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

Knowledge Graph: The Book of Why: The New Science of Cause and Effect (Judea Pearl & Dana Mackenzie, 2018)

Editorial spotlight: ↑ the Ladder of Causation — Pearl's framework for transcending correlation

Concepts

• Pearl's Ladder of Causation (importance 5): Three-rung hierarchy: association (seeing), intervention (doing), counterfactuals (imagining). Each rung requires different mathematical machinery and enables different kinds of reasoning.. Source: (from training memory of book).
• Rung One: Association (seeing) (importance 4): Statistical correlation. P(Y|X). What conventional ML does. Can detect patterns but cannot answer 'what if' questions.. Source: (from training memory of book).
• Rung Two: Intervention (doing) (importance 4): Effect of deliberate action. P(Y|do(X)). Requires causal models. Enables prediction under manipulation.. Source: (from training memory of book).
• Rung Three: Counterfactuals (imagining) (importance 4): Retrospective reasoning about alternatives. 'What if I had done X instead?' Requires full structural causal model.. Source: (from training memory of book).
• Structural Causal Model (SCM) (importance 4): Mathematical framework combining directed graph with functional equations. Each variable defined as function of its parents plus noise term.. Source: (from training memory of book).
• Pearl's causal effect (importance 4): Difference between outcomes under different interventions: E[Y|do(X=1)] - E[Y|do(X=0)]. Distinct from and often different than correlation.. Source: (from training memory of book).
• causal Bayesian networks (importance 4): Bayesian networks with arrows interpreted causally. Enables intervention queries via graph surgery (removing incoming arrows to intervened variable).. Source: (from training memory of book).
• confounding variable (importance 3): Hidden common cause of both treatment and outcome. Creates spurious correlation. Classic obstacle to causal inference.. Source: (from training memory of book).
• Bayesian networks (importance 3): Probabilistic graphical models. Pearl's early work. Can represent correlations efficiently but do not distinguish causation without additional semantics.. Source: (from training memory of book).
• collider bias (importance 3): Conditioning on a common effect (collider) creates spurious association between causes. Opposite of confounding. Example: talent and beauty in Hollywood.. Source: (from training memory of book).
• mediation analysis (importance 3): Decomposing total causal effect into direct and indirect (through mediator). Requires causal assumptions about sequential ignorability.. Source: (from training memory of book).
• Pearl's direct effect (importance 3): Causal effect not mediated by intermediate variables. Requires counterfactual reasoning. Defined via do-operator on treatment and mediator.. Source: (from training memory of book).
• Pearl's indirect effect (importance 3): Causal effect transmitted through mediator. Total effect = direct + indirect. Central to understanding mechanisms.. Source: (from training memory of book).
• causal identification (importance 3): Can causal quantity be computed from observational data and causal assumptions? Do-calculus provides complete algorithm.. Source: (from training memory of book).
• Pearl's actual causation (token causality) (importance 3): Did this specific event cause this specific outcome? Requires Rung 3 counterfactual reasoning. Legal attribution of responsibility.. Source: (from training memory of book).
• Pearl's causal inference engine (importance 3): Algorithmic framework combining causal diagrams + do-calculus to answer causal queries from data. Automated causal reasoning.. Source: (from training memory of book).
• causal explanation (importance 3): True explanation requires identifying causes. Correlation-based patterns are not explanations. Explanation is Rung 2-3 concept.. Source: (from training memory of book).
• well-posed causal questions (importance 3): Formulating precise causal query is first step. What intervention? What outcome? What assumptions? Vague questions get vague answers.. Source: (from training memory of book).
• causal representation learning (importance 3): Emerging field: learn causal variables and structure from raw data. Combines deep learning with causal discovery.. Source: (from training memory of book).
• agency as causal competence (importance 3): Acting intelligently requires knowing what your actions will cause. Agency lives at Rung 2. Current AI lacks agency.. Source: (from training memory of book).
• potential outcomes Y(0), Y(1) (importance 2): Neyman-Rubin framework. Each unit has hypothetical outcomes under each treatment. Causal effect = Y(1) - Y(0). Fundamental problem: only one is observed.. Source: (from training memory of book).
• natural direct effect (NDE) (importance 2): Direct effect allowing mediator to take its natural value under control. One of Pearl's mediation estimands.. Source: (from training memory of book).
• controlled direct effect (CDE) (importance 2): Direct effect fixing mediator at specific value. Simpler than NDE but less policy-relevant.. Source: (from training memory of book).
• strong ignorability assumption (importance 2): Treatment assignment independent of potential outcomes given covariates. Central assumption in Rubin framework. Pearl shows graphical equivalent.. Source: (from training memory of book).
• external validity (transportability) (importance 2): Generalizing causal conclusions across populations. Pearl developed calculus for transportability using selection diagrams.. Source: (from training memory of book).
• nonparametric identification (importance 2): Identification without functional form assumptions. Pearl's framework is fundamentally nonparametric.. Source: (from training memory of book).
• sufficient cause (importance 2): Condition that guarantees outcome. Pearl discusses in context of actual causation and legal reasoning.. Source: (from training memory of book).
• necessary cause (importance 2): Condition without which outcome would not have occurred. Requires counterfactual reasoning.. Source: (from training memory of book).
• probability of necessity (PN) (importance 2): P(Y_x=0 = 0 | X=1, Y=1). Probability outcome would not have occurred without treatment. Legal 'but-for' causation.. Source: (from training memory of book).
• probability of sufficiency (PS) (importance 2): P(Y_x=1 = 1 | X=0, Y=0). Probability treatment would have produced outcome if administered.. Source: (from training memory of book).
• causal Markov condition (importance 2): Each variable is independent of its non-descendants given its parents. Links causal graphs to probability distributions.. Source: (from training memory of book).
• faithfulness assumption (importance 2): All independence relations in distribution arise from graph structure. Needed for causal discovery. Can fail due to parameter cancellation.. Source: (from training memory of book).
• Markov equivalence class (importance 2): Multiple causal graphs can encode same independence relations. Fundamental limit of purely observational causal discovery.. Source: (from training memory of book).
• v-structure (collider) (importance 2): Pattern X → Z ← Y identifiable from independence tests. Key to causal discovery. Conditioning on Z creates dependence.. Source: (from training memory of book).
• testable implications (importance 2): Conditional independence constraints implied by causal diagram. Can falsify model with data. Derived via d-separation.. Source: (from training memory of book).
• noncompliance in trials (importance 2): Assigned treatment ≠ received treatment. Creates endogeneity. IV methods recover causal effect. Intent-to-treat vs. per-protocol.. Source: (from training memory of book).
• bad control (over-adjustment) (importance 2): Conditioning on mediator or collider biases causal estimate. Pearl's graphical rules identify bad controls.. Source: (from training memory of book).
• missing data as causal problem (importance 2): Missingness mechanism is causal process. MAR, MCAR, MNAR. Pearl shows graphical conditions for valid inference.. Source: (from training memory of book).
• legal attribution (but-for test) (importance 2): Would harm have occurred but for defendant's action? This is counterfactual causation. Pearl formalizes via PoN.. Source: (from training memory of book).
• heterogeneous treatment effects (importance 2): Causal effects vary across individuals. Average treatment effect insufficient. Need conditional average or individual-level estimation.. Source: (from training memory of book).
• imagination = counterfactual simulation (importance 2): Human ability to imagine alternatives is Rung 3 capacity. Creativity requires counterfactual reasoning. AI needs this for innovation.. Source: (from training memory of book).
• recoverability (missing data) (importance 1): Can full-data distribution be recovered from incomplete data? Graphical criterion involving missingness indicators.. Source: (from training memory of book).

Claims

• Pearl's Causal Revolution (importance 5): Pearl's thesis: statistics was trapped in correlation for a century. His work (1990s-2000s) created a mathematical language for causation, enabling scientists to formalize and answer causal questions.. Source: (from training memory of book).
• Pearl's AI causation gap (importance 5): Current AI systems are stuck at Rung 1 (pattern recognition). True intelligence requires climbing to Rungs 2-3. This is the central obstacle to AGI.. Source: (from training memory of book).
• deep learning's causal limitations (importance 4): Neural networks excel at curve-fitting but cannot reason about interventions or counterfactuals without causal structure. Data alone is insufficient.. Source: (from training memory of book).
• causal AI (Pearl's vision) (importance 4): Future AI systems must combine learning from data with causal reasoning. Requires explicit causal models, not just correlational patterns.. Source: (from training memory of book).
• strong AI requires causation (importance 4): Pearl's central thesis on AI: systems stuck at Rung 1 cannot achieve human-level intelligence. Must build causal reasoning capability.. Source: (from training memory of book).
• RCT as gold standard (challenged) (importance 3): Randomized controlled trials traditionally seen as only path to causation. Pearl argues causal diagrams + do-calculus can achieve same goal from observational data under assumptions.. Source: (from training memory of book).
• 'correlation ≠ causation' (mantra challenged) (importance 3): Traditional mantra that correlation does not imply causation. Pearl argues this is defeatist — we CAN infer causation from correlation given right assumptions encoded in causal model.. Source: (from training memory of book).
• understanding = causal model (importance 3): Pearl argues we understand a domain when we have causal model that answers what-if questions. Correlation alone is not understanding.. Source: (from training memory of book).
• limits of data-driven science (importance 3): Big data cannot compensate for lack of causal model. No amount of correlation recovers intervention effects without assumptions.. Source: (from training memory of book).
• causal models = transparent assumptions (importance 3): Causal diagrams make assumptions explicit and criticizable. Contrast with black-box statistical models where assumptions are hidden.. Source: (from training memory of book).
• scientific method = iterative causal modeling (importance 3): Science progresses by proposing causal models, deriving testable implications, testing, refining. Pearl formalizes this loop.. Source: (from training memory of book).
• causal revolution in epidemiology (importance 3): Epidemiology was first field to adopt causal diagrams at scale. Now standard in training. Model for other disciplines.. Source: (from training memory of book).
• causal inference transforms social science (importance 3): From 'correlation studies' to rigorous causal claims. Requires discipline: state assumptions, justify identification, test implications.. Source: (from training memory of book).
• causality enables transfer learning (importance 3): Causal models generalize across contexts. Correlational models fail under distribution shift. Causation is key to robust AI.. Source: (from training memory of book).
• interpretability = causal explanation (importance 3): Explaining AI decision requires identifying causes of output. Feature importance without causation is incomplete.. Source: (from training memory of book).
• policy evaluation = causal inference (importance 3): Did this policy cause this outcome? What would have happened without it? These are causal questions requiring formal framework.. Source: (from training memory of book).
• humans = intuitive causal reasoners (importance 3): Children build causal models from early age. Humans naturally think in terms of interventions and counterfactuals. AI must match this.. Source: (from training memory of book).
• common sense = causal world model (importance 3): What AI lacks is causal understanding of how world works. Objects, agency, time. This is Pearl's recipe for common sense AI.. Source: (from training memory of book).
• Turing test requires Rung 3 (importance 3): Passing real Turing test requires answering 'why' and 'what if' questions. Current chatbots fail mini-Turing test.. Source: (from training memory of book).
• free will as Rung 2 intervention (importance 2): Pearl argues free will is the subjective experience of do-operator. Choosing is intervening in the causal network.. Source: (from training memory of book).
• economics' resistance to causal diagrams (importance 2): Economists prefer potential outcomes framework. Pearl argues diagrams are more intuitive but adoption has been slower.. Source: (from training memory of book).
• fairness requires causal reasoning (importance 2): Defining fair algorithm requires distinguishing legitimate from illegitimate causal pathways. Pure prediction cannot resolve fairness.. Source: (from training memory of book).
• moral responsibility requires counterfactuals (importance 2): Responsibility attribution is Rung 3 question: would outcome have occurred without agent's action? Legal and moral reasoning depend on this.. Source: (from training memory of book).
• personalized medicine = individual causal effects (importance 2): Precision medicine aims to estimate treatment effect for this patient. Requires Rung 3 counterfactual reasoning.. Source: (from training memory of book).
• language encodes causal knowledge (importance 2): Verbs, prepositions, tense encode causal relations. 'Because', 'if', 'would have'. Language is causal reasoning externalized.. Source: (from training memory of book).

Empirical results

• Simpson's Paradox resolution (importance 3): Reversal of association when data is aggregated vs. stratified. Pearl shows causal diagrams resolve which view is correct by identifying proper conditioning set.. Source: (from training memory of book).
• probability of causation (PoC) (importance 3): Pearl's formula for P(Y_x=1 > Y_x=0 | X=1, Y=1). Probability that treatment caused outcome in this individual. Bounds computable from data.. Source: (from training memory of book).
• Pearl's mini-Turing test (importance 3): Can AI answer 'What if' and 'Why' questions? Current systems fail. True intelligence requires causal reasoning.. Source: (from training memory of book).
• birth weight paradox (importance 2): Smoking mothers have lower infant mortality among low-birth-weight babies. Confounding by collider bias. Classic example of misleading stratification.. Source: (from training memory of book).
• Berkson's paradox (importance 2): Selection bias creates negative correlation between independent causes of selection. Hospital admission example. Instance of collider bias.. Source: (from training memory of book).
• ecological fallacy (importance 2): Aggregate-level associations can differ from individual-level. Pearl shows this as special case of Simpson's paradox.. Source: (from training memory of book).
• Lord's paradox (importance 2): Adjusting vs. not adjusting for baseline gives opposite conclusions. Resolved by specifying causal question and graph.. Source: (from training memory of book).
• M-bias (importance 2): Conditioning on descendant of collider induces bias. Named for M-shaped graph. Counterintuitive failure of 'control for everything'.. Source: (from training memory of book).
• bounds under unmeasured confounding (importance 2): Even without identification, can derive bounds on causal effects. Manski bounds, sensitivity parameters.. Source: (from training memory of book).
• Z-bias (butterfly bias) (importance 1): Adjusting for collider of confounders can amplify bias. Named for Z-shaped graph pattern.. Source: (from training memory of book).

Methods

• Pearl's causal diagrams (DAGs) (importance 5): Directed acyclic graphs encoding causal assumptions. Nodes = variables, arrows = direct causal influence. Mathematical foundation for moving beyond correlation.. Source: (from training memory of book).
• Pearl's do-calculus (importance 5): Three rules for transforming observational distributions P(Y|X) into interventional distributions P(Y|do(X)). Enables causal inference from passive observation under assumptions.. Source: (from training memory of book).
• Pearl's back-door criterion (importance 4): Graphical rule for identifying sufficient adjustment sets to block confounding. If satisfied, can estimate causal effect from observational data.. Source: (from training memory of book).
• Pearl's graph surgery (do-operator) (importance 4): To compute P(Y|do(X)), delete all incoming arrows to X in the causal graph, then compute P(Y|X) in the modified graph.. Source: (from training memory of book).
• Pearl's front-door criterion (importance 3): Alternative identification strategy using mediators when confounders are unmeasured. Leverages chain structure in causal graph.. Source: (from training memory of book).
• instrumental variable (IV) (importance 3): Variable affecting treatment but not outcome except through treatment. Enables causal inference despite unmeasured confounding. Pearl shows graphical conditions for validity.. Source: (from training memory of book).
• Pearl's adjustment formula (importance 3): P(Y|do(X)) = Σ_z P(Y|X,Z) P(Z) when Z satisfies back-door criterion. Core result connecting intervention to observation.. Source: (from training memory of book).
• causal discovery algorithms (importance 3): Learning causal structure from data. Pearl discusses PC algorithm, FCI. Challenges: Markov equivalence, hidden confounders.. Source: (from training memory of book).
• Pearl's d-separation (importance 3): Graphical criterion for conditional independence in causal diagrams. Foundation for testable implications. Complete for DAGs.. Source: (from training memory of book).
• randomization (causal interpretation) (importance 3): Random assignment breaks incoming arrows to treatment. Graph surgery by design. RCT implements do-operator.. Source: (from training memory of book).
• Wright's path analysis (importance 2): Graphical method for decomposing correlations into direct and indirect effects. Pioneering but controversial in its era.. Source: (from training memory of book).
• Rubin's propensity score matching (importance 2): Method for causal inference by balancing covariate distributions. Pearl shows when it works via causal diagrams.. Source: (from training memory of book).
• regression discontinuity design (importance 2): Causal inference from sharp cutoff in treatment assignment. Local randomization near threshold. Pearl discusses graphical interpretation.. Source: (from training memory of book).
• difference-in-differences (importance 2): Compare change over time in treated vs. control group. Removes time-invariant confounding. Assumes parallel trends.. Source: (from training memory of book).
• linear regression (causal interpretation) (importance 2): Classic statistical tool. Pearl shows when regression coefficients estimate causal effects via graphical criteria.. Source: (from training memory of book).
• Pearl's selection diagrams (importance 2): Augmented causal diagrams with selection nodes representing sample selection. Enables reasoning about selection bias and generalization.. Source: (from training memory of book).
• intent-to-treat analysis (importance 2): Analyze by assigned treatment regardless of compliance. Preserves randomization. Estimates effect of treatment offer, not treatment itself.. Source: (from training memory of book).
• causal sensitivity analysis (importance 2): Bounding causal effects under violations of assumptions. Pearl discusses partial identification under unmeasured confounding.. Source: (from training memory of book).

Entities

• smoking-cancer controversy (1950s) (importance 3): Historical debate whether smoking causes cancer or genetic factor causes both. Fisher vs. epidemiologists. Resolved by accumulating evidence but lacked formal causal framework at time.. Source: (from training memory of book).
• Sewall Wright (path analysis) (importance 3): Geneticist who invented path diagrams in 1920s. Precursor to Pearl's causal diagrams but lacked formal inference rules.. Source: (from training memory of book).
• Rubin Causal Model (potential outcomes) (importance 3): Dominant framework in statistics. Defines causation via potential outcomes Y(0), Y(1). Pearl argues causal diagrams are more fundamental and transparent.. Source: (from training memory of book).
• Bertrand Russell (causation skeptic) (importance 2): Argued 'cause' has no place in physics. Pearl positions this as wrong turn that delayed causal reasoning in science.. Source: (from training memory of book).
• Monty Hall problem (importance 2): Famous probability puzzle. Pearl uses it to illustrate importance of causal reasoning — solution depends on why Monty opened the door.. Source: (from training memory of book).
• firing squad example (importance 2): Classic thought experiment. Captain orders, riflemen fire, prisoner dies. Illustrates chains of causation and redundancy.. Source: (from training memory of book).
• tuberculosis X-ray paradox (importance 2): Historical case where X-ray treatment appeared harmful in observational data. Confounding by disease severity. Causal diagram resolves paradox.. Source: (from training memory of book).
• Vietnam draft lottery (IV example) (importance 2): Natural experiment using lottery number as instrument for military service. Classic IV application to estimate effect of service on earnings.. Source: (from training memory of book).
• Hume's causation skepticism (importance 2): Argued causation is just constant conjunction. Pearl respects Hume but argues we can go beyond correlation with causal models.. Source: (from training memory of book).
• Galileo's experimental method (importance 2): Controlled experiments to establish causation. Pearl sees this as proto-do-operator — creating interventions to see effects.. Source: (from training memory of book).
• Bayes' theorem (importance 2): Foundation of probability updating. Pearl shows it operates at Rung 1 only — cannot handle interventions without causal extension.. Source: (from training memory of book).
• hypothetical smoking gene (importance 2): Fisher's defense of tobacco industry. If gene causes both smoking and cancer, correlation is spurious. Pearl shows how causal diagram resolves.. Source: (from training memory of book).
• climate models as causal (importance 2): Climate science builds causal models to predict intervention effects (e.g., emissions reduction). Inherently Rung 2 enterprise.. Source: (from training memory of book).

Relations

• Pearl's Ladder of Causation hierarchically-contains Rung One: Association (seeing)
• Pearl's Ladder of Causation hierarchically-contains Rung Two: Intervention (doing)
• Pearl's Ladder of Causation hierarchically-contains Rung Three: Counterfactuals (imagining)
• Rung One: Association (seeing) precedes Rung Two: Intervention (doing)
• Rung Two: Intervention (doing) precedes Rung Three: Counterfactuals (imagining)
• Pearl's causal diagrams (DAGs) enables Pearl's do-calculus
• Pearl's causal diagrams (DAGs) formalizes Structural Causal Model (SCM)
• Pearl's do-calculus enables Rung Two: Intervention (doing)
• Structural Causal Model (SCM) enables Rung Three: Counterfactuals (imagining)
• Pearl's back-door criterion addresses confounding variable
• Pearl's back-door criterion enables Pearl's adjustment formula
• Pearl's front-door criterion addresses confounding variable
• Pearl's causal effect exemplifies Rung Two: Intervention (doing)
• RCT as gold standard (challenged) relies-on randomization (causal interpretation)
• Simpson's Paradox resolution exemplifies confounding variable
• Simpson's Paradox resolution resolved-by Pearl's causal diagrams (DAGs)
• smoking-cancer controversy (1950s) exemplifies confounding variable
• Sewall Wright (path analysis) developed Wright's path analysis
• Wright's path analysis precedes Pearl's causal diagrams (DAGs)
• Bayesian networks generalized-to causal Bayesian networks
• causal Bayesian networks enables Pearl's graph surgery (do-operator)
• Pearl's graph surgery (do-operator) implements Pearl's do-calculus
• Rubin Causal Model (potential outcomes) based-on potential outcomes Y(0), Y(1)
• Rubin Causal Model (potential outcomes) alternative-to Pearl's causal diagrams (DAGs)
• Pearl's Causal Revolution enabled-by Pearl's causal diagrams (DAGs)
• Pearl's Causal Revolution enabled-by Pearl's do-calculus
• 'correlation ≠ causation' (mantra challenged) challenged-by Pearl's Causal Revolution
• Monty Hall problem exemplifies Rung Two: Intervention (doing)
• birth weight paradox exemplifies collider bias
• collider bias opposite-of confounding variable
• mediation analysis decomposes-into Pearl's direct effect
• mediation analysis decomposes-into Pearl's indirect effect
• natural direct effect (NDE) type-of Pearl's direct effect
• controlled direct effect (CDE) type-of Pearl's direct effect
• Berkson's paradox exemplifies collider bias
• instrumental variable (IV) addresses confounding variable
• Vietnam draft lottery (IV example) exemplifies instrumental variable (IV)
• Pearl's adjustment formula requires Pearl's back-door criterion
• strong ignorability assumption equivalent-to Pearl's back-door criterion
• external validity (transportability) addressed-by Pearl's selection diagrams
• causal identification determined-by Pearl's do-calculus
• Pearl's actual causation (token causality) requires Rung Three: Counterfactuals (imagining)
• probability of causation (PoC) quantifies Pearl's actual causation (token causality)
• probability of necessity (PN) quantifies Pearl's actual causation (token causality)
• probability of sufficiency (PS) quantifies Pearl's actual causation (token causality)
• Pearl's AI causation gap diagnosed-by Pearl's Ladder of Causation
• deep learning's causal limitations stuck-at Rung One: Association (seeing)
• causal AI (Pearl's vision) requires Pearl's causal diagrams (DAGs)
• causal AI (Pearl's vision) addresses Pearl's AI causation gap
• Pearl's mini-Turing test requires Rung Two: Intervention (doing)
• Pearl's mini-Turing test requires Rung Three: Counterfactuals (imagining)
• free will as Rung 2 intervention interpreted-as Rung Two: Intervention (doing)
• Bayes' theorem operates-at Rung One: Association (seeing)
• Pearl's causal inference engine implements Pearl's do-calculus
• causal discovery algorithms learns Pearl's causal diagrams (DAGs)
• causal Markov condition property-of Pearl's causal diagrams (DAGs)
• faithfulness assumption required-for causal discovery algorithms
• Markov equivalence class limits causal discovery algorithms
• v-structure (collider) enables causal discovery algorithms
• Pearl's d-separation derives testable implications
• randomization (causal interpretation) implements Pearl's graph surgery (do-operator)
• noncompliance in trials handled-by instrumental variable (IV)
• intent-to-treat analysis addresses noncompliance in trials
• ecological fallacy instance-of Simpson's Paradox resolution
• Lord's paradox resolved-by Pearl's causal diagrams (DAGs)
• bad control (over-adjustment) causes collider bias
• M-bias exemplifies bad control (over-adjustment)
• legal attribution (but-for test) formalized-by probability of necessity (PN)
• causal explanation requires Rung Two: Intervention (doing)
• understanding = causal model requires Pearl's causal diagrams (DAGs)
• strong AI requires causation blocked-by Pearl's AI causation gap
• limits of data-driven science overcome-by Pearl's causal diagrams (DAGs)
• causal models = transparent assumptions provided-by Pearl's causal diagrams (DAGs)
• scientific method = iterative causal modeling formalized-by Pearl's causal diagrams (DAGs)
• causal representation learning extends causal discovery algorithms
• causality enables transfer learning enabled-by Pearl's causal diagrams (DAGs)
• fairness requires causal reasoning requires Pearl's causal diagrams (DAGs)
• interpretability = causal explanation requires causal explanation
• agency as causal competence requires Rung Two: Intervention (doing)
• moral responsibility requires counterfactuals requires Rung Three: Counterfactuals (imagining)
• personalized medicine = individual causal effects requires heterogeneous treatment effects
• policy evaluation = causal inference estimates Pearl's causal effect
• humans = intuitive causal reasoners operates-on Pearl's Ladder of Causation
• imagination = counterfactual simulation exemplifies Rung Three: Counterfactuals (imagining)
• common sense = causal world model encoded-in Pearl's causal diagrams (DAGs)
• Turing test requires Rung 3 generalizes Pearl's mini-Turing test
• Pearl's causal diagrams (DAGs) enables Pearl's d-separation
• hypothetical smoking gene relates-to smoking-cancer controversy (1950s)
• tuberculosis X-ray paradox exemplifies confounding variable
• firing squad example exemplifies Pearl's actual causation (token causality)
• Rubin's propensity score matching implements Pearl's back-door criterion
• linear regression (causal interpretation) estimates-when Pearl's causal effect
• Hume's causation skepticism source-of 'correlation ≠ causation' (mantra challenged)
• Bertrand Russell (causation skeptic) opposed Pearl's Causal Revolution
• well-posed causal questions formalized-by Pearl's do-calculus
• causal revolution in epidemiology adopted Pearl's causal diagrams (DAGs)
• climate models as causal exemplifies Rung Two: Intervention (doing)
• language encodes causal knowledge encodes Rung Three: Counterfactuals (imagining)

© 2026 LOOMUS. Curation, importance scoring, edge typing, and visualization. Source paper © respective authors, quoted under fair use.