Before estimating a causal effect, ask whether the data can contain it

Pearl opens with fumigants, oat yields and eelworms rather than with an abstract treatment and outcome. Farmers decide which plots receive treatment, so assignment is not random. Their decision responds to last year’s eelworm population, which the investigator does not observe. Laboratory measurements can recover populations before treatment, immediately after treatment and at the end of the season. Predators affect later worm growth and may themselves reflect the earlier infestation. The system contains mediation, confounding, measurement at several times and missing variables. It is exactly the sort of case in which a regression can be numerically impeccable and causally wrong.

The proposed diagram is not a photograph of nature. It is a disciplined statement of what the investigator is prepared to assume. Nodes name quantities; arrows name direct causal influence relative to the included variables; hollow nodes stand for unmeasured quantities; absent arrows assert that certain direct influences are not present. From this qualitative object the investigator asks a sharply prior question: do the assumptions force one answer for the total effect of fumigation on yield, or can two possible causal systems fit every observed distribution while disagreeing about the effect?

That order reverses a common workflow. It is tempting to choose an estimator, run it, then debate the coefficient. Pearl makes identification precede estimation. If the target is not a unique function of the observed distribution, an infinite sample and a perfect optimiser will not solve the problem. If it is identified, the graph can produce an estimand: an expression in ordinary observed probabilities. Only then do sampling error, model fitting, confidence intervals and finite-data robustness become the operative concerns.

The eelworm example also prevents “adjust for everything” from masquerading as a principle. Pre-treatment, post-treatment and end-of-season variables occupy different causal positions. A variable affected by treatment may be dangerous to condition on in one graph yet indispensable to a multistage identification formula in another. The role of a measurement comes from paths and the target effect, not from its column label or its predictive power.

An arrow becomes causal only with a mechanism behind it

A directed acyclic graph by itself supports a probabilistic factorisation. If the variables are ordered so that parents come before children, the joint distribution factors into one conditional distribution for each variable given its parents. D-separation reads conditional independences from paths: conditioned non-colliders block, while a collider and its descendants behave in the opposite way. These are statistical properties. Pearl adds a structural interpretation by associating every observed variable with an equation whose inputs are its parents and a background disturbance.

The functions represent autonomous mechanisms. “Autonomous” does not mean eternal, morally independent or immune to context. It means the model licenses a particular modular thought experiment: one equation can be replaced while the others are retained. Mutually independent disturbances express the claim that any relevant common cause has been represented in the graph. If the same unmeasured disturbance enters two equations, the graph must show an appropriate latent common cause or otherwise encode the dependence.

Now the distinction between seeing and setting becomes precise. Conditioning on X=x keeps the equation that selected X and restricts attention to units or worlds in which it produced x. Intervention replaces that equation by the constant X=x. Incoming arrows are cut because the usual causes no longer determine treatment. Outgoing arrows remain because downstream mechanisms still receive the imposed value. Pearl originally used a check or hat notation; the now familiar do(x) notation makes the same separation visible.

In a fully observed Markovian model, this surgery yields a truncated factorisation: multiply every ordinary local conditional except the factor for the intervened variable. This is why P(y|do(x)) is not generally P(y|x). The observational conditional carries information about causes of X; the intervention distribution deliberately destroys that information channel. Equality is a result to be earned from the graph, not a default property of notation.

Identification is agreement across every compatible hidden world

A target is identifiable when it is uniquely determined by the observed distribution and the causal assumptions. Pearl’s definition quantifies over models, not over estimators. Take any two fully specified structural systems compatible with the graph. If they agree on all observed probabilities, they must also agree on the requested intervention distribution. Only then can the target be written as a function of observed quantities.

This definition exposes why predictive fit is not a causal certificate. Two models can reproduce the same joint distribution, pass held-out prediction and disagree about what would happen under an intervention. The disagreement is not a small-sample defect. It arises because observational equivalence leaves the relevant mechanism underdetermined. A bow pattern: X directly causes Y while an unobserved common cause also connects them: is the paper’s smallest warning. Without further assumptions, any part of the association can be reallocated between direct response and hidden confounding.

Pearl contrasts this nonparametric result with linear instrumental-variable identification. If Z changes X and is independent of the disturbance affecting Y, a ratio of covariances can identify a linear coefficient under the stated linear structure. But simply adding an instrument to a bow does not identify the entire unrestricted response distribution. Functional form is doing essential work. Calling the graph “identified” without naming the target and model class erases that distinction.

Identification can also be global rather than local. One need not identify the causal effect of every edge on a route to identify the total effect from X to Y. Conversely, separately identifying several marginal responses need not identify their joint response. Graph topology can make a compound distribution ambiguous even when its components are obtainable. The object of the query: single outcome, joint outcome, conditional effect or policy: must therefore be explicit.

The back door turns deconfounding into a path criterion

A back-door path from X to Y begins with an arrow pointing into X. It carries association created by causes of treatment rather than by the causal response to treatment. A set Z is admissible for Pearl’s criterion if it contains no descendant of X and blocks every such path. Under those conditions, intervention and observation agree within levels of Z, and the intervention response is recovered by averaging those stratum-specific outcomes over the ordinary population distribution of Z.

Each term has a role. P(y|x,z) compares treated and untreated observations within a stratum that blocks spurious entry into treatment. P(z), not P(z|x), restores the pre-treatment population mixture. The formula is sometimes described as “controlling for confounders”, but the graph sharpens both words. A useful set need not contain every cause of treatment or every cause of outcome. It must block the specified noncausal paths without opening new ones.

That last qualification matters. Conditioning on a collider can create association between its causes. Conditioning on a descendant of treatment can remove part of the total effect, induce selection or demand a different multistage formula. The prohibition on descendants in the simple criterion prevents a common error; it is sufficient, not a claim that every post-treatment measurement is forever useless. Later in the paper, Pearl gives identifying diagrams in which descendants are necessary inputs to nonstandard formulas. The lesson is not a calendar slogan but a graph-and-query discipline.

The front door recovers an effect even when treatment and outcome are secretly confounded

The paper’s most surprising construction begins with an unmeasured common cause U of X and Y, so ordinary adjustment cannot close the treatment–outcome back door. An observed mediator Z can nevertheless identify the total effect if three conditions hold. Z intercepts every directed path from X to Y. There is no unblocked back-door path from X to Z. And X blocks every back-door path from Z to Y. These conditions permit two smaller causal effects to be identified and composed.

Pearl’s illustration uses smoking X, tar deposits Z and lung cancer Y, with an unobserved genotype U affecting both smoking and cancer. The identifying assumptions are demanding: all causal influence of smoking on cancer passes through measured tar; genotype does not also confound smoking and tar; and conditioning on smoking blocks the noncausal route from tar to cancer. The formula is a theorem conditional on that structure, not evidence that the structure holds.

The inner sum estimates what would happen if the mediator were set. X serves as an adjustment variable for the Z–Y relation because it blocks Z←X←U→Y. The outer sum then weights that mediator response by the mediator distribution induced by X. This two-stage route is why a descendant of treatment can be essential without being naively “controlled for” in a single regression.

Three rules turn graph separation into symbolic permission

Back-door and front-door are memorable patterns. Pearl’s larger ambition is a calculus that can manipulate arbitrary intervention expressions. The graph is modified to represent the regime named in the query: incoming arrows are removed for variables already set; outgoing arrows can be removed when testing whether an action may be exchanged for an observation. A d-separation statement in that modified graph licenses one algebraic transformation.

The rules are sound: whenever their graphical condition holds, the two expressions are equal in every structural model represented by the graph. In 1995 Pearl presented them as a reduction system and stated that completeness remained open. Later work by Huang and Valtorta and independently by Shpitser and Pearl established the completeness of do-calculus for the relevant nonparametric identification problem. That later theorem is historically important, but it should not be silently inserted into the original paper’s claims.

The front-door derivation demonstrates the calculus as work rather than incantation. First, action on X can be exchanged for observation when calculating Z, yielding P(z|do(x))=P(z|x). Second, the effect of Z on Y is obtained by adjusting for X. Third, the remaining action on X can be removed under the corresponding separation. Ordinary probability then sums over the mediator. At each step the graph supplies both a licence and a clue about which transformation might help.

Failure to find a reduction was not, in the 1995 paper alone, a universal proof of nonidentifiability. Specific nonidentifying patterns were established by constructing observationally equivalent models with different effects. The later completeness result closes that logical gap for the general problem: if the identification algorithm fails in the relevant model class, no sequence of the three rules can produce an observational formula because none exists.

What the diagrams reveal about adjustment and instruments

The paper catalogues small identifying and nonidentifying graphs to build visual judgement. Removing arrows cannot hurt nonparametric identification because it adds restrictions; adding arrows cannot rescue a failed derivation because it removes d-separation statements. Inserting measured mediators along an edge can help. Yet a confounding arc from X to a child of X on a route to Y is a particularly strong obstruction. These monotonic observations apply to the unrestricted graphical model, not automatically to specialised parametric structures.

The instrumental-variable contrast is instructive. In a linear equation, a valid instrument can reveal a coefficient through a covariance ratio. In a nonparametric model with noncompliance, treatment assignment Z may be random while received treatment X and outcome Y share hidden causes. The graph alone does not generally identify the full distribution P(y|do(x)). Bounds may be available, and additional assumptions such as monotonicity or restricted interactions may identify more specific quantities. The instrument is not weak; the requested object is richer than the assumptions support.

The paper also distinguishes local and global reasoning. In one graph, the overall X→Y effect is identifiable although a parent’s local effect on Y is not. In another, separate effects on two intermediate outcomes are identifiable while their joint post-intervention distribution is not. The examples resist a tempting engineering shortcut: certify every edge locally and multiply. Causal identification belongs to a target query in a whole graph.

The discussion is part of the paper’s result

The published record continues beyond Pearl’s article through invited discussions and a rejoinder. The exchange matters because it separates formal power from empirical warrant. Cox and Wermuth acknowledge a powerful computational scheme while emphasising how strong the absence of an unobserved common cause can be. Dawid examines whether manipulation needs the full functional and counterfactual apparatus. Freedman insists that explicit causal assumptions remain assumptions: a graph may reveal them without supplying independent validation.

Imbens and Rubin translate parts of the framework into potential outcomes and use random assignment, treatment receipt and compliance to press distinctions that a simple treatment node may hide. Robins relates the formulas to g-computation and longitudinal regimes. Rosenbaum points to design, sensitivity analysis and corroboration across different weaknesses. Shafer and Sobel question aspects of graphical and structural interpretation, including overly broad claims connecting back-door conditions and ignorability. These are not footnotes to a finished technology. They identify the bridge that must be crossed between a formal causal model and an empirical study.

Pearl’s rejoinder does not claim that graphs manufacture causes from data. It argues that causal reasoning is already present whenever investigators discuss treatment, prevention, mechanism or policy, and that a language should make those commitments precise. Diagrams, equations and counterfactuals are presented as complementary views. Hypothetical interventions can define scientific questions even when a literal physical operation is unavailable, provided the intended change and the mechanisms held invariant are sufficiently specified.

Two limitations are explicit in the article. First, many causal assumptions cannot be tested from the observational distribution at hand, although some models have empirical implications and auxiliary experiments can probe specific claims. Second, identification treats the observed distribution as known. It provides an estimand, not its sampling distribution, confidence interval or robust estimator. Parametric modelling may be useful after identification, but it should not be allowed to blur which conclusion came from graph structure and which from functional form.

A modern deployment should add selection, measurement, interference, positivity and transport questions. Who enters the dataset? Does treatment mean the same operation across sites and time? Can each relevant stratum receive each treatment? Does one unit’s treatment affect another unit’s outcome? Are measured proxies being mistaken for the latent causes they imperfectly record? Does a policy change one equation cleanly, or does it reorganise several mechanisms? Pearl’s framework makes these extensions possible precisely because it treats regimes and assumptions as objects rather than atmospheric caveats.

Place in the canon

Paper 0026 follows Charniak’s Bayesian-network tutorial because it changes the semantics of the arrows. Paper 0025 showed how a directed factorisation organises uncertain belief and observational evidence. Pearl asks what additional structural commitments allow the same graphical language to answer a different class of question: what distribution would result if a mechanism were externally changed?

Earlier canon entries supply ingredients. Bayes distinguishes prior and inverse probability; Chow and Liu approximate dependence by a tree; Pearl’s propagation work makes local evidence computation tractable; Lauritzen and Spiegelhalter compile loopy graphs into junction trees; Charniak explains the representational package. None of those observational inference results alone makes conditioning equal intervention. Paper 0026 installs the missing regime distinction and turns identification into graph-guided symbolic reduction.

The next canon entry is Pearl’s distinct 1995 work on structural models, diagrams and causal effects. The two are close in time and vocabulary but not interchangeable. This entry is governed by the *Biometrika* article, its eelworm case, back-door and front-door criteria, intervention calculus, catalogue of identifying graphs, invited discussion and rejoinder. Keeping that bibliographic boundary matters because a canon is only useful if its claims can be traced to the work named.

For intelligence and agency, the contribution is a grammar of controlled change. An intelligent system should distinguish prediction under the current regime from consequence under a proposed action. It should know when its data and assumptions determine the answer, when only bounds are possible and when a new observation or experiment is required. That is a substantive gain in reasoning. It is not evidence of consciousness, and it does not delegate the choice of intervention, target population, utility or acceptable risk.

Front-door identification laboratory

The laboratory enumerates a binary structural model with hidden U→X and U→Y confounding, observed X→Z→Y mediation and no direct X→Y arrow. It computes the complete observed distribution P(X,Z,Y), the naïve association P(Y|X), the true intervention P(Y|do(X)) from the hidden model and the front-door estimand from observations only. Three parameter profiles change confounding and mediation strength while preserving the front-door structure.

A fourth toggle deliberately violates the first front-door condition by adding a direct X→Y contribution. The old formula then targets only the mediated structure and is expected to disagree with the oracle. This is a diagnostic, not a failed program: the numerical failure demonstrates that identification belongs to stated assumptions.

The laboratory verifies an identity over its chosen finite models, not the empirical truth of any real graph. Its oracle sees U and the structural regime; the front-door calculation does not. Their agreement under the valid profiles demonstrates the theorem in concrete arithmetic. It cannot test whether an actual mediator intercepts every treatment path or whether an unmeasured treatment–mediator confounder exists.

Memory studio

What remains after the arrows are drawn

Pearl’s 1995 paper gives causal inference a productive division of labour. Subject-matter knowledge proposes mechanisms and exclusions. The diagram makes them visible. Structural semantics define what an intervention changes. D-separation tests relations in ordinary and surgically modified graphs. The calculus determines whether a target can be reduced to observed probabilities. Statistical methods estimate the resulting expression. Design, sensitivity work and external evidence challenge the assumptions on which the chain rests.

The great achievement is therefore neither a colourful DAG nor a universal recipe for “controlling confounding”. It is the conversion of a vague modal question: what would happen if we changed this?: into a mathematical query whose answer can sometimes be certified and sometimes be shown unavailable. Nonidentification is useful knowledge. It tells an investigator not to polish an impossible point estimate, and it can indicate which measurement or surrogate experiment would change the answer.

The most dangerous reading reverses that discipline. It treats a plausible graph as validated, an estimand as an estimate, a fitted estimate as a transported policy effect and a policy effect as a decision. The published discussion anticipates this danger. Explicit language does not eliminate judgement; it gives judgement a surface on which disagreement can be located. Hidden-cause assumptions, intervention meaning, invariance, compliance and design remain empirical and scientific work.

For contemporary intelligent systems, that is the enduring lesson. A system that predicts well under passive observation may fail under action because action changes the data-generating process. A system that can represent the change, prove identification conditions, report assumption failures and ask for the right experiment is more capable and more honest. Its competence comes from maintaining distinctions: seeing and doing, association and effect, identification and estimation, calculation and warrant: not from collapsing them into one score.