You are here

The Book of Why: The New Science of Cause and Effect

Judea Pearl
Basic Books
Publication Date: 
Number of Pages: 
[Reviewed by
Brian Borchers
, on

Students of statistics are often reminded that “correlation is not the same as causation.” If two variables X and Y are correlated, it might be that X causes Y, or that Y causes X, or perhaps that a third confounding variable Z has causal effects on both X and Y. More complicated scenarios can easily be imagined. In classical statistics, we may say that X and Y are associated but avoid making statements about causation unless we can perform a randomized controlled experiment to confirm the causality.

For applications of statistics in education, psychology, medicine, and other fields, establishing causality and measuring the size of direct and indirect effects is an important goal. In many cases, controlled experiments are impractical or unethical and only observational data are available. Often there are important variables that cannot be observed and measured directly. Although making causal inferences directly from observational data without further assumptions is impossible, researchers have attempted to develop methods that incorporate causal assumptions into the statistical analysis. For example, structural equation models and instrumental variables are two approaches of this type.

Pearl has developed and championed an approach that uses graphical models to describe assumptions about causal relations between the variables. These causal graphs are then analyzed to determine whether the graph is consistent with available data and to develop specific strategies for controlling for confounding variables and estimating the direct and indirect effects of variables from observational data. Using this approach, the authors move from conventional statistical measures of association to answer questions about interventions such as “What will happen if I increase X?” or “How should I adjust X to increase Y?” He also considers counterfactual questions like “What would have happened to Y if X had been different?”

In the last chapter of the book, the authors consider the implications of this approach for artificial intelligence. Recent progress in AI has been based on the use of statistical methods from machine learning that operate at the level of correlation rather than employing causal inference. Pearl argues that these systems are fundamentally limited by the lack of causal inference and further progress towards strong AI will require researchers to develop systems that incorporate causal inference.

The Book of Why has been written for a general audience. There are very few formulas or formal definitions. The main ideas of Pearl’s approach are explained with examples from applications in the social and life sciences and illustrated with causal diagrams. Causal diagrams are a very intuitive way of communicating causal relationships to a non-technical audience that might have great difficulty in understanding the assumptions required by other approaches to causal inference.

The Book of Why should serve to introduce a very broad audience to the challenges of causal inference and Pearl’s approach. However, it is not particularly suitable for use as a course text or reference. Readers looking for a how-to guide to Pearl’s approach will find it in Causal Inference in Statistics: A Primer. Another book by Pearl, Causality: Models, Reasoning and Inference is recommended for those who want to dive deeply into the subject. It should also be noted that there are other popular approaches to causal inference, most notably Donald Rubin’s potential outcomes framework. Readers interested in alternatives to Pearl’s approach should consider Counterfactuals and Causal Inference by Morgan and Winship.


Stephen L. Morgan and Christopher Winship. Counterfactuals and Causal Inference: Methods and Principles for Social Research, 2nd edition. Cambridge University Press, 2014.

Judea Pearl. Causality: Models, Reasoning and Inference, 2nd edition. Cambridge University Press, 2009.

Judea Pearl, Madelyn Glymour, and Nicholas P. Jewell. Causal Inference in Statistics: A Primer. Wiley, 2016.

Brian Borchers is a professor of Mathematics at the New Mexico Institute of Mining and Technology. His interests are in optimization and applications of optimization in parameter estimation and inverse problems.

The table of contents is not available.