Skip to content

Causal Inference

Moving beyond correlation to cause — the tools for asking 'what if?' and 'why?', not just 'what is likely?'.

Most computer predictions answer what is likely — like noticing that people who carry umbrellas often get rained on. Causal inference is the science of answering a harder question: why, and what if we change something?

Analogy: imagine your bike and a friend's bike are always parked together, so seeing one usually means the other is nearby. That's a pattern, but moving your bike doesn't teleport your friend's. Two things happening together doesn't mean one causes the other.

Causal inference gives us careful tools to tell the difference — to figure out whether pressing a button actually makes something happen, or whether it just happens to appear alongside it. This matters whenever we act on a prediction: a doctor choosing a treatment, or a school changing a rule. Getting cause and effect right is how we make good decisions, not just good guesses.

The main ideas

  • Correlation vs causation — Why predictive accuracy alone isn't enough for decisions and interventions.
  • Causal graphs & do-calculus — Representing assumptions as DAGs and reasoning about interventions.
  • Counterfactuals — Estimating what would have happened under a different action.
  • Experiments & A/B testing — Randomized trials — the gold standard for causal claims.
  • Observational methods — Instrumental variables, matching, and difference-in-differences when you can't experiment.
  • Uplift modeling — Predicting the effect of an action per individual — who to treat, not just who will convert.

Machine Learning · Data & MLOps · Recommender Systems

Want to make things?

Head to AI School — AI camps where kids build their own games.