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 — three rungs that separate correlation machines from causal reasoners

Concepts

• Pearl's Ladder of Causation (importance 5): Three-level hierarchy: seeing (association), doing (intervention), imagining (counterfactuals). Each rung requires fundamentally different mathematics. Modern ML is stuck at rung one.. Source: (from training memory of book).
• Pearl's structural causal models (SCM) (importance 5): Mathematical objects combining causal graphs with functional equations. Each variable defined as a function of its parents plus noise. Foundation for counterfactual reasoning.. Source: (from training memory of book).
• Rung One: Association (seeing) (importance 4): Statistical correlation, pattern recognition, curve fitting. What current ML excels at. P(Y|X) — probability of Y given we observe X.. Source: (from training memory of book).
• Rung Two: Intervention (doing) (importance 4): Causal questions about actions. What happens if we do X? P(Y|do(X)) — probability of Y if we force X to happen. Requires causal models, not just data.. Source: (from training memory of book).
• Rung Three: Counterfactuals (imagining) (importance 4): Retrospective causal reasoning. What if I had done X instead? Why did Y happen? Requires structural causal models and the ability to rewind causal chains.. Source: (from training memory of book).
• causal identification problem (importance 4): Can we estimate causal effect from observational data given a causal graph? Do-calculus provides algorithmic answer. Not always possible.. Source: (from training memory of book).
• Pearl's counterfactual Y_x(u) (importance 4): Value that Y would take in unit u if we set X to x, keeping all else (background factors u) the same. Requires structural model.. Source: (from training memory of book).
• confounder (Pearl definition) (importance 3): Variable that affects both treatment and outcome, creating spurious association. Opens a backdoor path. Must be blocked to estimate causal effect.. Source: (from training memory of book).
• collider (Pearl definition) (importance 3): Variable caused by two others. Conditioning on a collider creates spurious association between its parents. The opposite of a confounder.. Source: (from training memory of book).
• confounding bias (importance 3): Difference between observed association and true causal effect. Arises from open backdoor paths. Zero when all confounders are measured and controlled.. Source: (from training memory of book).
• selection bias as collider conditioning (importance 3): Many forms of selection bias arise from conditioning on colliders. Recognizing the collider structure prevents spurious conclusions.. Source: (from training memory of book).
• probability of causation (PN, PS, PNS) (importance 3): Quantifying legal/attribution questions: probability that X was necessary for Y, sufficient for Y, or both. Computable from SCM even when not identifiable from data alone.. Source: (from training memory of book).
• external validity and transportability (importance 3): Can we generalize causal effect from study population to target population? Pearl's do-calculus provides conditions for valid transport across populations.. Source: (from training memory of book).
• causal discovery/learning (importance 3): Inferring causal structure from data. Harder than parameter estimation. Requires assumptions like causal sufficiency and faithfulness. Active research area.. Source: (from training memory of book).
• causal representation learning (importance 3): Learning representations that capture causal structure, not just correlations. Promises better generalization and interpretability. Active research combining deep learning with causality.. Source: (from training memory of book).
• causal modularity principle (importance 3): Each variable's mechanism is autonomous. Intervening on X doesn't change how Y's parents affect Y. Foundation of graphical surgery and transportability.. Source: (from training memory of book).
• mediator (importance 2): Variable on the causal path from treatment to outcome. Part of the mechanism. Controlling for it blocks the effect you're trying to measure.. Source: (from training memory of book).
• Rubin's conditional ignorability (importance 2): Assumption that treatment assignment is independent of potential outcomes given covariates. Corresponds to blocking all backdoor paths in Pearl's framework.. Source: (from training memory of book).
• natural direct effect (NDE) (importance 2): Effect of treatment on outcome while holding mediator at its natural value (what it would be without treatment). Rung 3 quantity.. Source: (from training memory of book).
• natural indirect effect (NIE) (importance 2): Effect transmitted through mediator. Change in outcome from changing mediator from control to treatment value while holding treatment at control. NDE + NIE = total effect.. Source: (from training memory of book).
• faithfulness assumption (importance 2): Parameters don't conspire to create conditional independencies beyond those implied by graph structure. Needed for causal discovery. Can be violated in practice.. Source: (from training memory of book).
• causal sufficiency assumption (importance 2): All common causes of measured variables are themselves measured. No unmeasured confounding. Strong assumption, often violated. Relaxed by allowing latent variables.. Source: (from training memory of book).
• Markov equivalence class (importance 2): Set of causal graphs that imply the same conditional independencies. Observational data alone often cannot distinguish between graphs in the same class. Need interventions or assumptions.. Source: (from training memory of book).
• effect modification (interaction) (importance 2): When causal effect of X on Y differs across levels of Z. Not the same as confounding. Important for personalized medicine and targeting.. Source: (from training memory of book).
• average treatment effect (ATE) (importance 2): Expected causal effect of treatment across population. E[Y|do(X=1)] - E[Y|do(X=0)]. Rung 2 quantity. Most common causal estimand.. Source: (from training memory of book).
• conditional ATE (CATE) (importance 2): Average treatment effect within subgroup defined by covariates. Used for personalization. E[Y|do(X=1),Z=z] - E[Y|do(X=0),Z=z].. Source: (from training memory of book).
• sparse mechanism shift principle (importance 2): When environment changes, only a few causal mechanisms change. Exploited for causal discovery and robust prediction. Causal models compose across contexts.. Source: (from training memory of book).

Claims

• Pearl's Causal Revolution thesis (importance 5): Statistics has been shackled by the ban on causal language since the 1930s. We now have the mathematics to answer causal questions rigorously. This changes everything.. Source: (from training memory of book).
• strong AI requires causal reasoning (importance 5): Current ML is curve fitting. Human-level intelligence requires asking why, planning interventions, learning from sparse data. All require climbing the ladder of causation.. Source: (from training memory of book).
• statistics' causal language taboo (1930s-1990s) (importance 4): Statistics textbooks taught 'correlation does not imply causation' as a commandment. Causal questions were deemed unscientific. Fisher enforced this doctrine.. Source: (from training memory of book).
• Bayesian networks → causal diagrams leap (importance 4): Pearl realized Bayesian network edges could represent causal relationships, not just conditional independence. The key insight that launched causal inference as a mathematical discipline.. Source: (from training memory of book).
• optimization ≠ causation (importance 4): Modern ML optimizes loss functions. This is curve fitting. Understanding requires causal models. You cannot optimize your way to causal knowledge.. Source: (from training memory of book).
• deep learning is stuck at rung 1 (importance 4): Neural networks are universal function approximators. But functions are rung 1 (association). To reach rung 2, you need a model of interventions. To reach rung 3, you need structural equations.. Source: (from training memory of book).
• data cannot replace causal assumptions (importance 4): No amount of observational data can tell you causal direction without assumptions. Big data does not solve causal inference. You need a model (graph or equations).. Source: (from training memory of book).
• AGI requires intervention reasoning (importance 4): To plan, an agent must predict consequences of actions. This is rung 2 reasoning. An agent that only predicts what will happen (rung 1) cannot choose actions strategically.. Source: (from training memory of book).
• AGI requires counterfactual reasoning (importance 4): To learn from mistakes, an agent must ask 'what should I have done?' To attribute credit, it must reason about alternative actions. Rung 3 is necessary for sophisticated learning.. Source: (from training memory of book).
• RCT limitations (Pearl argument) (importance 3): RCTs are often unethical, expensive, or impossible. Observational causal inference with do-calculus can answer the same questions from passive data when graph structure is known.. Source: (from training memory of book).
• Pearl-Rubin equivalence with graph transparency advantage (importance 3): Pearl argues causal diagrams make assumptions explicit and checkable, while Rubin's framework hides them in untestable ignorability conditions.. Source: (from training memory of book).
• Fisher's rejection of causal graphs (importance 3): R.A. Fisher opposed Wright's path diagrams, calling them unscientific speculation. Insisted only randomized experiments could establish causation.. Source: (from training memory of book).
• Pearl's graphical IV conditions (importance 3): An IV must: affect treatment, not affect outcome except through treatment, share no unmeasured causes with outcome. Easily verified on a causal diagram.. Source: (from training memory of book).
• causal graphs have testable implications (importance 3): d-separation implies conditional independencies that can be tested against data. Falsifiable predictions without knowing parameters. Distinguishes causal graphs from pure speculation.. Source: (from training memory of book).
• counterfactuals resolve legal causation (importance 3): Courts ask 'but-for' questions. Did defendant's action cause harm? Requires counterfactual reasoning. Pearl's framework formalizes what law has done intuitively.. Source: (from training memory of book).
• interventions identify causal direction (importance 3): Observational data cannot always determine arrow direction. Intervening breaks incoming arrows, revealing structure. This is why experiments are powerful.. Source: (from training memory of book).
• causality enables transfer learning (importance 3): Causal relationships are stable across environments; correlations are not. Agents that learn causal models can generalize to new contexts. Pure pattern matching fails.. Source: (from training memory of book).
• explanations are counterfactual (importance 3): To explain Y, we must say what would have happened without X. This is inherently counterfactual (rung 3). Black-box ML cannot explain because it cannot climb to rung 3.. Source: (from training memory of book).
• good controls vs bad controls (importance 3): Controlling for a confounder removes bias. Controlling for a mediator blocks the effect. Controlling for a collider creates bias. Graph structure determines which is which.. Source: (from training memory of book).
• interventional distributions factor on mutilated graph (importance 3): P(v|do(x)) factors according to the graph with X's incoming edges removed. Interventions preserve local mechanisms except for the intervened variable.. Source: (from training memory of book).
• attribution is a counterfactual question (importance 3): Did this specific instance of X cause this specific instance of Y? Requires unit-level counterfactual reasoning. Cannot be answered at population level (rung 2). Needs rung 3.. Source: (from training memory of book).
• causal meta-analysis (importance 2): Combining studies requires causal assumptions, not just statistical ones. Different populations may have different confounders. Causal diagrams make combination rules explicit.. Source: (from training memory of book).
• missing data as selection bias (importance 2): Missingness mechanism creates selection bias. Whether data are missing completely at random (MCAR), at random (MAR), or not at random (MNAR) corresponds to graph structure.. Source: (from training memory of book).
• free will requires counterfactual reasoning (importance 2): Choosing requires imagining consequences of different actions. If we could not reason about 'what if I do X?', we could not make deliberate choices. Rung 2 minimum for agency.. Source: (from training memory of book).
• moral responsibility requires rung 3 (importance 2): Holding someone responsible requires 'would the harm have occurred without their action?' This is counterfactual. Determinism does not eliminate responsibility if defined counterfactually.. Source: (from training memory of book).
• heterogeneous treatment effects (importance 2): Effects vary across individuals. Average effect may hide important subgroup differences. Requires rung 2 reasoning about specific groups or rung 3 for individuals.. Source: (from training memory of book).

Empirical results

• Simpson's Paradox (importance 4): Statistical phenomenon where an association reverses when data is stratified. Resolved by causal diagrams: depends on whether the stratifying variable is a confounder or mediator.. Source: (from training memory of book).
• birth weight paradox (importance 3): Smoking appears protective for low-birth-weight babies. Actually a collider bias: conditioning on birth weight creates spurious negative association.. Source: (from training memory of book).
• identification algorithm completeness (importance 3): Shpitser-Pearl 2006: complete algorithm for determining if causal effect is identifiable from graph. If algorithm fails, no formula exists. Settles the identification problem.. Source: (from training memory of book).
• Berkson's paradox (importance 2): Negative correlation between diseases in hospitalized patients even when uncorrelated in population. Hospital admission is a collider.. Source: (from training memory of book).
• Table 2 Fallacy (importance 2): Common mistake in epidemiology: controlling for everything in a regression. Often includes mediators and colliders, biasing causal estimates. Graphs prevent this.. Source: (from training memory of book).

Methods

• Pearl's do-calculus (importance 5): Three formal rules for manipulating causal expressions. Allows converting observational data into answers about interventions under certain graphical conditions.. Source: (from training memory of book).
• Causal diagrams (DAGs) (importance 5): Directed acyclic graphs encoding causal relationships. Nodes are variables, arrows are direct causal influences. The foundation of Pearl's framework.. Source: (from training memory of book).
• Pearl's do-operator (importance 5): Mathematical operator do(X=x) representing an intervention that sets X to x. The bridge from seeing to doing. Not reducible to conditioning. Requires causal model to evaluate.. Source: (from training memory of book).
• Pearl's backdoor criterion (importance 4): Graphical rule for identifying which variables to control for to estimate causal effects. Blocks all spurious paths between treatment and outcome.. Source: (from training memory of book).
• Pearl's Bayesian networks (1980s) (importance 4): Probabilistic graphical models for efficient inference. Pearl's early work on uncertainty in AI. Foundation for later causal diagrams.. Source: (from training memory of book).
• Pearl's d-separation (importance 4): Graphical criterion for conditional independence. If X and Y are d-separated given Z in the graph, they're conditionally independent in all compatible distributions.. Source: (from training memory of book).
• Pearl's frontdoor criterion (importance 3): Alternative identification strategy when backdoor fails. Uses mediating variables to estimate causal effects even with unmeasured confounders.. Source: (from training memory of book).
• randomized controlled trial (RCT) (importance 3): Gold standard for causal inference. Randomization breaks all backdoor paths by design. Implements do(X) in the real world.. Source: (from training memory of book).
• Rubin's potential outcomes framework (importance 3): Dominant statistical framework before Pearl. Focuses on what-if comparisons between treatment and control. Algebraically equivalent to Pearl's framework but less transparent about assumptions.. Source: (from training memory of book).
• Bradford Hill's causal criteria (importance 3): Nine guidelines for inferring causation from association: strength, consistency, specificity, temporality, dose-response, plausibility, coherence, experiment, analogy. Pre-Pearl heuristics.. Source: (from training memory of book).
• instrumental variable (IV) (importance 3): Variable that affects outcome only through treatment. Allows causal estimation despite unmeasured confounding. Must satisfy exclusion restriction.. Source: (from training memory of book).
• Pearl's twin network method (importance 3): Technique for answering counterfactual queries. Create parallel copy of causal graph, one for factual world and one for counterfactual intervention. Linked by shared noise terms.. Source: (from training memory of book).
• Pearl's causal mediation analysis (importance 3): Decompose total effect into direct effect and indirect effect through mediator. Answers 'how much of the effect goes through this mechanism?' Requires cross-world counterfactuals.. 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 backdoor criterion. Converts intervention to observational conditioning. First major do-calculus result.. Source: (from training memory of book).
• truncated factorization formula (importance 3): P(y|do(x)) = P(y|pa_Y) × [product over non-descendants of X]. Intervention removes edges into X, then compute probability normally. Fundamental to do-calculus.. Source: (from training memory of book).
• graph surgery (intervention as edge removal) (importance 3): do(X=x) implemented by removing all incoming edges to X and fixing its value. Mutilated graph. Shows interventions change the causal structure.. Source: (from training memory of book).
• Wright's path coefficient (importance 2): Quantitative measure of direct causal influence along a path in a diagram. Computed from correlations using structural equations.. Source: (from training memory of book).
• Rosenbaum-Rubin propensity score (importance 2): Probability of treatment given covariates. Matching or weighting by propensity score can estimate causal effects under ignorability. Pearl shows when this works graphically.. Source: (from training memory of book).
• encouragement design (importance 2): Randomize encouragement to take treatment rather than treatment itself. Encouragement serves as instrumental variable. Estimates effect of treatment on compliers.. Source: (from training memory of book).
• PC algorithm (Spirtes-Glymour-Scheines) (importance 2): Constraint-based causal discovery algorithm. Uses conditional independence tests to infer graph structure. Assumes causal sufficiency and faithfulness.. Source: (from training memory of book).
• difference-in-differences (importance 2): Quasi-experimental method comparing change over time in treatment group vs control group. Assumes parallel trends. Pearl shows graphical assumptions.. Source: (from training memory of book).
• regression discontinuity design (importance 2): Quasi-experimental method exploiting threshold in treatment assignment. Compares units just above and below threshold. Local randomization assumption.. Source: (from training memory of book).
• natural experiments (importance 2): Observational studies where nature or policy creates quasi-random assignment. Examples: draft lottery, policy changes. Approximate instrumental variables.. Source: (from training memory of book).
• sensitivity analysis for unmeasured confounding (importance 2): How much unmeasured confounding would it take to explain away the observed effect? Quantifies robustness. Pearl: easier to express graphically.. Source: (from training memory of book).
• frontdoor adjustment formula (importance 2): P(y|do(x)) = Σ_m P(m|x) Σ_x' P(y|x',m)P(x') when M is a frontdoor set. Eliminates unmeasured confounding via mediator.. Source: (from training memory of book).
• invariant causal prediction (importance 2): Method for finding causal parents of target variable by looking for predictive relationships that are stable across environments. Exploits that causation is stable, correlation is not.. Source: (from training memory of book).

Entities

• smoking-lung cancer controversy (1950s-60s) (importance 4): Tobacco industry argued correlation doesn't prove causation. Fisher proposed genetic confounders. Resolved by combining Hill's criteria with causal reasoning and RCTs on animals.. Source: (from training memory of book).
• Sewall Wright's path analysis (1920s) (importance 4): Geneticist who invented causal diagrams and path coefficients decades before statisticians. Rejected by Fisher and the statistics establishment. Pearl's intellectual ancestor.. Source: (from training memory of book).
• Berkeley admissions case (1973) (importance 3): Famous Simpson's Paradox example. Overall admission rates favored men, but within departments favored women. Department choice was a mediator, not a confounder.. Source: (from training memory of book).
• Galton's regression (1886) (importance 3): Discovery of regression to the mean. Galton interpreted it causally but couldn't distinguish causation from correlation. Birth of statistics as correlation-focused field.. Source: (from training memory of book).
• Karl Pearson's correlation coefficient (importance 3): Quantifies linear association. Pearson explicitly rejected causal interpretation, setting statistics on a century-long detour.. Source: (from training memory of book).
• Monty Hall problem (importance 2): Famous probability puzzle. Pearl uses it to illustrate distinction between seeing (P(car|opened-door)) and doing (P(car|do(switch))). Intervention changes the problem.. Source: (from training memory of book).
• firing squad example (importance 2): Pearl's favorite toy example. Captain's order causes both soldiers to fire, which causes death. Illustrates confounding, mediation, counterfactuals.. Source: (from training memory of book).
• Neyman-Rubin model (importance 2): Potential outcomes framework for causal inference. Jerzy Neyman's 1923 agricultural experiment paper, formalized by Donald Rubin in 1970s-80s. Dominant in statistics.. Source: (from training memory of book).
• Rubin Causal Model (RCM) (importance 2): Donald Rubin's formalization of potential outcomes. Emphasizes assignment mechanism and ignorability. Less transparent than Pearl's graphs about confounding structure.. Source: (from training memory of book).
• econometrics tradition (importance 2): Economics developed causal inference methods (IV, panel data, difference-in-differences) independently. Expressed in equations rather than graphs. Pearl shows graphical equivalents.. Source: (from training memory of book).

Relations

• Pearl's Ladder of Causation exemplifies Rung One: Association (seeing)
• Pearl's Ladder of Causation exemplifies Rung Two: Intervention (doing)
• Pearl's Ladder of Causation exemplifies Rung Three: Counterfactuals (imagining)
• Rung Two: Intervention (doing) builds-on Rung One: Association (seeing)
• Rung Three: Counterfactuals (imagining) builds-on Rung Two: Intervention (doing)
• Pearl's Causal Revolution thesis contradicts statistics' causal language taboo (1930s-1990s)
• Pearl's do-calculus enables Pearl's Causal Revolution thesis
• Pearl's do-calculus enables Rung Two: Intervention (doing)
• Causal diagrams (DAGs) requires Pearl's do-calculus
• Causal diagrams (DAGs) builds-on Sewall Wright's path analysis (1920s)
• Pearl's backdoor criterion exemplifies Pearl's do-calculus
• Pearl's backdoor criterion requires confounder (Pearl definition)
• Pearl's frontdoor criterion exemplifies Pearl's do-calculus
• Pearl's frontdoor criterion requires mediator
• Pearl's structural causal models (SCM) generalizes Causal diagrams (DAGs)
• Pearl's structural causal models (SCM) enables Rung Three: Counterfactuals (imagining)
• confounder (Pearl definition) enables confounding bias
• collider (Pearl definition) enables selection bias as collider conditioning
• collider (Pearl definition) contradicts confounder (Pearl definition)
• Simpson's Paradox exemplifies confounder (Pearl definition)
• Simpson's Paradox exemplifies mediator
• Berkeley admissions case (1973) exemplifies Simpson's Paradox
• randomized controlled trial (RCT) exemplifies Rung Two: Intervention (doing)
• randomized controlled trial (RCT) exemplifies Pearl's do-operator
• RCT limitations (Pearl argument) contradicts randomized controlled trial (RCT)
• RCT limitations (Pearl argument) motivates Pearl's do-calculus
• Rubin's potential outcomes framework exemplifies Rung Two: Intervention (doing)
• Pearl-Rubin equivalence with graph transparency advantage supports Rubin's potential outcomes framework
• Pearl-Rubin equivalence with graph transparency advantage supports Causal diagrams (DAGs)
• smoking-lung cancer controversy (1950s-60s) motivates Bradford Hill's causal criteria
• smoking-lung cancer controversy (1950s-60s) exemplifies confounder (Pearl definition)
• Bradford Hill's causal criteria precedes Pearl's do-calculus
• Sewall Wright's path analysis (1920s) precedes Causal diagrams (DAGs)
• Sewall Wright's path analysis (1920s) contradicts Fisher's rejection of causal graphs
• Wright's path coefficient exemplifies Sewall Wright's path analysis (1920s)
• Galton's regression (1886) precedes Karl Pearson's correlation coefficient
• Karl Pearson's correlation coefficient motivates statistics' causal language taboo (1930s-1990s)
• statistics' causal language taboo (1930s-1990s) exemplifies Fisher's rejection of causal graphs
• Fisher's rejection of causal graphs motivates randomized controlled trial (RCT)
• Pearl's Bayesian networks (1980s) precedes Causal diagrams (DAGs)
• Bayesian networks → causal diagrams leap builds-on Pearl's Bayesian networks (1980s)
• Bayesian networks → causal diagrams leap enables Pearl's Causal Revolution thesis
• causal identification problem motivates Pearl's do-calculus
• confounding bias exemplifies causal identification problem
• Rubin's conditional ignorability exemplifies Pearl's backdoor criterion
• Rosenbaum-Rubin propensity score requires Rubin's conditional ignorability
• Rosenbaum-Rubin propensity score exemplifies Pearl's backdoor criterion
• instrumental variable (IV) enables confounder (Pearl definition)
• Pearl's graphical IV conditions supports instrumental variable (IV)
• Pearl's graphical IV conditions requires Causal diagrams (DAGs)
• encouragement design exemplifies instrumental variable (IV)
• Monty Hall problem exemplifies Rung One: Association (seeing)
• Monty Hall problem exemplifies Rung Two: Intervention (doing)
• firing squad example exemplifies confounder (Pearl definition)
• firing squad example exemplifies mediator
• birth weight paradox exemplifies collider (Pearl definition)
• Berkson's paradox exemplifies collider (Pearl definition)
• selection bias as collider conditioning requires collider (Pearl definition)
• Pearl's d-separation requires Causal diagrams (DAGs)
• causal graphs have testable implications enables Pearl's d-separation
• Pearl's counterfactual Y_x(u) requires Pearl's structural causal models (SCM)
• Pearl's counterfactual Y_x(u) exemplifies Rung Three: Counterfactuals (imagining)
• Pearl's twin network method enables Pearl's counterfactual Y_x(u)
• probability of causation (PN, PS, PNS) requires Pearl's counterfactual Y_x(u)
• counterfactuals resolve legal causation motivates probability of causation (PN, PS, PNS)
• counterfactuals resolve legal causation requires Rung Three: Counterfactuals (imagining)
• Pearl's causal mediation analysis requires mediator
• Pearl's causal mediation analysis requires Rung Three: Counterfactuals (imagining)
• natural direct effect (NDE) exemplifies Pearl's causal mediation analysis
• natural indirect effect (NIE) exemplifies Pearl's causal mediation analysis
• external validity and transportability requires Pearl's do-calculus
• causal meta-analysis requires external validity and transportability
• missing data as selection bias exemplifies selection bias as collider conditioning
• strong AI requires causal reasoning requires Pearl's Ladder of Causation
• strong AI requires causal reasoning motivates Pearl's Causal Revolution thesis
• causal discovery/learning enables Causal diagrams (DAGs)
• faithfulness assumption requires causal discovery/learning
• causal sufficiency assumption requires causal discovery/learning
• PC algorithm (Spirtes-Glymour-Scheines) exemplifies causal discovery/learning
• PC algorithm (Spirtes-Glymour-Scheines) requires Pearl's d-separation
• Markov equivalence class exemplifies causal discovery/learning
• interventions identify causal direction enables Markov equivalence class
• causality enables transfer learning supports strong AI requires causal reasoning
• explanations are counterfactual requires Rung Three: Counterfactuals (imagining)
• free will requires counterfactual reasoning requires Rung Two: Intervention (doing)
• moral responsibility requires rung 3 requires Rung Three: Counterfactuals (imagining)
• Neyman-Rubin model precedes Rubin's potential outcomes framework
• Rubin Causal Model (RCM) builds-on Neyman-Rubin model
• Rubin Causal Model (RCM) motivates Pearl-Rubin equivalence with graph transparency advantage
• econometrics tradition exemplifies instrumental variable (IV)
• difference-in-differences exemplifies econometrics tradition
• regression discontinuity design exemplifies econometrics tradition
• natural experiments exemplifies instrumental variable (IV)
• sensitivity analysis for unmeasured confounding requires confounding bias
• effect modification (interaction) exemplifies heterogeneous treatment effects
• average treatment effect (ATE) exemplifies Rung Two: Intervention (doing)
• conditional ATE (CATE) generalizes average treatment effect (ATE)
• heterogeneous treatment effects motivates conditional ATE (CATE)
• good controls vs bad controls exemplifies confounder (Pearl definition)
• good controls vs bad controls exemplifies mediator
• good controls vs bad controls exemplifies collider (Pearl definition)
• Table 2 Fallacy exemplifies good controls vs bad controls
• Pearl's adjustment formula requires Pearl's backdoor criterion
• Pearl's adjustment formula exemplifies Pearl's do-calculus
• truncated factorization formula exemplifies Pearl's do-operator
• truncated factorization formula requires graph surgery (intervention as edge removal)
• frontdoor adjustment formula requires Pearl's frontdoor criterion
• identification algorithm completeness supports causal identification problem
• identification algorithm completeness supports Pearl's do-calculus
• optimization ≠ causation motivates strong AI requires causal reasoning
• deep learning is stuck at rung 1 exemplifies Rung One: Association (seeing)
• deep learning is stuck at rung 1 motivates strong AI requires causal reasoning
• data cannot replace causal assumptions requires Causal diagrams (DAGs)
• data cannot replace causal assumptions supports Pearl's Causal Revolution thesis
• AGI requires intervention reasoning requires Rung Two: Intervention (doing)
• AGI requires counterfactual reasoning requires Rung Three: Counterfactuals (imagining)
• causal representation learning builds-on causal discovery/learning
• causal representation learning contradicts deep learning is stuck at rung 1
• invariant causal prediction exemplifies causal discovery/learning
• sparse mechanism shift principle exemplifies causal modularity principle
• Pearl's do-operator enables Rung Two: Intervention (doing)
• Pearl's do-operator requires Pearl's do-calculus
• graph surgery (intervention as edge removal) exemplifies Pearl's do-operator
• interventional distributions factor on mutilated graph supports graph surgery (intervention as edge removal)
• causal modularity principle requires Pearl's structural causal models (SCM)
• causal modularity principle enables graph surgery (intervention as edge removal)
• attribution is a counterfactual question requires Rung Three: Counterfactuals (imagining)
• attribution is a counterfactual question motivates probability of causation (PN, PS, PNS)

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