Beyond Correlation: A Review of Pearl’s Causality

The Book of Why tells you the ladder exists. This one hands you the machinery and expects you to keep up. d-separation, for reading independences straight off a graph. Back-door and front-door criteria. The do-calculus. Structural counterfactuals. Probability of causation. Bounds for when your experiment is imperfect.

The centerpiece is the do-calculus, and it is three rules. Three rules for rewriting an expression containing do() into one containing only quantities you can observe.

What makes that remarkable came later. In 2006 Huang and Valtorta, and Shpitser and Pearl independently, proved the three rules are complete. If a causal effect is identifiable from your graph and data at all, some sequence of the three will find it. If no sequence works, it is not identifiable, full stop. A finite rule set that provably exhausts its problem.

The piece I keep returning to is the front-door criterion, and it pairs with 011.

Fisher argued smoking might not cause cancer, because a genetic confounder could produce the correlation. The front-door criterion answers him. If a mediator sits on the path, tar in the lung, and the confounder doesn’t touch it, you can identify the effect of smoking on cancer without ever measuring the gene. Unmeasured confounder, and you still get the number. That shouldn’t be possible. It is.

Where I’d push back.

It reads as what it is, twenty years of papers assembled into a book. Chapter 3 is teachable. Chapter 7, on structural counterfactuals, is a different difficulty class and the notation shifts under you.

Chapter 11 in this edition is sixty pages of Pearl answering his critics. As intellectual history it’s fascinating. In a textbook it’s unusual, and the book argues with you about as often as it teaches you.

The comparison with Rubin’s potential outcomes runs throughout and is useful, but it is Pearl’s reading of Rubin. Read Imbens and Rubin alongside for the other half.

And chapter 2, on causal discovery, is the thinnest chapter in the book. It’s also the one I need most. Same complaint as 011: the identification machinery is magnificent once you already have the graph.

Verdict: 5 out of 5 as a reference. 3 as something you read front to back. Nobody reads it front to back.

Reach for it when you hit a question the Primer can’t answer. Start with the Primer if you haven’t.