The gap between 'moves with' and 'makes happen'
Correlation says two variables travel together. Causation says intervening on one would move the other. The gap between them is filled with third variables (confounders) that drive both, reverse causation (the effect pushing the apparent cause), and sheer coincidence in a world awash with data. Every classic example is a confounder in a costume: ice-cream sales correlate with drowning because summer drives both; no lifeguard was ever saved by banning ice cream.
Our running dataset: study hours and marks, r = 0.995. The causal story (hours produce marks) is plausible, and still the correlation alone cannot exclude the confounders: students who slept well both study more and score higher; motivated students do both; an easy test inflates marks for the keenest studiers. The experiment that would settle it assigns hours randomly: half the class is made to study two extra hours, half is not, and the mark gap between randomized groups is then a causal estimate, because randomization breaks the confounders' links to the treatment.
Why experiments are the gold standard
Random assignment makes the groups alike on everything except the treatment, in expectation: measured confounders, unmeasured ones, even unknown ones. That is the entire trick, and no statistical adjustment of observational data fully reproduces it. Where experiments are impossible (smoking, poverty, climate), causation is argued from converging evidence: dose-response, mechanism, timing, natural experiments, and studies whose confounders point in different directions.
The verb slide in headlines: a study finds a correlation, the report says "linked to", the headline says "causes". Also its mirror: dismissing any observational finding with "correlation is not causation" as if that ended thought. The slogan is the start of the analysis (which confounder? which direction?), not a substitute for it.
What correlation is still good for
Prediction does not need causation: hours predict marks well enough to flag students at risk, whatever the mechanism. Correlation is also the screening tool that tells researchers where causal questions are worth the expense. It fails only when asked to do the causal job alone: setting policy, assigning credit, or blaming.
Frequently asked questions
What is a confounder, exactly?
A variable that influences both of the correlated variables, creating association between them without either causing the other. Summer temperature confounds ice cream and drowning; motivation confounds study hours and marks.
Can causation exist with zero correlation?
Yes. A U-shaped effect (too little and too much both bad) can produce r near 0 while X strongly causes Y. Cancellation can too: two causal paths of opposite sign. Zero correlation rules out linear association, not causation.
What evidence upgrades correlation toward causation?
Randomized experiments first. Short of that: the cause preceding the effect, a dose-response gradient, a plausible mechanism, consistency across studies with different weaknesses, and confounders measured and adjusted for.
If I control for confounders in regression, is it causal then?
Only if you controlled for all the important ones and modelled them correctly, claims that are hard to verify. Adjustment strengthens an argument; randomization is what settles one.