causal inference

How do we infer causality from observational data? This question is important in science and is closely related to the progress of off-policy RL.

It's not even obvious that a passive observer should be able to learn causal effects. Wouldn't they be

The simplest way is to observe someone conduct a randomized controlled trial. This establishes that it's clearly possible. But you need a model of the situation, e.g., you need to understand that the treatments are assigned randomly.

More generally, we can observe a natural experiment in which we're willing to model some factor as being endogenously manipulated: for example, in 1854, some water pumps in London were fed from upstream of a sewage discharge point, and others fed from downstream, and a cholera outbreak was concentrated around those pumps (notably the Broad St pump: 1854 Broad Street cholera outbreak - Wikipedia).

Relevant frameworks:

• potential outcomes
• do-calculus
• instrumental variables