Amey-Thakur Correlation vs causation
"X correlates with Y" is easy to measure and easy to over-read. The leap to "X causes Y" is where analysis most often goes wrong, driving real decisions off a relationship that is coincidence, confounding, or backwards. Knowing when you can and cannot claim causation is a core data-science guardrail.
Method
- Default to "correlation is not causation". A relationship in data has several possible explanations: X causes Y, Y causes X (reverse), a third factor Z causes both (confounding), selection into the sample created it, or pure chance. The correlation alone cannot distinguish them. Start skeptical of the causal story.
- Hunt the confounder. The classic trap: ice cream sales correlate with drownings (summer causes both). Before claiming X causes Y, ask what Z could drive both. Confounders are everywhere (seasonality, age, wealth, the underlying trend); an uncontrolled comparison is a confounder waiting to be named (see mental-models' first-principles).
- Check the direction. Does X cause Y, or does Y cause X? "Users of feature A retain better" may mean A drives retention, or that already-engaged users (who would retain anyway) adopt A. Reverse and bidirectional causation masquerade as forward causation constantly.
- Suspect selection effects. If the sample was chosen in a way related to the variables (survivors, opt-ins, the treated group differing from the control at baseline), a relationship can appear from the selection itself, not any causal link. Ask how the data was collected and who is missing (see the sampling caution in statistical-inference).
- Reserve causal claims for causal evidence. The strong evidence for causation is a randomized experiment (randomization breaks confounding; see ab-test-design, experiment-design-basics). Without it, causal inference needs careful methods (natural experiments, controlling for known confounders, instrumental variables) and still carries caveats. Observational correlation supports "associated with", not "causes".
- Frame conclusions honestly. Say "associated with" or "predicts" for correlational findings, "causes" only when the evidence earns it. The discipline is in the wording: a decision-maker who hears "causes" acts as if intervening on X will move Y, which a mere correlation does not promise.
Boundaries
- Prediction does not require causation: a model can use correlated features to predict well without any causing the target, and that is fine for forecasting (see ml-problem-framing). The trap is only when you act on the relationship as if intervening.
- Absence of proven causation is not proof of no causal link; correlational evidence plus mechanism and dose-response can build a strong (if not certain) causal case where experiments are impossible.
- Causal inference from observational data is a deep field; this skill is the guardrail against the common leap, not the full methodology (see experiment-design-basics, statistical-inference).