core Estimated learning time: 4 h

1.23 Correlation, causation and confounding

You can say when a model's finding supports an action.

Before:00. Orientation & SetupUnlocks:03. Data Handling & Analysis04. Classical AI — Agents, Search & Knowledge Representation

Models find correlations; decisions need causes, and this topic is the bridge to cross before recommending action from any model. Confounders and Simpson's paradox show how a true correlation can point the wrong way entirely, and causal graphs give a notation for thinking it through. It sits near the end of statistics because it keeps everything the models later produce honest. The confusion that costs money is assuming a strong predictor is a lever — a variable can predict beautifully and move nothing when pushed.

Work through these

  • Spurious correlation and Simpson's paradox

    Two quantities can move together for reasons that have nothing to do with each other, and aggregated data can reverse a relationship present in every subgroup. Both are common and both are counter-intuitive.

  • Confounders, colliders, mediators

    Three kinds of third variable, each of which requires opposite handling: one you must adjust for, one you must not, and one that sits on the path itself. Getting this backwards produces confident nonsense.

  • Causal graphs (DAGs) as a thinking tool

    Drawing the assumed causal structure as a diagram makes it possible to reason about what may be adjusted for. It is a thinking tool rather than a statistical procedure.

  • Natural experiments, IV, difference-in-differences

    When an experiment is impossible, certain natural situations approximate one, and there are standard designs for exploiting them. These are how causal questions get answered outside a laboratory.

Sign in to keep your progress.

Free resources

Links last checked 29 Aug 2026.

Stuck here?

Ask a mentor. A real person answers, and they can see exactly which topic you're on. Usually within a couple of working days.

Checking your session…

Topics shown in module order.