Causal Inference
Working out what would happen if you intervened, rather than what tends to occur together — the distinction that decides whether a model can support a decision.
Reviewed July 16, 2026Stable
When not to use it
- When you only need ranking or forecasting and no intervention follows, where predictive accuracy is the right target.
- As a claim of causality from observational data without stating the identifying assumptions.
Reach for something else instead
- Randomised experiments, which answer the question directly when feasible.
- Predictive modelling, when the decision does not involve changing anything.
- Sensitivity analysis, which bounds how strong unobserved confounding would need to be to overturn a finding.
Sources & further reading
- Pearl (2009), Causality: Models, Reasoning and Inference — the graphical framework and the do-operator.
- Caruana et al. (2015), Intelligible Models for HealthCare — the asthma and pneumonia case in detail.
- Angrist & Pischke (2008), Mostly Harmless Econometrics — the applied observational toolkit.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Controlling for every available variable, which introduces collider bias rather than removing confounding.
- Treating a coefficient in a predictive regression as a causal effect.
- Assuming that more data resolves confounding. It narrows confidence intervals around a biased estimate.
At a glance
FieldMachine Learning
Answerswhat if I intervene
Gold standardrandomisation
Key hazardcollider bias from over-controlling
Prior questionidentification
DifficultyTechnical
Where this sits
A destination. 3 concepts lead here, and nothing in the corpus depends on it.
3Levelsteps in
3Needs firstconcepts
0Opens upnothing further
1Areastays here
Learn these firstRegression
Computed from the prerequisite graph, not assigned. How this works