Most important business and policy questions are causal: what will happen if we change something?
Data alone usually shows only association, which is why this distinction matters so much.
Treating a correlation as a cause can lead to spending money on changes that do nothing.
A correlation between A and B has 4 broad explanations: A causes B, B causes A, something causes both, or chance.
Reverse causation is easy to miss, as when people who fall ill stop exercising, making exercise look more protective than it is.
A shared cause is the most common trap in data that was not produced by an experiment.
Chance correlations are everywhere in large datasets, since testing enough pairs guarantees some strong ones.
Time order helps, because a cause must come before its effect, but it is not enough on its own.
A dose response, where more of A goes with more of B, strengthens a causal case without proving it.
A plausible mechanism, explaining how A could produce B, makes a causal story more convincing.
Correlation is still useful, because it can predict well even when it says nothing about cause.
For example, umbrella sales predict rain well, but buying umbrellas does not cause rain.
Experiments settle causal questions most directly, and careful designs without experiments can come close.
In practice, a causal claim should state its design and the other explanations it has ruled out.
The usual slogan says correlation is not causation, but the fuller truth is that correlation alone cannot say which cause it is.