confounders and reverse causation; ice cream sales and drownings both rise in summer
Correlation is not cause because the numerical link between two variables does not encode how the data were generated. The correlation coefficient only measures how tightly points cluster around a straight line, and that same number can be produced by a direct cause, a shared cause, a reverse cause, or pure chance.
The correlation coefficient is symmetric: the correlation between ice cream sales and drownings is identical to the correlation between drownings and ice cream sales. It carries no arrow of time and no direction of influence. Causal claims need extra information—from randomization, controlled comparisons, or known temporal sequence—that the correlation itself never provides. Without that information, any observed association is a clue to investigate, not a verdict.
Consider twenty days of hypothetical data: ten cool days and ten hot days. On cool days, ice cream sales range from 90 to 110 cones while drownings range from 1 to 2. On hot days, sales range from 190 to 210 while drownings range from 3 to 4. Pool all twenty days together: higher sales are clearly accompanied by more drownings. But look only at the ten cool days. The day with 110 cones has no more drownings than the day with 90 cones. The same is true among the hot days. Once temperature is held fixed, the correlation vanishes. Temperature is the confounder: it pushes both variables at once, and within any single temperature the two move independently.
The mistake is to think that a strong correlation means a strong cause, or that a weak correlation means no cause. Neither follows. A strong correlation can be inflated by one powerful confounder, and a weak correlation can hide a consistent effect that is averaged out across subgroups. Another misunderstanding is that every correlation must have a hidden third variable. Sometimes the hidden explanation is nothing more than chance, especially when many relationships are examined and only the striking ones are remembered.
The warning applies to observational data, where you did not assign who got which exposure. In a randomized experiment, random assignment breaks the link between confounders and the treatment, so the correlation between group assignment and outcome is direct evidence of a causal effect—provided the randomization was done properly and participants stayed in their assigned groups. Even then, the strength of the correlation matters less than the confidence that no other systematic difference between groups remains. Correlation is not cause is not the same as correlation is never useful. It is the starting clue that tells you where to dig; the digging requires a controlled comparison.
Cram When ice cream sales go up, drownings do too. So ice cream causes drowning?
Rep No. Correlating does not prove cause. Both jump with summer heat, a third factor drives both.
Cram So the ice cream is innocent.
Rep Usually. But the trap holds everywhere. Two lines moving together feel like proof, so the question is what else changed at the same time.
Cram How do I spot the hidden cause?
Rep Ask what could push both at once. For sales and drowning the omitted cause is heat. Find that and the link fades.
Cram Could there be more than one explanation?
Rep Always a few. A confounder is anything connected to both sides that you failed to measure. The direction can hide too.
Cram What do you mean direction can hide?
Rep Maybe more drownings make lifeguards sell more ice cream nearby. Cause can flow backwards from what you assumed.
Cram So the story could even be flipped.
Rep Yes. The arrow can run either way, or a third variable owns the whole link. Correlation only says the two tend to move together.
Cram Then what can correlation ever tell me?
Rep It is the starting clue. It tells you where to look. But until you control the confounders and test the direction, it does not prove cause.