a randomized controlled experiment — where the variable suspected to be the cause is deliberately and randomly assigned, specifically breaking any link to confounding variables, since random assignment ensures confounders are, on average, equally distributed across both groups being compared — or, when randomized experiments aren't ethically or practically possible, rigorous quasi-experimental and causal inference methods (instrumental variables, regression discontinuity, difference-in-differences) that attempt to approximate a randomized experiment's logic using carefully structured observational data instead. This is precisely why randomized controlled trials remain the gold standard specifically for establishing causation (not just correlation) in fields like medicine and policy evaluation — the randomization itself, not the statistical analysis applied afterward, is what does the actual work of ruling out confounding as an alternative explanation.

Read "correlation does not imply causation" not as a vague disclaimer but as a precise statement about what a correlation coefficient can and cannot mathematically distinguish: confounding, reverse causation, and pure coincidence are all real, well-documented, and common explanations for an observed correlation that has no actual causal relationship behind it at all, and only a randomized experiment — or carefully constructed causal inference methodology standing in for one — can reliably rule those alternative explanations out, which is exactly why establishing causation is treated as a fundamentally harder and more demanding claim than establishing correlation.