Overview
This post reviews *The Book of Why: The New Science of Cause and Effect* (2018) by Judea Pearl & Dana Mackenzie. Pearl — inventor of Bayesian networks and Turing Award winner — argues that causal inference is not an appendix to statistics but an independent science, and that statistics crippled itself by refusing to discuss causation.
Key points
1. The Ladder of Causation
Pearl's framework has three levels:
- Level 1 — Association: "If I observe X, what about Y?" —
P(Y | X). This is where classical statistics and nearly all current machine learning live. It can only find co-occurrence patterns. - Level 2 — Intervention: "If I do X, what about Y?" —
P(Y | do(X)). Observation is not intervention: seeing a patient take medicine may indicate severity of illness (confounding); intervention means assigning treatment (as in an RCT). - Level 3 — Counterfactuals: "If X had not happened, what would Y have been?" — the basis of legal, medical, and moral reasoning.
P(Y|X): probability of Y after observing X; X may be caused by a confounder Z that also affects Y.P(Y|do(X)): probability of Y after forcing X's value, severing all causal paths into X, eliminating confounding.- What LLMs can do: learn statistical associations from massive text (Level 1), imitate human language patterns.
- What LLMs cannot do: answer counterfactuals (Level 3), predict under unseen interventions (Level 2), distinguish "observing X" from "doing X."
- A/B test alternatives: estimate intervention effects from observational data when randomization is impossible (e.g., pricing changes).
- Ad attribution: was the purchase caused by the ad or would it happen anyway? Difference-in-differences and synthetic control are standard tools.
- Medical evidence: observational data suggested hormone replacement therapy lowered heart disease risk; RCTs showed the opposite (confounder: healthier, wealthier women used HRT). Without causal reasoning, observational data can kill.
- Policy evaluation: Card & Krueger's natural experiment showed minimum wage increases did not reduce employment, contradicting naive observational correlations — reshaping labor economics.
> Pearl's key claim: all current machine learning, including LLMs, is stuck on Level 1. LLMs find associations but cannot do intervention or counterfactual reasoning — which is why they hallucinate: they have statistical correlation, not causal models.
2. The do-operator
Traditional probability does not distinguish P(Y|X) from P(Y|do(X)). Pearl insists they are fundamentally different:
Graphically, on a causal DAG, do(X) is equivalent to deleting all edges pointing into X and computing probabilities on the modified graph. RCTs realize do(X) physically, but Pearl's operator lets you estimate causal effects from observational data when the causal graph structure is known.
3. The back-door criterion
To estimate the causal effect of X on Y with confounders Z, find a set Z such that:
1. Z contains no descendants of X (does not block the X→Y path), 2. Z blocks all back-door paths from X to Y.
Then the back-door adjustment formula applies:
P(Y|do(X)) = Σ_z P(Y|X, Z=z) · P(Z=z)
This means: no RCT needed — measure all confounders, know the causal graph, and you can compute causal effects from observational data. Revolutionary for medicine, economics, and sociology where RCTs are impossible (e.g., you cannot randomize smoking).
4. Why statisticians rejected causation
Early 20th-century statistics, entangled in philosophical debates (Pearson vs. Yule), abandoned causation as "metaphysics" and spoke only of correlation. The absurd result: statistics could say "smoking and lung cancer are correlated" but not "smoking causes lung cancer." Pearl spent thirty years proving causation can be rigorously mathematized with graphs and probability — no philosophy required.
5. Implications for AI and LLMs
Pearl's verdict: LLM intelligence is the ultimate curve fitting — a statistical distribution of human language without a causal model. The next generation of AI needs a causal world model, an architecture-level breakthrough, not just bigger Transformers.
6. Practical value
7. Pearl's legacy
Pearl's framework (SCMs + do-calculus + counterfactuals) competes with Rubin's potential outcomes framework and Robins's marginal structural models, but has two unique advantages:
1. Graphical representation: causal assumptions become visual and auditable. 2. Mathematized counterfactuals: computable via structural equation models.
Pearl's ultimate vision: causal inference is the bridge that upgrades data science from curve fitting to science. After reading the book, you stop asking "are X and Y correlated?" and start asking "why — does X cause Y, does Y cause X, or does a third variable Z cause both?" That questioning is the starting point of scientific thinking.
---
Book: *The Book of Why: The New Science of Cause and Effect* (2018) Authors: Judea Pearl & Dana Mackenzie
Link shared in the original post for the books mentioned: https://b23.tv/4vCEQYn
*Tags in the original post: #CausalInference #JudeaPearl #CorrelationIsNotCausation #Counterfactuals #ClassicBooks*