[memexy]

I guess it's tricky because the real world is full of feedback loops. If you want a causal model for fake news then your model needs to include some representation of incentives for ad revenue and clickbait. How does the causal inference framework handle feedback loops?

[wenc]

I believe DAG-based causal inference isn’t able to handle feedback loops (acyclic) or nonlinearity (linear). Nonlinearities include stuff like deadbands and delays.

Control theory models handle these things just fine. But control models are hard to apply to sociological/epidemiological domains, where causal inference dominates.

From what I gather, causal inference is useful for designing studies. I’m not sure if they’re used for prediction — would appreciate if someone in the know could chime in.

[marmaduke]

> DAG-based causal inference isn’t able to handle feedback loops (acyclic) or nonlinearity (linear)

I don't think this should be true, and if it is, then "causal inference" should be qualified to refer only to a specific modeling framework. As a counter example, it's possible to formulate a nonlinear differential equation model of Covid spread and infer the parameters to construct a plausible, causal, generative model.

[memexy]

Interesting. I wonder if someone has tried to combine the two. I guess modern deep reinforcement learning is one such combination because it combines feedback (reinforcement) and probabilistic descriptions but maybe there are other interesting combinations of probability, causality, and feedback.