Judea Pearl (2009), 2nd edition, Cambridge: Cambridge University Press — DOI-10.1017/CBO9780511803161
Key takeaways
From a skim of four excerpts, the contents, the Preface to the Second Edition, §§5.3.2–5.4.1 and the Epilogue (NOTE-696). The rest of the book was not read.
- The do-operator. E(Y | do(x)) is “the controlled expectation” and E(Y | x) “the standard conditional or observational expectation” (p. 158). Under do(X = x) the equation for X “should be overruled and replaced with the equation X = x”, and the solution of the new set is the new equilibrium (p. 158). “The passive observation X = x should not violate any of the equations” (p. 158). In probability, “the vertical bar stands for the phrase: ‘given that we see’” (p. 421).
- Intervention as surgery. “Intervention amounts to a surgery on equations (guided by a diagram) and causation means predicting the consequences of such a surgery” (p. 417). It presupposes that each equation is an autonomous mechanism (pp. 414, 420), and it is undefined for equations rewritten into an equivalent form that loses “the equation for Y” (p. 417).
- What a structural equation means. Definition 5.4.1 (p. 160): y = βx + ε is structural if, “in an ideal experiment where we control X to x and any other set Z of variables … to z, the value y of Y is given by βx + ε, where ε is not a function of the settings x and z”. So β = ∂E[Y | do(x)]/∂x (5.24), which “has nothing to do with the regression coefficient” (p. 161), and the error term is ε = y − E[Y | do(x)] (5.25).
- Two components, three rules. “Scientific activity … consists of two basic components: Observations and interventions” (p. 421). The do-calculus has three rules: ignore an irrelevant observation, exchange an action with an observation of the same fact, ignore an irrelevant action, each licensed by the diagram (pp. 422–423). Their derivation is in chapter 3, which was not read.
- What the second edition added (p. xix): corrections in all ten original chapters, end-of-chapter summaries of new developments with annotated references, and a new chapter 11 on issues readers found “perplexing, objectionable, or in need of elaboration”.
Standing in the record
Registered on 2026-10-10 so that the record’s causal claims name a source. CLAIM-127 (version 2) separates conditioning on a communicative situation, intervening on it and intervening on what the interpreter is told. CLAIM-130 draws the experimental consequence of the last two. Both rest on the calculus only for the definition of do(); neither uses an identification result.
It was filed Deferred, unread. NOTE-696 is the skim of 2026-10-10,
of four excerpts posted on the author’s site, and it placed the work:
Active. The excerpts bear out the record’s use. Observing a value is
conditioning on it, and do() replaces the variable’s equation and keeps
the others. They also sharpen it, and each claim carries a dated note
saying how. Surgery needs the variable to have an equation of its own,
and in A129’s model, as CLAIM-127 quotes it, the framing variable
I has none. For an exogenous variable that an
experiment assigns, conditioning and intervening coincide. A fuller
reading would take §4.4 (sequential plans), §7.2.3 (“Causal Explanations,
Utterances, and Their Interpretation”) and §11.4 (on the do-operator).