Correlation Is Not Causation — Formally

Judea Pearl's framework separates three rungs that are routinely conflated:

• Rung 1 — Seeing (association): P(Y | X). 'Users who use feature X retain better.' Every ML model, every dashboard, lives here.
• Rung 2 — Doing (intervention): P(Y | do(X)). 'If we force users onto feature X, does retention move?' This is the question every product decision actually asks — and rung-1 data cannot answer it.
• Rung 3 — Imagining (counterfactual): 'Would this user have churned anyway?' Necessary for attribution, fairness questions, and postmortems.

The whole discipline in one line: observing X→Y differs from setting X and watching Y, and the difference is confounding.