The ladder is the argument

The New Science of Cause and Effect, co-written with Dana Mackenzie. Filed as popular science, but really a manifesto with footnotes.

Pearl won the Turing Award for Bayesian networks, then spent thirty years arguing the field he helped build was stuck. This book is that argument, without the math.

The central claim is the Ladder of Causation. Three rungs. Association is seeing: what does this symptom tell me about that disease? Intervention is doing: what happens if I give the drug? Counterfactuals are imagining: would this patient have recovered anyway?

Almost all machine learning lives on rung one. Pearl’s point is that no quantity of data or parameters moves you up. You climb only by adding assumptions the data cannot contain, drawn as a graph. That’s the book.

What it gets right is hard to overstate.

His treatment of Simpson’s paradox is the best I’ve read anywhere. Same numbers, and whether you aggregate the groups or split them flips the conclusion. The data cannot settle which is correct. You need a causal story from outside the data to know which table to read. I’d seen it explained a dozen times and never seen anyone say plainly that it isn’t a statistics problem at all.

The history is good too. Fisher, the finest statistician of the century, argued smoking might not cause cancer because a genetic confounder could explain it. He wasn’t being stupid. The tools to answer him didn’t exist yet.

Now where I’d push back.

It’s cranky. Pearl is fighting a long war with statisticians and it bleeds through. The Neyman-Rubin potential outcomes framework, which most working statisticians and economists actually use, gets a few dismissive pages. You’re reading one side of a live feud presented as settled history.

Philosophers have also asked whether rungs two and three are really distinct, since an intervention is already defined counterfactually. Pearl’s answer is technical and, by his own admission, not obvious from the book.

And the gap I keep hitting in my own work: he shows beautifully what a causal graph buys you, and far less about where the graph comes from. That is the whole problem. My current project leans on CRISPR perturbation screens as interventional ground truth because observational data won’t orient the edges. The book gave me the vocabulary for what I was already trying to do. Not the method.

Verdict: 4.5 out of 5.

Read it if you fit models to data and have ever said “correlation isn’t causation” without being able to finish the sentence. Skip it if you want a textbook. This is a case, not a course.