Skip to main content

The Book of Why — Pearl & Mackenzie

Review of Pearl & Mackenzie’s The Book of Why: causal inference for a general audience, the ladder of causation, and why every ML practitioner should understand do-calculus.

finished certainty: likely importance: 8

The Book of Why: The New Science of Cause and Effect (2018) by Judea Pearl and Dana Mackenzie is Pearl’s attempt to explain his life’s work on causal inference to a general audience. Pearl is not a modest man—he believes, with some justification, that he has solved one of the oldest problems in science, and that most of statistics has been doing it wrong for a century. The book radiates this conviction on every page, which makes it simultaneously inspiring and occasionally irritating.1

The Core Argument

Pearl’s central claim is that statistics, as traditionally practiced, is fundamentally limited because it deals only in associations—correlations, regressions, conditional probabilities—and cannot answer causal questions. “Does smoking cause cancer?” is not a question that any amount of observational data can answer using standard statistical tools, because correlation famously does not imply causation. Pearl argues that to reason about causation, you need a formal language that goes beyond probability theory, and he has spent decades building one.

The key framework is the ladder of causation, which distinguishes three levels of cognitive ability:

  1. Association (seeing): What is the probability of Y given that I observe X? This is standard statistics.

  2. Intervention (doing): What happens to Y if I do X? This requires the do-calculus and causal diagrams (DAGs).

  3. Counterfactual (imagining): What would have happened to Y if X had been different? This requires structural causal models.

Pearl contends that most of machine learning operates exclusively at level 1, and that this is a fundamental limitation—not just a practical one but a theoretical ceiling on what purely associational systems can learn.2 A model trained on observational data alone cannot, in general, predict the effects of interventions, no matter how much data it has or how flexible its architecture.

What’s Great

The historical narrative is genuinely compelling. Pearl traces the “causal revolution” through a series of intellectual battles: the rejection of causality by early statisticians like Karl Pearson and Ronald Fisher, the development of path analysis by Sewall Wright (a geneticist, not a statistician—Pearl notes this pointedly), the formalization of randomized controlled trials, and finally Pearl’s own work on Bayesian networks and causal diagrams.

The smoking-cancer case study is the book’s best chapter. Pearl shows how the tobacco industry exploited the legitimate philosophical difficulty of inferring causation from observational data, and how the tools he developed—specifically the front-door criterion—could have settled the debate decades earlier. It is a masterful example of how abstract mathematical theory connects to life-and-death policy questions.3

The explanation of confounding and Simpson’s paradox is the clearest I’ve encountered in any source. Pearl shows that Simpson’s paradox is not a statistical curiosity but a symptom of the fundamental inadequacy of purely associational reasoning: the same data can tell opposite stories depending on the causal structure, and no amount of statistical sophistication can resolve the ambiguity without causal assumptions. This is a deep point that is often hand-waved in statistics courses.4

What’s Problematic

Pearl is fighting a war on multiple fronts—against traditional statistics, against the Rubin causal model (potential outcomes framework), against machine learning’s purely predictive orientation—and this combativeness sometimes overwhelms the exposition. He is uncharitable to the potential outcomes framework, which in practice produces similar answers to his approach for many common problems, just with different notation.5 The rivalry between the Pearl and Rubin schools has generated more heat than light, and the book amplifies rather than resolves this.

The popular-science format occasionally works against the material. Pearl’s technical work is genuinely deep and mathematically precise; the book necessarily simplifies it, and sometimes the simplifications are misleading. The do-calculus is presented as almost magical—“just draw the diagram and read off the answer”—without adequate discussion of how hard it is in practice to get the diagram right. The causal diagram must be specified by the analyst based on domain knowledge; it is not learned from data (at least not without strong assumptions). This is the Achilles’ heel of the entire framework, and the book underplays it.6

The writing can be repetitive. Pearl makes the same basic point—statistics can’t do causation, you need causal models—many times in slightly different ways. By the midpoint of the book, the reader who is already convinced feels somewhat bludgeoned.

Relevance to Machine Learning

Pearl’s critique of ML as “glorified curve fitting” is provocative and partially correct. Modern ML systems excel at association (level 1 on the ladder) but struggle with interventional and counterfactual reasoning. This matters practically:

  • Recommendation systems that optimize for observed click-through rates are confounded by selection effects—they learn what users were shown, not what users would prefer.

  • Medical AI trained on observational health records cannot directly predict the effects of treatments without causal adjustments.

  • Fairness and bias in ML is fundamentally a causal question: does the model discriminate because of protected attributes, or merely correlate with them?

  • Robustness to distribution shift is arguably a causal problem: systems that learn causal mechanisms generalize better than those that learn spurious correlations.

The emerging field of causal ML tries to bridge Pearl’s framework and modern deep learning. Notable work includes causal representation learning, invariant risk minimization (Arjovsky et al. 2019), and the use of do-calculus for debiasing. Pearl’s book is the accessible entry point to the ideas motivating this research.

The Verdict

The Book of Why is an important book that makes a genuine intellectual contribution accessible to non-specialists. The core argument—that causal reasoning requires tools beyond standard probability and statistics—is correct and important. The ladder of causation is a powerful conceptual framework. The historical narrative is engaging.

It would be a better book if Pearl were less combative and more honest about the limitations of his framework (especially the reliance on analyst-specified DAGs). It would also benefit from more engagement with the cases where causal inference is genuinely hard—where the DAG is uncertain, where there are unmeasured confounders, where the system is too complex for clean causal diagrams.7

For ML practitioners, this is essential context. Understanding why your predictive model might fail when deployed in a new environment, or why A/B testing is necessary even when you have mountains of observational data, requires the causal perspective Pearl champions. You don’t need to become a do-calculus expert, but you need to understand the basic framework.

Rating: 7/10. Read it for the ideas, tolerate the polemics, then read Pearl’s more technical Causality (200026ya/200917ya) or Causal Inference in Statistics: A Primer for the actual tools.8

Further Reading