Summary of "The Book of Why"

Core Idea

Causation requires three distinct levels of reasoning: association (observing patterns), intervention (doing experiments), counterfactuals (imagining alternatives)—each needs different tools; data alone cannot answer causal questions
Causal diagrams are mandatory: Draw your causal model BEFORE analyzing data; this single step determines which variables to control for and whether your analysis will succeed or fail
The do-operator is your decoder: Distinguish between observing P(Y|X) and intervening P(Y|do(X)); when they differ, confounding exists—diagrams reveal it instantly

Block all "back-door" paths (arrows pointing INTO treatment) by controlling for confounders only
Never control for mediators (blocks true causal effect) or colliders (opens spurious paths)
Three junction types determine flow: chains, forks (confounders), colliders—identify each before choosing control variables

Use front-door adjustment when confounders are unmeasured but a mediator exists with no direct confounder
Deploy instrumental variables (e.g., genes in Mendelian randomization) when you have randomized assignment to treatment proxy
Run experiments if do-calculus can't express P(Y|do(X))—observational data won't identify the effect

Three-step process: (1) Abduct hidden factors from data, (2) Apply do-operator to modify model, (3) Predict outcome
Use natural direct/indirect effects instead of regression when interactions exist; apply mediation formula explicitly
Distinguish necessary ("would not happen without X") vs. sufficient ("X alone makes outcome likely") causation

Simpson's Paradox: Aggregated vs. stratified data give opposite conclusions; causal diagrams determine which is correct
Collider bias (Monty Hall, birth-weight paradox): Conditioning on colliders creates spurious correlation—diagrams expose it
Baron-Kenny mediation: Fails when treatment-mediator interaction exists; replace with counterfactual formulas

Test external validity: Mark which variables differ between populations; use do-calculus to determine if effects transport or need recalibration
Combine studies via data fusion when individual studies measure different variables; don't assume identical results without proof
Big data limitation: Correlation patterns only suggest hypotheses; causal models still required for "what if" answers

Before analyzing: Sketch causal diagram with all relevant variables and directional arrows; label confounders, mediators, colliders explicitly
Choose adjustment method: Identify which variables to control for using back-door criterion; verify no mediators or colliders are included
Test assumptions: For instrumental variables, verify monotonicity; for front-door, confirm mediator is shielded from confounders
Validate findings: If do-calculus cannot express your causal effect from diagram, design an experiment instead of forcing observational analysis
Communicate causally: Use "do-operator" language with stakeholders—frame as interventions ("if we did X"), not just correlations