Causality Discussion Group - Abstracting Causal Models
Details
Title : Causality Discussion Group - Abstracting Causal Models Author(s): Sander Beckers Link(s):
Rough Notes
Problem setting
Causal models describe causal relations between variables - but what are "good" variables?
Some context-dependent criteria:
- What can I learn?
- What can I compute?
- What can I observe?
- What can I do? (Manipulation variables)
- What do I want to achieve? (Outcome variables)
- How accurate do I want to be? (Approximation)
Suppose we get some variables \(V\) from these criteria. Is there a single causal model that contains all of \(V\)?
Some general criteria:
- Values of a variable partition all possibilities.
- Independent manipulability.
The problem is now: What should we do if we cannot meet both sets of criteria (context-dependent and general criteria)?
Causal abstractions
Causal abstractions are 1 kind of solution to the problem above.
The solution involves finding two sets of variables such that:
- Separately they meet the general criteria.
- Combined they meet the context-dependent criteria.
- There is an asbtraction function between them.
See also Causal Models with Constraints, Beckers et al CLEAR 2023.