Saturday, August 02, 2025

Thoughts on Pearl's Book of Why: Lots of Good Intuition on Bayesian Networks, Sometimes with Strong Opinions

The Book of Why by Judea Pearl with great interest. The author is undoubtedly one of the most brilliant thinkers in recent years on causative reasoning and artificial intelligence. The book provides a detailed, somewhat technical, and historical account of causation. I particularly appreciated how it delved into key concepts such as Bayesian networks and confounders, topics that Pearl has explored extensively. 

I also enjoyed his discussion of smoking, the Bradford Hill criteria, and how we needed to move beyond mere correlation and association to make informed decisions. Pearl’s ability to place these discussions in a historical context, highlighting the contributions of figures like Fisher and Sewall Wright, was impressive. I particularly appreciated how Pearl explained the three key graphical elements in Bayesian networks, which he referred to as chains (A leads to B, which leads to C), colliders (A and C both point to B), and forks (B leads to both A and C). He explained how these motifs differ in terms of the independence of the three variables and the extent to which they remain independent when conditioned on B. He provided intuitive examples to help readers easily grasp these concepts and emphasized the fundamental problem of confounding in statistical design, as well as how randomized controlled trials address this issue.  

 However, at times, I found it challenging to distinguish between background information on established concepts and Pearl’s more polemical arguments advocating for what he calls the “causal revolution,” which may be less widely accepted in the statistical and AI communities. This occasionally led to some confusion, but overall, I enjoyed the book immensely and would highly recommend it to any reader.