Correlation-Causation Fallacy
Ice cream sales and drownings rise together, but ice cream doesn't drown anyone. This fallacy is assuming that when two things happen together, one must cause the other.
Two things rise and fall together, so one must be causing the other, right? Not necessarily. That leap is the correlation-causation fallacy. A correlation only tells you two things are related somehow. A third factor might be pulling both strings, or the link might be pure coincidence, the kind of random pattern our brains love to over-read (see the clustering illusion).
Two classic examples:
Ice cream sales and drowningsAs ice cream sales climb, so do drownings. Does ice cream make people drown? Of course not. Both go up in summer, when it's hot, people flock to beaches and pools, and a cone sounds great. The warm weather is the common factor, not the ice cream.
Shoe size and reading abilityKids with bigger shoes tend to read better. Big feet don't make better readers, though. Both track the child's age: older kids have larger feet and more advanced reading skills. Age is doing the work, not shoe size.
So when you spot a correlation, ask what else could explain it before you decide one thing causes the other.
Nobody seriously argues that correlation proves causation. The pushback is about the slogan itself: "correlation is not causation" can be misused as badly as the fallacy it warns against.
The opposite fallacyDismissing a correlation out of hand is its own mistake. Correlation isn't proof, but it is a clue, and often the first one. Many causal discoveries began as a suspicious correlation that someone took seriously enough to investigate.
The smoking lessonIn the 1950s the great statistician Ronald Fisher repeatedly argued that the link between smoking and lung cancer was only a correlation. He suggested that early cancer might make people want to smoke, or that a shared genetic trait might cause both. The tobacco industry welcomed the argument. Richard Doll and Austin Bradford Hill answered not with a single experiment but with converging evidence: a dose-response pattern (more cigarettes, more cancer), the right timing, and consistency across many studies. History sided with them.
Observational data can support causal claimsRandomized trials aren't always possible or ethical. Epidemiologists use the Bradford Hill criteria as a checklist for weighing causal evidence, and modern causal inference (Judea Pearl's work, natural experiments, counterfactual methods) offers formal tools for drawing causal conclusions from non-experimental data when the assumptions are made explicit.
A conversation-stopperOnline, the phrase is often deployed to wave away any finding someone dislikes, a handy shield for confirmation bias. Asking "what's the confounder, and how big would it need to be?" is far more useful than repeating the slogan.
The core warning stands: a correlation alone can't tell you which way the arrow points, or whether there's an arrow at all. But the right response is to dig further, not to shrug.
Resources
The correlation-causation fallacy, also called cum hoc ergo propter hoc (Latin for "with this, therefore because of this"), is the error of concluding that because two variables move together, one must be driving the other. Correlation can be evidence that a relationship exists, but it does not establish causation, and reading cause straight off a correlation is a reliable route to wrong conclusions.
The fallacy sits at the centre of several related ideas:
Spurious correlationsCorrelations produced by chance, or by a shared underlying factor (a confounding variable), rather than by any causal link. They are the raw material of the fallacy, because people readily mistake these coincidental relationships for cause and effect. The same pattern-hunting drives the clustering illusion, and in animals it can be seen in Skinner's superstitious pigeons, which repeated whatever they happened to be doing when food arrived.
Confounding variablesFactors related to both the independent and dependent variables that distort the relationship observed between them. Accounting for them is essential for establishing causality, since a confounder can manufacture a causal-looking link or hide a real one.
Experimental designGood design, above all the randomized controlled trial, can establish causality by controlling for potential confounders and isolating the effect of the independent variable on the dependent one. Observational studies, by contrast, can only show correlations, so causal claims from them need additional evidence or analysis.
Granger causalityA statistical test of whether one time series helps predict another, which can offer evidence of a causal relationship. It is useful for flagging candidate causes in time series data, but it cannot settle causality on its own, because other factors may still confound the relationship.
Knowing how the fallacy connects to these concepts helps researchers and decision-makers resist over-reading correlational data and go looking for the more rigorous evidence that causal claims require.
References
- Holland, P. W. (1986). Statistics and causal inference. Journal of the American Statistical Association, 81(396), 945-960.
- Pearl, J. (2009). Causality: Models, Reasoning, and Inference (2nd ed.). Cambridge University Press.
- Granger, C. W. J. (1969). Investigating Causal Relations by Econometric Models and Cross-spectral Methods. Econometrica, 37(3), 424-438.
- Shadish, W. R., Cook, T. D., & Campbell, D. T. (2002). Experimental and Quasi-Experimental Designs for Generalized Causal Inference. Houghton Mifflin.