0
$\begingroup$

After having read Book Of Why and Causal Inference In Statistics - A Primer, I was reading Causality - all by Judea Pearl.

Yet I found that there were quite some points which I was not understanding. Particularly, Markov and Bayesian networks, axiomatization of probabilistic system using information relevance axioms, etc. However, having followed the references given to Probabilistic Reasoning In Intelligent Systems by Pearl, they were much more clear.

Hence, is Probabilistic Reasoning In Intelligent Systems, or parts of it, a pre-requisite of Causality? If yes, which parts might be good to go through?

Improve this question
$\endgroup$
2

1 Answer 1

Reset to default
2
$\begingroup$

I have only read Causality (2nd ed) from Judea Pearl. I would say that, at a minimum. a graduate level course in probability and statistics is the only "pre-req". I would say each of Judea's texts are intended for different audiences, rather than meant to comprise a didactic arc. Judea's approach in Causality is highly technical, and follows much the same line of reasoning as Rubin's Potential Outcomes Framework https://www.tandfonline.com/doi/abs/10.1198/016214504000001880. Judea and Rubin specifically develop an idea of causality as an "algebraic" concept, i.e. I can "do" anything conceptually. For a softer take that represents a different perspective, consider Hernan and Robin's book "Causal Inference" which I consider to be excellent https://www.hsph.harvard.edu/miguel-hernan/causal-inference-book/

Improve this answer
$\endgroup$
8
  • 1
    $\begingroup$ I disagree. Without an understanding of Bayesian statistics and Bayesian networks, most readers would be lost in Causality. $\endgroup$
    – Adrian Keister
    Commented Jul 21, 2022 at 16:55
  • $\begingroup$ Pearl himself has admitted his references to these networks in Causality are purely incidental. See related question here: stats.stackexchange.com/questions/271103/… and his twitter interactions seem to suggest the concept is a work-in-progress twitter.com/yudapearl/status/1203502110286811136?lang=en $\endgroup$
    – AdamO
    Commented Jul 21, 2022 at 17:01
  • $\begingroup$ Not seeing the relevance in those two links. $\endgroup$
    – Adrian Keister
    Commented Jul 21, 2022 at 17:04
  • $\begingroup$ @AdrianKeister in other words, you can develop a causal framework based on DAG by assuming WLOG that a modeling procedure exists for which the residuals for a particular node are conditionally independent of its ancestors (some combination of which are available as covariate inputs). Whether Bayesian, non-parametric, or parametric SEM turns out not to matter. $\endgroup$
    – AdamO
    Commented Jul 21, 2022 at 17:04
  • 1
    $\begingroup$ Well, that's very nice, I would agree. But you're speaking more from the perspective of someone already "in the know". I, for example, found Causality nearly unreadable because I don't have the prerequisites I have mentioned in my answer here (stats.stackexchange.com/questions/568281/…), and I suspect many other readers would feel the same way. When I read a book, I am always driving towards "depth-first" or total understanding. Without Bayesian networks, I don't see how that's possible. $\endgroup$
    – Adrian Keister
    Commented Jul 21, 2022 at 17:08

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.