Zack complicates the story in Causal Diagrams and Causal Models, in an indirect way. There’s a bit of narrative thrown in for fun. I enjoyed this in 2023 but less on re-reading.
I don’t know if the fictional statistics have been chosen carefully to allow multiple interpretations, or if any data generated by a network similar to the “true” network would necessarily also allow the “crazy” network. Maybe it’s the second, based on Wentworth’s comment that there are “equivalent graph structures” (e.g. A → B → C vs A ← B ← C vs A ← B → C). But the equivalent structures have the same number of parameters and in this one the “crazy” network has an extra parameter, so I don’t know. The aside about “the correctness of the family archives” adds doubt. A footnote would help.
The post is prodding me to think for myself and perhaps buy a textbook or two. I could play with some numbers to answer my above doubt. Those are all worthy things, but it’s less clear that they would be worthwhile. Unlike Learn Bayes Nets! there’s no promise of cosmic power on offer.
The alternate takeaway from this post is a general awareness that causal inference is complicated and has assumptions and limitations. Perhaps just that Bayesian Networks Aren’t Necessarily Causal. Drilling into the comments on both posts adds more color to that takeaway.
Overall I’m left feeling slightly let down in a way I can’t quite put my finger on. Like there’s something of value here that I’m not getting, or the author didn’t express in quite the right way for me to pick up on it. Sorry, this is frustrating feedback to get, but it’s the best I can do today.
Zack complicates the story in Causal Diagrams and Causal Models, in an indirect way. There’s a bit of narrative thrown in for fun. I enjoyed this in 2023 but less on re-reading.
I don’t know if the fictional statistics have been chosen carefully to allow multiple interpretations, or if any data generated by a network similar to the “true” network would necessarily also allow the “crazy” network. Maybe it’s the second, based on Wentworth’s comment that there are “equivalent graph structures” (e.g. A → B → C vs A ← B ← C vs A ← B → C). But the equivalent structures have the same number of parameters and in this one the “crazy” network has an extra parameter, so I don’t know. The aside about “the correctness of the family archives” adds doubt. A footnote would help.
The post is prodding me to think for myself and perhaps buy a textbook or two. I could play with some numbers to answer my above doubt. Those are all worthy things, but it’s less clear that they would be worthwhile. Unlike Learn Bayes Nets! there’s no promise of cosmic power on offer.
The alternate takeaway from this post is a general awareness that causal inference is complicated and has assumptions and limitations. Perhaps just that Bayesian Networks Aren’t Necessarily Causal. Drilling into the comments on both posts adds more color to that takeaway.
Overall I’m left feeling slightly let down in a way I can’t quite put my finger on. Like there’s something of value here that I’m not getting, or the author didn’t express in quite the right way for me to pick up on it. Sorry, this is frustrating feedback to get, but it’s the best I can do today.