Causality: Models, Reasoning, and Inference — Chapter-by-Chapter Outline

Author: Judea Pearl First published: 2000 Edition covered: Second Edition (Cambridge University Press, 2009). The second edition adds a new Chapter 11, "Reflections, Elaborations, and Discussions with Readers," which revisits all ten original chapters with eight years of post-publication results and reader correspondence. The Epilogue ("The Art and Science of Cause and Effect"), originally a 1996 public lecture, is retained from the first edition.

Central thesis

Causality cannot be reduced to probability. The tools of classical statistics — correlation, regression, conditional probability — operate entirely within what Pearl calls the "associational" layer of reasoning: they describe how variables co-vary in a fixed, passive world. But science, medicine, law, and everyday reasoning demand two higher layers: answering questions about the effects of interventions ("what happens if we do X?") and about counterfactuals ("what would have happened had we done otherwise?"). Pearl argues that these three layers form a strict hierarchy — the Ladder of Causation — and that no amount of statistical data alone can lift you from the first rung to the second or third.

The book's central project is to provide a rigorous, unified mathematical language for the upper two rungs. That language has three components: Structural Causal Models (SCMs), which are sets of functional equations encoding direct causal relationships; causal diagrams (directed acyclic graphs), which make the assumptions of an SCM visible and manipulable; and the do-calculus, a set of three inference rules that determines — from a diagram and observed data alone — whether and how the effect of an intervention or a counterfactual can be computed. The framework unifies and extends structural equation modeling (from econometrics), path analysis (from genetics and social science), probabilistic graphical models (from AI), and the potential-outcomes framework (from statistics), placing all of them in a common formal setting.

Can the effect of an action be determined from observed data, without running an experiment — and if so, how?

Chapter 1 — Introduction to Probabilities, Graphs, and Causal Models

Central question

What mathematical vocabulary is needed to represent causal knowledge, and how does it differ from the vocabulary of probability alone?

Main argument

Probability theory as the first layer. Pearl opens by grounding the reader in the basics of probability: events, random variables, joint and conditional distributions, Bayes' theorem, and the chain rule of factorization. These are the tools of pure observation — they describe what we see but not what we can do.

Graphs encode conditional independence. The chapter introduces directed acyclic graphs (DAGs) and the concept of d-separation: a graphical criterion that reads off conditional independence relations from the graph structure. Two variables X and Y are d-separated by a set Z if every path between them in the graph is "blocked" by Z (either through a non-collider in Z or a collider with no descendant in Z). The Markov condition links graph structure to probability: in a Bayesian network, every variable is conditionally independent of its non-descendants given its direct parents.

Bayesian networks. A Bayesian network is a DAG paired with conditional probability tables for each node given its parents, allowing the joint distribution over all variables to be factored compactly. The sprinkler/rain/wet-grass example illustrates how inference flows "upstream" (explaining away) and "downstream" (predicting). Crucially, a Bayesian network supports probabilistic inference — computing conditional probabilities — but does not by itself support causal inference.

Causal Bayesian networks. Pearl distinguishes an ordinary Bayesian network from a causal Bayesian network: in the latter, the arrows are interpreted as direct causal relationships, and the model supports a specific intervention operation. This distinction introduces the do(x) operator: P(Y | do(X = x)) asks not for the conditional probability of Y given that we observe X = x, but for the distribution of Y when we force X to take the value x by external manipulation. The intervention "surgically" removes all incoming arrows to X in the graph.

Functional causal models. Beyond Bayesian networks, Pearl presents Structural Causal Models (SCMs): each endogenous variable V is determined by a function of its parents and an exogenous noise term U — written V = f(parents(V), U_V). The exogenous terms are assumed jointly independent. SCMs support counterfactual reasoning (the third rung of the hierarchy) in a way that Bayesian networks alone cannot.

Causal versus statistical terminology. The chapter closes with a table contrasting causal and statistical terminology — a recurring theme throughout the book. Terms like "effect," "influence," "intervention," and "confounding" belong to the causal vocabulary; they cannot be defined or measured in purely statistical terms.

Key ideas

• D-separation is the key bridge between causal graph structure and probabilistic independence; it is the only way to test causal assumptions in observed data.
• The Markov condition allows the joint distribution to be factored as a product of local conditional probabilities, one per node.
• The do-operator formalizes interventions by a graph surgery: setting X = x removes all arrows into X and fixes its value.
• SCMs go beyond Bayesian networks by assigning deterministic functions to each variable, enabling counterfactual queries about individual units.
• Causal statements are not derivable from statistical statements alone; they require causal assumptions encoded in the graph.
• The distinction between observing X = x and doing X = x is the fundamental conceptual divide the entire book is built to formalize.

Key takeaway

Graphs, probability, and structural equations form a joint vocabulary for causal reasoning — and the do-operator marks the precise point where that vocabulary departs from classical statistics.

Chapter 2 — A Theory of Inferred Causation

Central question

Can causal structure be discovered — inferred from purely observational, non-experimental data — and if so, under what assumptions and to what extent?

Main argument

The causal discovery problem. Statistical data record associations. Multiple causal graphs can be compatible with the same joint distribution: a → b → c and a ← b → c are Markov-equivalent and produce the same conditional independencies. The question is whether, even so, some subset of causal relationships can be uniquely recovered, and what assumptions are needed.

Model preference and Occam's Razor. Pearl introduces a preference criterion: among all DAG models consistent with a given set of conditional independence relations in the data, prefer the minimal (most parsimonious) one. This causal Occam's Razor narrows the candidate set but does not uniquely identify a single graph in general — it identifies a Markov equivalence class.

Stable (faithful) distributions. The analysis depends on the Stability (or Faithfulness) assumption: that all and only the conditional independence relations that hold in the probability distribution are those entailed by d-separation in the true causal graph. Unstable distributions — where independencies arise by "accidental" cancellations of path coefficients — are treated as a measure-zero exception. Under stability, the data can pin down the graph up to its Markov equivalence class.

Recovering DAG structures: the IC algorithm. Pearl presents the Inductive Causation (IC) algorithm, which proceeds in three steps: (1) identify the skeleton (undirected graph of dependencies) by testing conditional independence; (2) orient v-structures (colliders, which are identifiable from the data because they are the only configurations that are not Markov equivalent to their reversal); (3) propagate orientation constraints. The output is a pattern — a partially directed graph representing the equivalence class.

Recovering latent structures. When unmeasured common causes (latent variables) are allowed, the equivalence class expands. Pearl develops the IC* algorithm for this case, which outputs a MAG (maximal ancestral graph) — a representation of the observable distribution that encodes which independence relations hold even with hidden variables.

Local criteria and non-temporal causation. The chapter examines whether temporal information (knowing which variable came first) is necessary. It is not: v-structures can break temporal symmetry, allowing causal directions to be inferred without observing time. Pearl introduces the concept of statistical time — a partial ordering derived from the causal structure itself — which can sometimes reverse statistical time (past can appear to depend on future in certain observational setups).

Key ideas

• Observational data alone can identify causal structure up to Markov equivalence, given the stability assumption.
• V-structures (X → Z ← Y with no direct X–Y edge) are the only configurations uniquely identifiable from conditional independence data.
• The IC and IC* algorithms provide a systematic procedure for structure recovery from independence tests.
• Hidden (latent) common causes cannot be directly observed but leave distinctive independence patterns that allow their presence to be inferred.
• Causal direction can sometimes be inferred without temporal information, purely from the topology of dependence.
• Stability/faithfulness is a strong assumption: violations (path cancellations) cannot be ruled out purely from data.

Key takeaway

Causal structure can be partially recovered from observational data — up to a Markov equivalence class — using conditional independence tests and the stability assumption, but this recovery has irreducible limits that only experimental data or prior knowledge can overcome.

Chapter 3 — Causal Diagrams and the Identification of Causal Effects

Central question

Given a causal diagram and observational data, when and how can the causal effect of one variable on another be calculated without performing a randomized experiment?

Main argument

Intervention in Markovian models. In a Markovian model (no unmeasured confounders, all error terms independent), the effect of setting X = x on Y is identified by the truncated factorization (also called the G-formula or the manipulation theorem): simply remove the term for X from the joint distribution's product factorization and condition on do(X = x). The result is P(Y = y | do(X = x)) = Σ P(Y = y | X = x, paX = pa) · P(paX = pa), summing over the parents of X.

Controlling confounding bias. In non-Markovian models, hidden common causes (shown as bidirected arcs or dashed arrows in the graph) prevent naive conditioning from recovering the interventional distribution. The chapter introduces the backdoor criterion: a set of observed variables Z satisfies the backdoor criterion relative to the pair (X, Y) if (i) Z blocks all "backdoor paths" from X to Y — paths that enter X through an arrowhead — and (ii) no member of Z is a descendant of X. When such a Z exists, the causal effect is identified by the backdoor adjustment: P(Y = y | do(X = x)) = Σ_z P(Y = y | X = x, Z = z) · P(Z = z).

The frontdoor criterion. For situations where no backdoor adjustment set exists (because the confounders are unmeasured), Pearl introduces the frontdoor criterion: if a set M of mediating variables intercepts all directed paths from X to Y and is itself unconfounded with Y given X, the causal effect can still be identified by first estimating the effect of X on M (unconfounded), then estimating the effect of M on Y (using backdoor adjustment on X), and combining the two estimates.

A calculus of intervention: the do-calculus. The chapter presents the do-calculus — three inference rules operating on probability expressions mixing do(·) and conditional notation. These rules allow systematic rewriting of interventional expressions into purely observational ones whenever the graph structure permits. The three rules govern: (1) insertion/deletion of observations, (2) action/observation exchange, and (3) insertion/deletion of actions. The do-calculus is sound and, as later proven by others, complete: every identifiable causal effect can be identified using these three rules.

Graphical tests of identifiability. Beyond specific criteria, the chapter develops graphical conditions for determining whether P(y | do(x)) is identifiable at all. The key structure that blocks identification is the bow-arc (a direct arrow from X to Y alongside a bidirected confounding arc), which creates irreducible non-identification.

Key ideas

• The truncated factorization formula identifies interventional distributions in Markovian (no hidden confounder) models.
• The backdoor criterion provides a graphical test for valid adjustment sets; any set satisfying it yields an unconfounded estimate.
• The frontdoor criterion shows that unmeasured confounding does not always block identification — a mediated path can be exploited.
• The do-calculus provides a complete algebraic system for causal identification; if the do-calculus cannot identify an effect, nothing can.
• Identifiability is a property of the graph, not the data; it is determined by the causal assumptions encoded in the diagram.
• The distinction between seeing (observing X = x) and doing (do(X = x)) is operationalized precisely through the graph surgery interpretation.

Key takeaway

Causal diagrams, the backdoor and frontdoor criteria, and the do-calculus together provide a complete, algorithmic answer to the question of when and how interventional effects can be inferred from observational data.

Chapter 4 — Actions, Plans, and Direct Effects

Central question

How are complex interventions — conditional policies, sequential plans, and direct effects — formalized and identified within the causal modeling framework?

Main argument

Conditional actions and stochastic policies. Chapter 3 treated primitive interventions: setting a variable to a fixed value. Chapter 4 extends this to more realistic policies. A conditional action is a rule of the form "set X to g(Z)" — the value assigned to X is a deterministic or stochastic function of observed covariates Z. Pearl shows that identifiability of the effect of a conditional policy reduces to identifiability of the effect of the corresponding primitive intervention P(Y | do(X = x), Z = z), and provides graphical conditions.

When is the effect of an action identifiable? The chapter develops refined graphical conditions beyond the backdoor and frontdoor criteria for identifying action effects in the presence of selection bias, partial observability, and complex policy structures. A key result is that the effect of an action is identifiable if and only if the interventional distribution P(Y | do(X)) can be expressed as a function of observational quantities — and the chapter gives graphical tests for when this holds.

The identification of plans. A plan is a sequence of conditional actions: policies that trigger sequentially as the agent acquires information. The identification of the effect of a plan (a "g-computation") requires carefully tracking which information is available at each decision point and which paths are blocked or opened by earlier interventions. Pearl links this to the G-computation formula of Robins (though not by that name), showing it follows naturally from the SCM framework.

Direct effects and their identification. Perhaps the most technically subtle part of the chapter concerns direct effects — the effect of X on Y not mediated through any of X's effect on other variables. Conventional statistics cannot define "direct effect" without specifying which mediators to hold fixed, and doing so naively creates collider bias. Pearl formalizes controlled direct effects (holding mediators fixed by intervention: do(M = m)) versus natural direct effects (holding mediators at whatever value they would have taken absent the treatment). Natural direct effects require counterfactual reasoning — they cannot always be identified from experiments alone and need both experimental and observational data combined.

Critique of evidential decision theory. Pearl uses the framework to critique Evidential Decision Theory (EDT), the view that rational agents should take the action for which the conditional probability of a good outcome is highest. EDT fails to distinguish doing from seeing: an action changes the world through a causal mechanism, not merely through the evidence it provides. Causal Decision Theory (CDT), which conditions on do(X) rather than X = x, is the correct framework, and the causal model makes the distinction precise.

Key ideas

• Conditional and stochastic policies are extensions of primitive interventions; their identifiability reduces to graphical conditions on the underlying causal diagram.
• Sequential plans interact with evolving information sets; their effects require the G-computation formula.
• Direct effects cannot be defined without causal language; the natural direct effect requires counterfactual operations beyond standard experiments.
• Evidential Decision Theory is shown to conflate correlation with causation; causal decision theory avoids the conflation.
• Identification of complex interventional effects is not monotone: adding more covariates to condition on can sometimes destroy identifiability if they are descendants of the treatment.

Key takeaway

The causal modeling framework extends naturally from single interventions to complex policies and direct-effect decompositions, but this extension requires counterfactual tools that go beyond what either statistics or ordinary experiments can supply.

Chapter 5 — Causality and Structural Models in Social Science and Economics

Central question

How does the causal modeling framework unify and clarify the tradition of structural equation modeling (SEM) in econometrics and social science, and what does graph theory add to classical SEM?

Main argument

The historical legacy of SEM. Structural equation modeling dates to Sewall Wright's path analysis (1920s genetics) and was extended by economists (Haavelmo, Koopmans, Cowles Commission) through the 1940s–1950s. SEM assigns coefficients to directed paths in a diagram, and the goal is to identify ("solve for") these coefficients from observed covariance data. But the causal interpretation of SEM had eroded by the 1970s–1980s, when social scientists began using SEMs as purely statistical fitting tools without causal commitment.

Wright's path coefficients and graphical identification. Pearl recasts Wright's graphical rules for computing covariances in terms of d-separation and graph-theoretic identification. A key result is that a path coefficient is identified from observational data if and only if there is no bidirected arc directly connecting the two nodes (no unmeasured direct confounding), and the graphical criterion generalizes to systems of equations. This makes identification testable by inspection of the diagram.

Model testing through missing links. In an SEM, every absent edge corresponds to a testable implication: the two variables it connects must be conditionally independent given some set of others. These overidentifying restrictions can be tested empirically. Pearl shows that d-separation provides a complete set of testable implications: all and only the conditional independencies implied by a DAG follow from d-separation. This means the causal assumptions encoded in a graph are partially falsifiable, and empirical data can refute but cannot fully confirm a proposed causal structure.

Graphs and identifiability in the non-parametric case. The chapter moves from linear parametric SEM to the non-parametric setting, showing that identifiability results derived graphically carry over. The main tool is the do-calculus applied to the structural equations. Key identifiability theorems (including the backdoor and frontdoor criteria) are illustrated with economic examples: simultaneous equations, demand-supply systems, and instrumental variables.

Conceptual underpinnings: what do structural equations mean? Pearl directly addresses the decades-long debate about the interpretation of structural equations. He argues that a structural equation V = f(paV, UV) is not a statistical regression equation — it is an autonomous mechanism: a claim about what V's value would be under any hypothetical manipulation of its parent variables. This autonomy interpretation explains why SEMs support intervention analysis and why regression equations — which lack this autonomy — do not.

Key ideas

• Wright's path diagrams encode both causal assumptions and testable constraints; d-separation makes both explicit.
• Structural equations are claims about autonomous mechanisms, not regression summaries of observed data.
• Absent edges in a SEM generate testable conditional independence restrictions; every missing path is an overidentifying restriction.
• Graphical identification conditions generalize the older algebraic rank-order conditions of classical econometrics.
• The linear-parametric SEM is a special case of the non-parametric SCM framework; results proved in the non-parametric case automatically apply to linear SEMs.
• Instrumental variables — variables that affect X but have no direct path to Y — are a special case of the frontdoor-like criterion; the graph makes their validity conditions precise.

Key takeaway

Causal diagrams rehabilitate structural equation models as genuine causal tools by making their assumptions explicit and testable, and by providing a non-parametric foundation that the older parametric tradition lacked.

Chapter 6 — Simpson's Paradox, Confounding, and Collapsibility

Central question

Why does confounding resist purely statistical definition, and what does the correct causal definition of confounding reveal about the limits of statistical analysis?

Main argument

Simpson's paradox: an anatomy. The chapter opens with a detailed dissection of Simpson's paradox: a situation in which an association between two variables reverses (or disappears) when a third variable is taken into account. The kidney-stone treatment example is instructive: Treatment A outperforms Treatment B for small stones and for large stones separately, yet Treatment B appears better in the aggregate — because large-stone patients (who are sicker and were preferentially given Treatment A) have lower recovery rates overall. The paradox shows that aggregation and disaggregation of statistical data can produce contradictory apparent conclusions.

The paradox is not a paradox. Pearl argues that the "paradox" is not a logical contradiction but a consequence of ignoring causal structure. The correct answer (should you prefer A or B?) cannot be determined from the numerical reversal alone — it depends on the causal role of the third variable. If the third variable (stone size) is a confounder (a common cause of both treatment and outcome), disaggregate. If it is a collider or a mediator, disaggregating is wrong. Only a causal diagram can adjudicate.

Why there is no purely statistical test for confounding. Pearl proves that confounding is not a statistical concept: no function of the observable joint distribution alone can identify whether a variable is a confounder. His argument uses the fact that the same statistical distribution can be generated by multiple causal structures, some of which involve confounding and some of which do not. This means that any "statistical test for confounding" either relies on hidden causal assumptions or will give wrong answers in some model.

How the associational criterion fails. The chapter examines various proposed statistical surrogates for confounding — marginal independence, conditional independence, collapsibility, and exchangeability — and shows precisely where each fails. A variable can satisfy exchangeability conditions yet still produce bias if the model has certain structures. Conversely, a variable can be unconfounded in the causal sense yet fail associational tests.

Stable versus incidental unbiasedness. Pearl distinguishes stable unbiasedness (the adjustment produces an unbiased estimate for all parameterizations of the causal model consistent with the graph) from incidental unbiasedness (the adjustment happens to work for the particular parameter values observed but fails under perturbation). Only the stable case is generalizable and scientifically trustworthy.

Key ideas

• Simpson's paradox is resolved not by numerical analysis alone but by asking which causal role the stratifying variable plays in the underlying causal diagram.
• Confounding is a causal concept: a confounding variable is a common cause of treatment and outcome; it cannot be defined purely statistically.
• No statistical criterion for confounding is universally valid; every such criterion relies on causal assumptions that the data cannot validate.
• Collapsibility (whether stratum-specific and aggregate estimates agree) is neither necessary nor sufficient for unconfoundedness.
• The backdoor criterion gives the correct causal definition of a sufficient adjustment set; it subsumes and corrects all the deficient statistical proposals.
• Behind every causal conclusion lies a causal assumption not visible in the data.

Key takeaway

Simpson's paradox is a symptom of a deeper truth: confounding is irreducibly causal, and no statistical criterion can replace a causal diagram in deciding which variables to control for.

Chapter 7 — The Logic of Structure-Based Counterfactuals

Central question

How can counterfactual statements ("Y would have been y had X been x, even though X was actually x′") be given a precise, operational meaning, and how do structural causal models provide this?

Main argument

Structural model semantics for counterfactuals. Pearl defines counterfactuals formally within SCMs using a three-step procedure: abduction (update the distribution over exogenous variables U given the observed evidence), action (intervene by modifying the relevant structural equation), and prediction (compute the distribution of the outcome in the modified model with the updated U). The notation Y_x(u) denotes the value that Y would have taken in unit u had X been set to x by intervention. This is a potential outcome in the sense of Rubin and Neyman, but now derived from — and given content by — the structural model.

Deterministic and probabilistic counterfactuals. In a deterministic SCM, Yx(u) is a fixed value for each unit u. In a probabilistic SCM (with distributions over U), Yx is a random variable, and queries like P(Y_x = y | X = x′, Y = y′) — the probability that Y would have been y had X been x, given that X was actually x′ and Y was actually y′ — are well-defined. These are the probability of necessity (PN), probability of sufficiency (PS), and probability of necessity and sufficiency (PNS) queries treated further in Chapter 9.

Applications and interpretation. The chapter works through applications: determining whether a defendant's action was the cause of the plaintiff's injury (legal counterfactuals), evaluating the effect of treatment on the treated (ETT), and understanding the difference between controlled and natural direct effects. These applications show that counterfactual reasoning is not merely philosophical speculation but a practical tool for answering legally and scientifically important questions.

Axiomatic characterization. Pearl provides an axiomatic characterization of structural counterfactuals, showing they satisfy three basic axioms: composition (if X = x is already the case, the intervention is trivially satisfied), effectiveness (the intervened variable equals the intended value), and reversibility (interventions can be composed in sequence). These axioms uniquely characterize the logic of structural models and distinguish them from purely probabilistic frameworks.

Structural versus similarity-based counterfactuals. Pearl compares his structural semantics to David Lewis's possible-worlds semantics, which evaluates "had X been x" by finding the closest possible world in which X = x and checking what Y is. For recursive (acyclic) models, the two accounts agree on all counterfactual truth values. But the structural account is more parsimonious: it does not require any notion of inter-world similarity metrics or possible-world ontology, and it is directly linked to the empirical mechanism that generates the data.

Structural versus probabilistic causality. The chapter contrasts the structural account with purely probabilistic theories of causality (e.g., those based on probability-raising). Probabilistic causality fails to handle overdetermination, preemption, and symmetric cases — situations where structural models handle all cases uniformly.

Key ideas

• The abduction-action-prediction (AAP) algorithm gives a mechanical procedure for evaluating any counterfactual query from an SCM.
• Potential outcomes (Y_x) are derived quantities in the SCM framework, not primitive undefined objects; the SCM provides their semantics.
• Counterfactual queries can be answered when both experimental and observational data are combined, even without full parametric knowledge of the model.
• The three axioms (composition, effectiveness, reversibility) uniquely characterize structural counterfactuals.
• Lewis's closest-world semantics is extensionally equivalent to structural counterfactuals in acyclic models, but the structural account is computationally and conceptually more tractable.
• Counterfactual reasoning is the third and highest rung of the Ladder of Causation; it requires SCMs, not just causal graphs.

Key takeaway

Structural causal models give counterfactual statements precise, operational meaning through the abduction-action-prediction algorithm, unifying the potential-outcomes and possible-worlds traditions while remaining directly linked to empirical mechanisms.

Chapter 8 — Imperfect Experiments: Bounding Effects and Counterfactuals

Central question

When a randomized experiment is contaminated by noncompliance or other imperfections, what can still be inferred about causal effects — and how tight are the resulting bounds?

Main argument

The noncompliance problem. A controlled experiment assigns subjects randomly to treatment (Z = 1) or control (Z = 0), but subjects may not comply: some assigned to treatment don't take it, some assigned to control take it anyway. The observed outcome is determined by the actual treatment received (X), not the assignment (Z). The causal effect of assignment (the intent-to-treat effect) is identifiable from the experiment, but the causal effect of actual treatment — what policymakers want — is not directly identifiable under arbitrary noncompliance.

Four response types and binary partitioning. Pearl's analysis partitions the subject population into four response types based on their potential treatment uptake: always-takers (take treatment regardless of assignment), never-takers (never take treatment regardless), compliers (comply with assignment), and defiers (do the opposite of assignment). The frequencies of these four types are not directly observed but are constrained by the experimental data. Each type has a different causal effect distribution, and the mixture determines the average causal effect.

Bounding causal effects. Pearl (following Balke and Pearl 1997) shows that the causal effect P(Y = 1 | do(X = 1)) − P(Y = 1 | do(X = 0)) is not point-identified but can be bounded using linear programming: the experimental data impose linear constraints on the response-type frequencies, and the causal effect is a linear function of those frequencies. The resulting Balke-Pearl bounds are typically much tighter than the naive bounds derived from the intent-to-treat effect alone.

Counterfactuals and legal responsibility. The chapter shows that the probability of causation queries (PN, PS, PNS) — e.g., "was the defendant's action a necessary cause of the plaintiff's injury?" — are also bounded rather than point-identified from experimental data alone, and derives the tight bounds.

A test for instruments. Pearl derives a testable implication of the instrumental variable (IV) assumption: if Z is a valid instrument (affects X but has no direct path to Y, and is independent of unmeasured confounders), then the observed frequencies must satisfy certain inequalities. This provides a falsifiability test that is typically overlooked in IV analysis.

Bayesian approaches. The chapter extends to finite-sample inference: rather than bounding, a Bayesian places a prior over response-type distributions and updates on the experimental data to obtain a posterior over causal effects. This bridges the frequentist bounding approach and the Bayesian estimation approach.

Key ideas

• Noncompliance in experiments prevents point identification of the actual treatment effect; bounds are the correct inferential object.
• The four-response-type decomposition captures all possibilities for noncompliance behavior and links experimental observables to unobservable response distributions.
• Balke-Pearl bounds (derived via linear programming) are sharp: no further tightening is possible without additional assumptions.
• The IV assumptions generate testable inequality constraints on observed probabilities, providing a falsifiability test.
• Combining experimental data (which controls Z) with observational data (which observes X without control) can tighten bounds on causal effects.
• Monotonicity (no defiers) is a common additional assumption that converts bounds into point estimates (the LATE, local average treatment effect).

Key takeaway

Even when a randomized experiment is corrupted by noncompliance, tight probabilistic bounds on causal effects can be derived algebraically, and the framework provides testable conditions for instrumental variable validity.

Chapter 9 — Probability of Causation: Interpretation and Identification

Central question

When an event Y occurred and X had previously occurred, what is the probability that X was a necessary cause of Y — that Y would not have occurred but for X — and when can this be estimated from data?

Main argument

Three notions of causation. The chapter distinguishes three counterfactual causal concepts important in law, medicine, and policy:

• Probability of Necessity (PN): P(Y_0 = 0 | X = 1, Y = 1) — the probability that Y would not have occurred had X not occurred, given that both X and Y did occur. This is the "but-for" standard in tort law.
• Probability of Sufficiency (PS): P(Y_1 = 1 | X = 0, Y = 0) — the probability that Y would have occurred had X occurred, given that neither did. This captures the general tendency of X to produce Y.
• Probability of Necessity and Sufficiency (PNS): P(Y1 = 1, Y0 = 0) — the probability that X both causes Y when present and prevents Y when absent; this is the strongest causal relationship.

Conditions of identification. These quantities are counterfactual and depend on the joint distribution of potential outcomes P(Y0, Y1) — a distribution that is never directly observable from a single dataset. The chapter establishes when they can be identified:

• If both experimental and observational data are available, PNS, PN, and PS can be point-identified under the assumption of monotonicity (no "prevention": if X = 1 can never decrease Y, then Y1 ≥ Y0 almost surely). Under monotonicity, PN = (P(Y = 1 | X = 1) − P(Y = 1 | X = 0)) / P(Y = 1 | X = 1).
• Without monotonicity, only bounds are available. The chapter derives sharp bounds from experimental data alone and from combined experimental plus observational data.

Examples and applications. The tobacco-lung cancer example: from randomized studies (if available), PN gives the probability that a specific smoker's cancer was caused by smoking rather than some other factor. This is exactly the quantity needed for legal attribution. Similarly, in vaccine trials, PS measures the probability that vaccination would have prevented a specific individual's illness.

Non-monotonic models. The chapter addresses models where prevention is possible (some units have Y1 < Y0 — the treatment actually reduces the outcome). Identification in non-monotonic models requires stronger assumptions or results only in bounds.

Key ideas

• PN, PS, and PNS are three distinct causal concepts; legal "but-for" causation corresponds to PN specifically.
• These quantities are not identifiable from observational data alone and not identifiable from experimental data alone; they require both combined.
• Under monotonicity, PN can be estimated from experimental data alone and equals the standard attributable risk fraction.
• Sharp bounds (tightest possible without additional assumptions) on PN, PS, and PNS are derivable from experimental data using linear programming.
• The framework provides a formal answer to questions of individual causal attribution, which courts implicitly ask but statistics typically cannot answer.

Key takeaway

The probability that a specific past event was a necessary or sufficient cause of another event is formally definable in the SCM framework and is identifiable — or tightly boundable — from combined experimental and observational data under mild monotonicity assumptions.

Chapter 10 — The Actual Cause

Central question

When we ask what "actually caused" a particular outcome in a specific instance — rather than what generally tends to produce outcomes — how can this be formalized to handle the hard cases of overdetermination, preemption, and redundancy?

Main argument

The insufficiency of necessary causation. The counterfactual "but-for" test (X caused Y if Y would not have occurred but for X) fails to identify actual causes in cases of causal overdetermination. Example: two fires simultaneously reach a house and burn it down; neither fire alone passes the but-for test (the house would have burned even without fire 1, because fire 2 was sufficient). Yet we want to say both fires were causes. A theory of actual causation must handle these cases.

Production, dependence, and sustenance. Pearl develops a new notion of actual causation built on three concepts:

• Production: X produces Y if X initiates or sustains Y through an active causal mechanism.
• Dependence: standard counterfactual dependence — Y depends on X if altering X alters Y.
• Sustenance: a variable's current setting sustains an outcome if it is sufficient to maintain the outcome given the rest of the model's structure. Formally, X = x sustains Y = y in a submodel M if Y = y holds in M under some perturbation of variables not in the sustaining path.

Causal beams. Pearl introduces causal beams — subgraphs of the full causal model that carry the active mechanism linking X to Y. A causal beam from X to Y is a set of path assignments consistent with the observed facts such that X = x is sufficient to guarantee Y = y along that beam. The concept is used to define sustenance computationally.

Handling preemption. Early preemption (one cause prevents the other from operating): Assassin A poisons the victim; assassin B was also en route but never arrived. A's action caused the death even though, if we imagine A hadn't acted, B would have (so the but-for test fails for some readings). Causal beams correctly identify A as the cause because A's mechanism is active in the actual world. Late preemption (both causes operate to a point, then one "takes over"): equally handled by the beam framework.

Path-switching causation. Cases where X causes Y not by directly changing Y but by switching which mechanism operates: the causal beam formalism handles these by tracking which structural equations are active.

Key ideas

• The standard counterfactual (but-for) test fails for overdetermination; actual causation requires a concept stronger than counterfactual necessity.
• Sustenance combines aspects of necessity and sufficiency at the level of specific event tokens.
• Causal beams are the formal tool that identifies which causal paths are "live" in a given scenario.
• The framework handles preemption (early and late), overdetermination, and path-switching — the three canonical hard cases.
• Probabilities of actual causation can be computed from SCMs, connecting singular-causation judgments to statistical quantities.
• The framework is intended for retrospective analysis (legal attribution, explanation) rather than prospective prediction.

Key takeaway

By grounding actual causation in the concept of sustenance and causal beams — tracking which mechanisms are active in the specific world under consideration — the structural framework resolves the longstanding puzzles of overdetermination and preemption that defeat purely counterfactual accounts.

Chapter 11 — Reflections, Elaborations, and Discussions with Readers

Edition note: Chapter 11 appears only in the second edition (2009). It was written after eight years of post-publication engagement with the research community and draws on correspondence, debates, and new results from 2000–2008.

Central question

What unresolved issues, common confusions, and new results emerged in the eight years between the first and second editions, and how do they refine the framework?

Main argument

Causal, statistical, and graphical vocabulary. The chapter opens by revisiting the fundamental distinction between causal and statistical language — a persistent source of confusion in applied work. Pearl explains why regression coefficients, partial correlations, and standardized effects are not causal quantities and why treating them as such has led to errors in epidemiology, economics, and social science.

Reversing statistical time. Pearl addresses a counterintuitive finding from Chapter 2: observational data can, in certain configurations, make it appear that "effects precede causes" statistically. He elaborates on why this does not violate physical causation and how the structural framework resolves the apparent paradox.

Estimating causal effects — elaborations. This is the longest sub-section, addressing reader questions and new results on causal effect identification:

• The intuition behind the backdoor criterion: why blocking all backdoor paths removes all spurious correlation between X and Y.
• Demystifying strong ignorability: Pearl shows that the "strong ignorability" condition used in potential-outcomes literature (Y_x ⊥ X | Z) is equivalent to the backdoor criterion when Z satisfies certain graphical conditions; the equivalence is not obvious from the definitions.
• Data vs. knowledge in covariate selection: an elaboration on when domain knowledge (encoded in the graph) is necessary for valid covariate selection and when it can be avoided.
• Understanding propensity scores: Pearl examines when balancing propensity scores (Rosenbaum and Rubin's method) is sufficient for confounding removal and when it is not — specifically, when propensity score adjustment fails even when strong ignorability holds, because of certain collider structures.
• The intuition behind do-calculus: Pearl provides a conceptual walkthrough of why the three rules of the do-calculus are complete.
• The validity of G-estimation: Robins's G-estimation procedure is shown to follow from the SCM framework as a special case of the G-computation formula.

Policy evaluation and the do-operator. The chapter addresses questions about dynamic treatment regimes (sequences of treatments adapted to evolving patient state) and shows how the do-operator handles time-varying interventions systematically.

New results on identification. Pearl previews results proved by Shpitser and Pearl (2006) establishing a complete graphical characterization of identifiability — the ID algorithm — which goes beyond the backdoor and frontdoor criteria to handle all identifiable causal effects.

Key ideas

• Strong ignorability in the potential-outcomes framework is equivalent to the graphical backdoor criterion; the two traditions define the same concept in different notation.
• Propensity score methods can fail even when strong ignorability holds if there are colliders in the graph that propensity conditioning opens.
• The G-computation formula is a consequence of the SCM framework, not an independent methodological choice.
• The complete characterization of identification (the ID algorithm) was proved after the first edition; Chapter 11 previews its significance.
• Dynamic treatment regimes (sequential interventions based on evolving state) fit naturally within the SCM framework.

Key takeaway

Chapter 11 reconciles the SCM framework with the potential-outcomes tradition by showing the two are formally equivalent for identification purposes, while clarifying why certain common methods (propensity scores, G-estimation) succeed or fail under specific graph structures.

The book's overall argument

1. Chapter 1 (Introduction to Probabilities, Graphs, and Causal Models) — establishes the three-layer Ladder of Causation (association, intervention, counterfactual) and introduces SCMs and causal diagrams as the joint language needed for the upper two layers; the do-operator marks the formal departure from pure statistics.
2. Chapter 2 (A Theory of Inferred Causation) — asks how much causal structure can be recovered from passive observation; the IC algorithm shows partial recovery is possible up to Markov equivalence under the stability assumption, setting the boundaries of observational causal discovery.
3. Chapter 3 (Causal Diagrams and the Identification of Causal Effects) — delivers the book's central technical contribution: the backdoor criterion, the frontdoor criterion, and the do-calculus, which together provide a complete algorithmic solution to the identification problem in non-experimental data.
4. Chapter 4 (Actions, Plans, and Direct Effects) — extends identification from primitive interventions to conditional policies, sequential plans, and direct effects; introduces natural direct effects, which require counterfactual tools beyond experiments.
5. Chapter 5 (Causality and Structural Models in Social Science and Economics) — applies the framework to SEMs in econometrics and social science, showing that graphical methods unify and extend the classical tradition by making assumptions explicit and testable.
6. Chapter 6 (Simpson's Paradox, Confounding, and Collapsibility) — uses Simpson's paradox to prove that confounding is an irreducibly causal concept, defying all purely statistical definitions, and shows the backdoor criterion gives the correct causal standard.
7. Chapter 7 (The Logic of Structure-Based Counterfactuals) — formalizes the third rung of the ladder via the abduction-action-prediction algorithm, showing that potential outcomes are derived, not primitive, and that structural and possible-worlds semantics agree on recursive models.
8. Chapter 8 (Imperfect Experiments: Bounding Effects and Counterfactuals) — shows that noncompliance in experiments, far from making inference impossible, admits tight algebraic bounds derivable by linear programming, and provides a falsifiability test for instrumental variables.
9. Chapter 9 (Probability of Causation: Interpretation and Identification) — formalizes legal and medical attribution questions (PN, PS, PNS) as counterfactual quantities and shows they are identifiable or tightly bounded when experimental and observational data are combined.
10. Chapter 10 (The Actual Cause) — extends the framework from population-level causal effects to token-level actual causation using sustenance and causal beams, resolving hard cases (overdetermination, preemption) that defeat purely counterfactual accounts.
11. Chapter 11 (Reflections, Elaborations, and Discussions with Readers) — reconciles the SCM framework with the potential-outcomes tradition, previews the complete ID algorithm, and clarifies when common methods (propensity scores, G-estimation) succeed or fail.
12. Epilogue (The Art and Science of Cause and Effect) — a panoramic historical survey of causality in science, philosophy, and statistics, making the case that the difficulties were conceptual rather than technical, and that the framework offered in the book resolves them.

Common misunderstandings

Misunderstanding: "Pearl's framework only applies to DAGs (no feedback loops)"

The formal results in the book are developed for acyclic (DAG) models because counterfactual identification is cleanest there. But Pearl notes that the do-operator and the do-calculus extend to cyclic graphs (models with feedback) with appropriate modifications, and his structural equation interpretation supports simultaneous equations with cycles. The DAG restriction is a modeling convenience, not a fundamental limitation of the framework.

Misunderstanding: "The do-calculus is just regression adjustment"

Regression adjustment (controlling for covariates) is one special case — the backdoor adjustment formula. But the do-calculus handles cases where no observed adjustment set exists (e.g., the frontdoor criterion, IV identification, and more complex identifiable cases) and also establishes non-identification when no algebraic derivation can work. It is a complete formal system, not a generalization of regression.

Misunderstanding: "Counterfactuals are metaphysical speculation, not scientific content"

Pearl argues that counterfactuals are the content of causal claims: when we say "aspirin cures headaches," we mean that in a counterfactual world where the headache sufferer did not take aspirin, their headache would have persisted. Counterfactuals are not philosophical extras — they are the actual meaning of the causal statements scientists, doctors, and lawyers make routinely. The structural model gives them precise, empirically grounded content.

Misunderstanding: "The potential-outcomes framework (Rubin/Neyman) is a competitor to Pearl's SCM framework"

The two frameworks are formally equivalent for identification purposes, as Chapter 11 establishes in detail. The potential-outcomes framework takes Yx as a primitive notation without specifying the mechanism that generates it; the SCM framework derives Yx from the structural equations and provides a graphical language for reasoning about it. The SCM adds the causal diagram (enabling identification analysis) and the mechanism interpretation (enabling counterfactual reasoning beyond the specific contrasts pre-specified in the potential-outcomes setup).

Misunderstanding: "The back-door criterion tells you which covariates to control for in a regression"

The backdoor criterion identifies a sufficient adjustment set — variables that, when conditioned upon, remove confounding bias. But it does not say the adjustment must be done by ordinary regression. In non-linear models, the adjustment formula involves non-linear conditional expectations. The criterion is about graphical validity of adjustment, not about the functional form of the estimation procedure.

Misunderstanding: "A large, representative observational study is as good as a randomized trial"

Pearl's framework shows exactly why this is wrong: no amount of observational data can identify the interventional distribution P(Y | do(X)) if the causal diagram contains unmeasured common causes of X and Y without a valid adjustment set or frontdoor path. The gap between observational and experimental inference is structural (determined by the graph), not a matter of sample size.

Central paradox / key insight

The central paradox the book resolves is this: causal reasoning is so fundamental to human thought that every child masters it effortlessly, yet generations of scientists and statisticians failed to give it a mathematical foundation — and many 20th-century statisticians argued that causation was simply too vague or metaphysical to be part of rigorous science.

Pearl's key insight is that the difficulty was conceptual, not mathematical. The tools of probability and statistics were designed to describe a passive world — one in which you observe but do not intervene. To reason about causes, you need at minimum to describe an active world — one in which you can intervene. The do-operator, ∑ P(Y | do(X = x)), is the minimal mathematical addition that bridges these two worlds. Once this operator is defined via graph surgery on a structural causal model, nearly all of causal inference — identification, confounding, counterfactuals, direct effects, attribution — follows from straightforward algebraic rules.

"Causal analysis requires the user to make a subjective commitment — a choice of model — before any analysis can begin. The reward for that commitment is the ability to answer questions no purely statistical analysis could address."

The paradox is that statistics, precisely because of its ambition to be assumption-free and purely data-driven, was incapable of answering the causal questions that motivate data collection in the first place. The SCM framework resolves this by making the causal assumptions explicit (in the graph) and then deriving from those explicit assumptions exactly what can and cannot be inferred from any given dataset.

Important concepts

Structural Causal Model (SCM)

A mathematical object consisting of a set of endogenous variables V, exogenous variables U (with a joint distribution P(U)), and a set of structural equations, one per endogenous variable, of the form Vi = fi(pai, Ui), where pai are the direct causes of Vi. SCMs support all three levels of the Ladder of Causation.

Causal diagram (DAG)

A directed acyclic graph in which nodes represent variables and directed edges represent direct causal relationships. Each missing edge is a causal assumption (no direct effect); each present edge is an assumption of potential influence. Bidirected arcs (↔) represent unmeasured common causes.

d-separation

A graphical criterion for reading conditional independence relations off a DAG. A set Z d-separates X from Y if Z blocks every path between X and Y, where blocking means: for non-collider nodes on the path, the node is in Z; for collider nodes on the path, neither the node nor any of its descendants is in Z.

Do-operator, P(Y | do(X = x))

The interventional distribution: the distribution of Y when X is set to x by external manipulation (as opposed to observed to be x). Computed via graph surgery — removing all incoming arrows to X, setting X = x, and computing the resulting distribution.

Backdoor criterion

A set Z satisfies the backdoor criterion for estimating the effect of X on Y if: (1) Z blocks all paths from X to Y that have an arrow into X ("backdoor paths"), and (2) no element of Z is a descendant of X. When satisfied, P(Y = y | do(X = x)) = Σ_z P(Y = y | X = x, Z = z) P(Z = z).

Frontdoor criterion

A set M of mediating variables satisfies the frontdoor criterion for (X, Y) if: M intercepts all directed paths from X to Y; there are no unblocked backdoor paths from X to M; and all backdoor paths from M to Y are blocked by X. When satisfied, the causal effect is identified even with unmeasured X–Y confounding.

Do-calculus

Three inference rules for transforming expressions involving P(· | do(·)) into expressions involving only observational probabilities P(· | ·), valid whenever certain d-separation conditions hold in modified versions of the causal graph. The do-calculus is complete: every identifiable causal effect can be computed using these rules.

Ladder of Causation

Pearl's three-level hierarchy of causal reasoning: (1) Association — P(Y | X), asking "what do I see if X is observed?"; (2) Intervention — P(Y | do(X)), asking "what happens if I do X?"; (3) Counterfactuals — P(Y_x | X = x′), asking "what would have happened had I done X differently?". Each level strictly subsumes the one below.

Markov condition

In a causal Bayesian network, every variable is conditionally independent of its non-descendants given its direct parents. The Markov condition links the graph structure to probabilistic independence and allows the joint distribution to be factored as a product of local conditionals.

Faithfulness (stability)

The assumption that all conditional independence relations in the data arise from d-separation in the true causal graph — not from accidental cancellations of structural parameters. Required for causal discovery algorithms (IC, IC*) to recover the correct Markov equivalence class.

Confounding

A bias that arises when treatment and outcome share a common cause (measured or unmeasured). Confounding is a causal, not statistical, concept; it can only be defined and corrected using a causal diagram, not from data alone.

Potential outcomes, Y_x

The value that the outcome Y would take for a specific unit if X were set to x by intervention. In the SCM framework, Yx(u) = Y(Mx, u), where M_x is the modified structural model with X fixed to x and u is the specific value of the exogenous variables for that unit.

Probability of Necessity (PN)

PN = P(Y_0 = 0 | X = 1, Y = 1) — the probability that Y would not have occurred had X not occurred, given that both actually occurred. The formal counterpart of legal "but-for" causation.

Probability of Sufficiency (PS)

PS = P(Y_1 = 1 | X = 0, Y = 0) — the probability that Y would have occurred had X occurred, given that neither actually occurred. Captures the tendency of X to produce Y as a type-level claim.

Causal beams

Subgraphs of the full causal model that carry an active mechanism from X to Y in the specific scenario under consideration. Used to define sustenance-based actual causation, handling overdetermination and preemption.

Sustenance

A causal concept used in the definition of actual causation: X = x sustains Y = y if X's current value is sufficient to maintain Y = y through an active causal mechanism, even under certain hypothetical perturbations of other variables.

Simpson's paradox

A statistical phenomenon in which an association between two variables reverses when a third variable is controlled for. In the SCM framework, the paradox is resolved by asking whether the third variable is a confounder (in which case disaggregate) or a mediator or collider (in which case do not disaggregate). The paradox has no resolution in purely statistical terms.

References and Web Links

Primary book and edition information

• Pearl, Judea. Causality: Models, Reasoning, and Inference. Cambridge University Press, 2nd edition, 2009. ISBN 978-0-521-89560-6.

  • Cambridge University Press book page
  • Amazon listing (2nd edition)
  • First edition (2000) on Internet Archive

• Author's official book page with table of contents (first edition / pre-2009)

Background and overview

• Wikipedia: Causality (book)
• Pearl's Causal Hierarchy — Emergent Mind overview
• LessWrong: Causality — a chapter by chapter review

The do-calculus and intervention

• Pearl, Judea. "A Probabilistic Calculus of Actions." UAI-1994.
  • Preprint at UCLA
• Causal inference overview by Pearl (2009) — Princeton course archive

Causal discovery (Chapter 2 background)

• Pearls of Causality #8: Inferred Causation — blog post with IC algorithm walkthrough
• Pearl and Verma (1991): On Pearl's Hierarchy — technical report

Backdoor, frontdoor, and identification (Chapter 3 background)

• Causal diagrams for empirical research (Pearl 1995, R218-B)
• Introduction to Causal Inference — PMC/NIH overview

Simpson's paradox (Chapter 6 background)

• Pearl, Judea. "Understanding Simpson's Paradox." UCLA Technical Report.
  • PDF at UCLA
• Stanford Encyclopedia of Philosophy: Simpson's Paradox

Counterfactuals (Chapter 7 background)

• Pearl, Judea. "Structural Counterfactuals: A Brief Introduction." Cognitive Science, 2013.
  • Wiley Online Library
• Causal and Counterfactual Inference (Pearl, R485)

Probability of causation (Chapter 9 background)

• Pearl, Judea. "Probabilities of Causation: Three Counterfactual Interpretations and Their Identification." Synthese, 1999.
  • Reprint PDF
• David Salazar's blog: Causality — Probabilities of Causation

Epilogue source

• Pearl, Judea. "The Art and Science of Cause and Effect." UCLA Faculty Research Lecture, 1996.
  • PDF at UCLA

Additional study resources

These are secondary summaries and should be used alongside, rather than instead of, the original book.

• Fabian Dablander: An Introduction to Causal Inference (blog)
• PhilPapers entry for Pearl's Causality
• ACM Digital Library entry