- [[confounding variable vs risk factor]]
# Idea
Unmeasured attributes or variables that influence a relationship between two variables are known as **confounders**, **lurking variables**, or **unobserved heterogeneity**. They are often variables that undermine [[causal inference]]. In general, it's extremely difficult or even impossible (in the social sciences) to know what all the confounders are.
![[20250216180243.png]]
Confounders cause/affect both the treatment and outcome. $X$ is a confounder:
```mermaid
graph LR
X --> T
X --> Y
T --> Y
```
If there are unmeasured/unobserved confounders that affect both the treatment assignment and outcome, we can get biased estimates of causal effects. The [[ignorability assumption]] is violated. But we can use methods like [[instrumental variables|instrumental variable analysis]] that do not rely on the [[ignorability assumption]].
To "address" the problem of unobserved confounders, researchers try to use methods that don't require them to identify or measure all potential confounders, and [[experiment|experimentation]] is a solution.
# References
- [[Gerber 2012 field experiments]]
- [\| Codecademy](https://www.codecademy.com/courses/learn-the-basics-of-causal-inference-with-r/lessons/potential-outcomes-framework/exercises/confounders)