---
tag: V3QJFN5800
kind: concept
name: Causal inference
msc: [62D20, 62A01]
sources:
  - url: https://en.wikipedia.org/wiki/Causal_inference
same_as:
  wikidata: Q5054566
  wikipedia: Causal inference
---

## Summary

Causal inference is the part of statistics that asks what would happen to an outcome if one intervened on a variable.
This differs from asking which values tend to occur together.
It makes precise the difference between correlation and causation, through frameworks such as the [[Rubin causal model|23ATBZMJHR]] of potential outcomes and causal graphs.
In a randomised experiment causal effects can be estimated by simple comparisons, thanks to [[randomization|734Q6Q51X7]].
With observational data one needs extra assumptions, such as the absence of unmeasured confounders, or special designs such as [[instrumental variables|6ETK64HYK6]].

## Explanations

### L1 human

People who carry umbrellas get wet more often than people who do not, but umbrellas do not cause rain.
Causal inference studies when data can tell us what a change would actually do, for example whether a drug lowers blood pressure.
The key problem is that each person either takes the drug or does not, so we never see both outcomes for the same person.
A coin flip that decides who gets the drug makes the two groups comparable, so the difference of their averages measures the effect.
Without such a coin flip, hidden differences between the groups can create or hide an effect, as in [[Simpson's paradox|T060JHSH05]].