Bayesian Vs Causal Modeling | Aleksander Molak, Thomas Wiecki, Carlos Trujillo | PyMC Labs

[@bo44arov]

thank you for discussion!

[@galenseilis5971]

I see value in being able to generate a PyMC model from DAGs. Would be interesting to see how distributional and functional aspects would be handled in a somewhat general way.

[@PyMCLabs]

I agree that would be interesting, but it would require adding more information than what's in the DAG.
The other direction I think is easier, going from a PyMC model to a high-level DAG.
-Thomas

[@galenseilis5971]

@@PyMCLabs Strongly agree. Going from DAG-to-PyMC model would be more appropriate to some more specific family of problems since it requires special assumptions.
I myself would benefit from generating a structural causal model from a PyMC model. Combined with PyMC's do operator it would be easier to generate diagrams for the factual and counterfactual comparison I am trying to communicate to a target audience.
Anyway, thanks for keeping this channel up! Always interesting to see expositions and discussions.

[@PyMCLabs]

@@galenseilis5971 Oh interesting, not sure how these counterfactual diagrams would look like. Want to open a feature request github issue?
-Thomas

[@galenseilis5971]

​@@PyMCLabs The diagrams I am alluding to would just be directed acyclic graphs drawn (with Graphviz or other tools) to represent structural causal models.
For context around that, recall Molak's comment that causal DAGs are not the complete picture of Pearl's causality. When you apply Pearl's do-calculus operators to a causal graph you get a new causal graph.
So if you solve the problem of generating a causal DAG from a PyMC model, and you continue support for the do operator, you should be able to show diagrams for various counterfactuals. Unless I'm confused, there should be no need for a feature request or pull request of a new type of diagram beyond that.