The Book of Why: Review

Joseph Y. HalpernarXiv 2019 · arXiv

Abstract

Joseph Y. Halpern reviews Judea Pearl’s The Book of Why, a landmark text that argues data alone is insufficient for understanding the world. The review highlights how Pearl’s "causal revolution" provides the mathematical tools—like graphical models and the "do-operator"—to move beyond the cliché that "correlation is not causation" and into the realm of formal causal inference.

The Problem: Data is "Dumb"

For decades, the field of statistics and modern machine learning have operated under the assumption that all necessary information is buried within the data. The prevailing logic suggests that with enough data and computing power, any relationship can be uncovered.

Halpern’s review focuses on Pearl’s central challenge to this view: Data is fundamentally "dumb." Data can reveal that two variables, $X$ and $Y$, move together, but it cannot explain why without a pre-existing causal model. The paper addresses the gap between passive observation ("seeing") and active intervention ("doing"), explaining why we need structural models to answer "what if" questions and avoid common pitfalls in data analysis.

Conceptual Frameworks for Causality

Pearl introduces several conceptual frameworks to formalize causality, moving it from a philosophical debate to a mathematical discipline:

  • Causal Diagrams (Directed Acyclic Graphs): These use nodes for variables and arrows to represent direct effects, allowing researchers to visualize the flow of influence.
  • The "do-operator": A mathematical notation, $Pr(Y = b \mid do(X = a))$, representing an intervention where we force $X$ to take a certain value, rather than just observing it.
  • The Back-door Criterion: A graphical method to determine which variables must be "controlled for" to isolate the causal effect of $X$ on $Y$. This provides a formula to calculate interventional results using only observational data.
A causal diagram illustrating a confounding variable Z that affects both X and Y.

Key Insights: The Danger of Over-Controlling

A major contribution of the book is clarifying when not to control for variables. A common mistake in research is the belief that controlling for more variables is always better. Halpern highlights that controlling for a "collider"—a common effect of two variables—can actually create a spurious causal connection where none exists.

A diagram showing a collider structure where B is a common effect of A and C.

Traditional statistics only deals with "seeing" (conditional probability), whereas causal inference introduces "doing" (interventions). The review emphasizes that one cannot derive a "do" from a "see" without a model of the underlying system. The "Back-door Formula" allows us to compute the effect of an intervention by looking at conditional probabilities in observational data, provided the back-door criterion is met.

Critical Assessment and Future Directions

Halpern’s assessment of the book is largely positive, though he offers technical critiques regarding the balance between accessibility and rigor. He suggests that including formal equations earlier might have benefited more technical readers. Furthermore, he notes that while the book focuses on Pearl’s journey, other pioneers like the "CMU school" (Spirtes, Glymour, and Scheines) deserve significant credit for their work in graphical causal modeling.

A major ongoing challenge remains "causal discovery": the process of inferring causal graphs directly from data. While Pearl focuses on what to do given a model, the field is still working on how to automatically infer model structures from raw data. Halpern concludes that the "causal revolution" is still in its early stages, and its full integration into AI and science is an ongoing endeavor.

Summary updated

BibTeX
@misc{halpern2019book,
  author = {Joseph Y. Halpern},
  title = {{The Book of Why: Review}},
  year = {2019},
  eprint = {1909.13485},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1909.13485},
}
Figure