The publication of the first edition of this book was a landmark in the study of causality [1]. It provided a coherent statement of more than two decades of previous research by Pearl and his students and colleagues. Pearl’s approach rigorously distinguishes causality--which accounts for dynamics and change--from statistics--which necessarily describes a static state of affairs--and provides a formal mathematical foundation for causality, in terms of surgery on inference graphs. The systematic exposition of this approach in the original monograph stimulated a much broader discussion of these issues, warranting a second edition that takes into account new developments.

This edition is a gentle extension of the first one. Almost all section headings are the same, in the same order, and even start on the same page numbers. Additional material relevant to individual chapters has been included as postscripts to each chapter. In addition, a new chapter of more than 60 pages addresses questions Pearl received from readers and students since the first edition’s publication. The bibliography has been thoroughly extended to account for literature published through 2009.

Two main themes dominate the book: confounding variables and counterfactuals. After introducing basic statistical notions and the relation of causality to Bayesian networks in chapter 1, Pearl turns to the problem of identifying causal relations in chapters 2 and 3. Confounding variables--variables that lie on a causal pathway that is being evaluated--are a major challenge to this program and manifest themselves in varieties of Simpson’s paradox, in which an effect at the level of a population appears to be reversed when one looks at nonintersecting but covering subsets of the population. This theme surfaces repeatedly in chapters 3 to 6, with chapter 6 offering a summary discussion of the paradox and ways of resolving it. Chapter 7 introduces the notion that causality can be defined in terms of counterfactuals; this theme dominates the subsequent chapters on imperfect experiments, the probability of causation, and the actual cause.

Pearl is not only concerned with setting forth a mathematical theory of causation, but also with probing the history and sociology of the field. He traces the graphical formalism with which he resolves the problem of confounding variables to the earlier use of structural equation models in sociology and economics. Although more recent users of these models eschew causal interpretation, Pearl shows how the intent of their originators was clearly causal. He traces the other big idea, counterfactuals, to the potential outcome framework of Neyman and Rubin (and less formally, to Hume and Mill). More broadly, he introduces readers to the ongoing tension between those who have asserted the completeness of statistics as a language for describing science and those who have insisted on the need for a distinct representation for causality. Pearl’s great contribution is reinforcing this distinction and providing a workable mathematical formalism in which to discuss it.

Pearl’s career has been motivated by problems of artificial intelligence, but the implications of this book are much broader. The distinctions he raises and the mathematical foundation he assembles are critical for every field of scientific endeavor. This updated edition of a modern classic deserves a broad and attentive audience.