The question before the estimate
A causal claim needs a language in which action can be expressed
Statistics is exceptionally good at describing a distribution once the variables and sampling process have been specified. It can tell us how often an outcome appears among exposed and unexposed people, how one measurement predicts another, and how uncertainty changes after evidence arrives. Yet the question that animates policy, medicine and engineering is usually different: what distribution would result if an agent changed the process that generates one variable? The word if hides a new regime. Conditioning alone does not name it.
Pearl’s 1995 proceedings paper opens by distinguishing two kinds of conditioning. Ordinary Bayesian conditioning, written here as P(y | X=x), updates a state of knowledge after learning that X happened to equal x. Causal conditioning, now conventionally written P(y | do(X=x)), describes a distribution under an external operation that fixes X. The notation is compact, but the conceptual difference is large. One filters worlds compatible with an observation; the other changes the world-producing machinery and asks what follows.
This difference makes the paper more than a catalogue of adjustment formulas. Its central contribution is a contract between equations, diagrams and interventions. A structural equation says how a variable responds to its direct causes and a disturbance. A directed graph records which mechanisms feed which others. An intervention replaces one mechanism and preserves the remainder. Separation in appropriately modified graphs then grants permission to add, remove or exchange observations and interventions. The resulting calculus is an algorithmic language for causal reasoning under explicit assumptions.
The paper organizes that language around practical research tasks. It asks which observed covariates are sufficient to control confounding, how to derive an estimand from observational data, which measurements could make randomization unnecessary, when an experiment on a surrogate can answer a question about a different variable, and what can be learned when treatment assignment meets imperfect compliance. The ordering matters. One first settles whether the target is logically determined by available distributions; only then does one choose an estimator and confront finite samples.
Probe 1 · Translate the verb
Take the sentence “raise the threshold to 80.” Is that an observation of units whose threshold already equals 80, a deterministic intervention that fixes every unit to 80, or a policy that changes the probability of thresholds conditional on context? Write the operator before calculating. If the intervention cannot be stated, the target is not yet defined.
Seeing carries information about causes; setting deliberately erases it
Suppose a hidden severity variable U influences whether a patient receives treatment X and also influences recovery Y. Among naturally treated patients, X=1 conveys information about U. Even if the treatment helps every severity stratum, the treated group may recover less often because it contains harder cases. The observational quantity P(Y=1 | X=1) mixes the treatment response with the selection process. The interventional quantity P(Y=1 | do(X=1)) forces treatment without using severity to choose it. That operation destroys the diagnostic information carried by the natural treatment decision.
This is why replacing causal language with a large regression is not a semantic repair. More flexible prediction estimates the observed distribution with less approximation; it does not decide which parts of that distribution belong to the assignment mechanism and which belong to the outcome mechanism. Even infinite observational data cannot distinguish causal models that produce exactly the same observed joint distribution but respond differently when X is set. Data solve sampling uncertainty. Identification assumptions solve equivalence among hidden causal worlds.
Pearl’s diagrams make the distinction operational. In a recursive structural model, each observed variable is generated by an equation from its parents and a disturbance. To represent do(X=x), delete the equation for X and substitute the constant x. In the graph, remove arrows entering X. The remaining equations retain their original form. This is sometimes called graph surgery, but the surgical metaphor should not obscure the strong premise: mechanisms not targeted by the action are invariant across regimes.
Worked example 1 · Treatment assigned by severity
Let U be equally likely to be 0 or 1. A severe patient (U=1) receives treatment with probability 0.9; a mild patient receives it with probability 0.1. Suppose treatment raises recovery probability by 0.20 at either severity, from 0.70 to 0.90 for mild patients and from 0.10 to 0.30 for severe patients. Under intervention, averaging over severity gives recovery probabilities 0.40 without treatment and 0.60 with treatment: a benefit of 0.20. Under observation, treated cases are mostly severe, so their recovery rate can be lower than that of untreated cases. The sign reversal is not a paradox. The two operators answer different questions.
A structural coefficient is a response to controlled change, not a correlation
Section 6 of the proceedings paper gives structural equations an operational reading. Consider Y=bX+εᵧ. Read merely as a regression, b might be chosen to summarize the conditional mean of Y given X. Read structurally, b says how the expected value of Y changes when X is externally varied while the mechanism represented by εᵧ remains in place. The disturbances may be correlated with causes elsewhere in the system. What gives b causal meaning is not an orthogonality convention but the equation’s claimed stability under the relevant intervention.
This distinction also changes how an omitted variable is interpreted. If Z does not appear in the structural equation for Y, the claim is not necessarily that Y and Z are conditionally independent in observational data. Paths through other variables or correlated disturbances can preserve association. The structural omission says that, once the variables appearing on the right-hand side are externally held fixed, manipulating Z has no further influence on Y. It is a claim about controlled response, not merely about a zero partial correlation.
Likewise, a disturbance is not just unexplained residual noise. It gathers background factors relevant to a mechanism at the chosen level of description. Correlation between disturbances can represent latent common causes. Different levels of aggregation can produce different disturbances and different invariance claims. A structural model is therefore an empirical theory about modules: what each variable responds to, what unmeasured factors are shared, and which equations survive a specified family of actions.
Probe 2 · What does the missing variable mean?
You are shown Y=bX+ε and told that Z is absent. Ask whether the author means “Z has zero regression coefficient in this dataset” or “holding X fixed by intervention, changing Z has no effect on Y.” Those statements can diverge. Name an observational path that preserves association even when the second claim is true.
Worked example 2 · A policy changes selection, not response
A hospital adopts a triage rule that treats high-risk patients more aggressively. The biological response equation for recovery is unchanged, but the slope from a naïve regression of recovery on treatment moves because the treated population changes. If a coefficient is meant structurally, it should track the stable controlled response, not the moving observational contrast. The laboratory later on implements this exact test by changing only P(X|U) and checking that P(Y|do(X)) stays fixed.
The graph is an assumption ledger and a search device
A directed acyclic graph gives a qualitative map of the factorization induced by the structural model. Directed arrows encode direct functional dependence at the selected level of description. Bidirected or latent connections can represent dependent disturbances. More importantly, the graph supports separation tests in modified versions of itself. Those tests turn causal questions into graphical obligations that can be inspected before numbers enter.
The back-door criterion is the best-known example. A set Z is admissible for estimating the effect of X on Y if no member of Z descends from X and Z blocks every path between X and Y that enters X through an incoming arrow. Under those conditions, stratifying on Z removes the spurious association carried into treatment and the effect is recovered by averaging conditional outcome distributions over the ordinary distribution of Z.
The formula is not a ritual instruction to control every available covariate. Conditioning on a collider can open a path that was closed; conditioning on a descendant of treatment can remove part of the effect or create selection. The diagram tells the analyst which paths matter and why. Measurement strategy becomes a design problem: find a sufficient blocking set that is temporally and substantively credible, then ensure it is measured with adequate support and accuracy.
The front-door criterion is more surprising. Suppose hidden causes confound treatment X and outcome Y, so no measured back-door set exists. A measured mediator M can nevertheless identify the total effect if it intercepts every directed path from X to Y, the X→M relation is unconfounded, and all back-door paths from M to Y are blocked by X. Identification then proceeds in two stages: learn how treatment changes the mediator, and learn how the mediator changes the outcome after controlling the mediator’s back doors.
Worked example 3 · The mediator does not “control away” the effect
Suppose smoking X changes tar deposition M, which changes disease Y, while an unmeasured disposition U affects both smoking and disease. Simply adding M to an outcome regression would block the pathway one wants to measure. The front-door derivation instead estimates P(m|x), identifies how M changes Y after conditioning on X to block M←X↔Y, and averages the two pieces. The mediator is used as a bridge between two identifiable causal relations, not as a generic adjustment variable.
Three rules make causal algebra conditional on graph surgery
Pearl’s calculus of intervention is a set of transformation rules. Each rule pairs an algebraic move with a d-separation statement in a graph modified to represent the relevant intervention or observation. In contemporary notation the rules allow, under their stated graphical conditions, deletion or insertion of observations, exchange between intervention and observation, and deletion or insertion of interventions. The details of which incoming or outgoing edges are removed matter; a verbal appeal to “do-calculus” is not a derivation.
Rule 1 · Observe
Add or remove an observation Z when Y is separated from Z given the remaining information in the graph where incoming edges to the intervention variables have been removed.
Rule 2 · Exchange
Replace an action on Z with an observation of Z when the modified graph blocks the paths that would make selection into Z differ from setting it.
Rule 3 · Act
Add or remove an action on Z when, after the specified surgery, Z has no remaining route to Y given the other variables.
The distinction between derivation and estimation is critical. A successful derivation ends in a functional of distributions available from observation or experiment. It says that, if the model is right and the needed probabilities exist, every compatible causal model agrees on the target. Estimation asks how to learn that functional from finite, noisy and possibly selected samples. Flexible models, regularization, uncertainty intervals and diagnostics enter at this later stage. An estimator cannot rescue a target for which no identifying functional exists.
The front-door formula is a compact demonstration. Begin with a query containing do(x). Expand over the mediator. Use graphical conditions to replace the effect of setting X on M with the observed P(m|x). Transform the effect of setting M on Y by controlling the back-door path through X. Sum over the ordinary distribution of X. At each line a graphical separation licenses a change. The finished expression contains only observational quantities, but its causal meaning comes from the entire chain of permissions.
Probe 3 · Audit a one-line formula
Someone presents Σ_z P(y|x,z)P(z) as “the causal effect after adjustment.” Require them to provide the graph, the target intervention, the separation statement and the reason Z contains no treatment descendants. Without that ledger the formula is a candidate, not an identified estimand.
The calculus can redesign the study instead of merely analyzing its data
One of the paper’s most productive shifts is from retrospective adjustment to prospective design. If the desired intervention cannot be randomized, perhaps a different variable can. Suppose cholesterol X is the target cause of heart disease Y, while diet Z is manipulable. An experiment on diet does not automatically answer the cholesterol question. It becomes useful only under structural restrictions: the effect of diet on disease must operate through cholesterol in the relevant model, and the remaining relations must allow the target to be identified when experimental information about Z is combined with observation.
The diagram therefore helps select surrogate experiments. It asks which mechanism an experiment should disturb so that the unavailable intervention can be reconstructed. This is a sharper question than searching for a correlated proxy. A surrogate is valuable because a derivation links its experimental distribution to the target, not because it predicts the target variable well.
Worked example 4 · Randomize diet, ask about cholesterol
Imagine cholesterol cannot ethically or practically be assigned directly, but diets can be randomized. Collect P(x|do(z)) from the trial and observational information relating X, Z and Y. If diet has a direct effect on disease not mediated by cholesterol, or if hidden paths violate the required separations, the proposed bridge fails. The graph states exactly which substantive facts must be defended before the surrogate experiment answers the cholesterol intervention.
Imperfect compliance reveals another boundary. A randomized encouragement Z may influence received treatment X, while hidden factors affect both treatment and outcome Y. In linear models, an instrumental-variable ratio can identify a structural coefficient under suitable exclusions and relevance. Pearl stresses that the analogous fully nonparametric model does not generally identify the entire causal effect distribution. Randomization of the encouragement does not equal randomization of treatment. Bounds or additional assumptions may still be available, but the full target is not supplied by random assignment alone.
Probe 4 · What exactly was randomized?
In an encouragement trial, replace every occurrence of “randomized treatment” with the actual assigned variable. Draw noncompliance explicitly. Then ask whether the claimed estimand is the effect of assignment, a local effect among a defined compliance type, a structural coefficient, or the population effect of treatment. The randomization validates only claims linked to its true target and additional assumptions.
Identification can be tested graphically, but the model still faces the world
The paper describes graphical tests and algorithms intended to decide whether causal effects are identifiable. For singleton outcomes in the model class considered, graphical reduction searches for a sequence of transformations that removes intervention operators. The broader question: whether the three rules are complete for every relevant identification problem: was explicitly left open in 1995. Later work established completeness results and systematic identification algorithms. Historical precision matters: the paper introduced a sound calculus and significant algorithms; it did not claim every later theorem.
Even a complete identification algorithm is complete relative to a formal model class, not relative to nature. The graph can be wrong. A supposedly invariant mechanism can change when the policy changes. Variables can be measured after treatment, selected into the sample, coarsened or defined inconsistently across regimes. Interference can make one unit’s outcome depend on another unit’s treatment. Feedback can violate the acyclic simplification. Positivity can fail where the policy asks for actions never taken in some strata.
Pearl sketches extensions to stochastic policies and cyclic systems. A policy may set a variable according to a response function rather than a fixed constant. Feedback systems require equilibrium or dynamic semantics; ordinary acyclic d-separation and one-pass substitution are no longer sufficient without additional stability assumptions. These extensions reinforce the core lesson: an intervention is a specified modification of mechanisms. “Change X” is incomplete until the policy, timing, affected module and preserved modules are defined.
Historical and empirical boundary
In 1995 the calculus was sound and its general completeness remained open; later results supplied completeness for major identification settings. Neither result validates a proposed graph. The analyst still owes evidence for temporal order, exclusion restrictions, absence or handling of hidden paths, intervention versions, consistency, positivity, sampling and transport. A derivation can be flawless and its empirical answer irrelevant if the intervention described by the model is not the intervention actually contemplated.
Why this companion follows the causal-diagrams paper
Canon 0026 was governed by Pearl’s Biometrika article “Causal diagrams for empirical research,” including its invited discussion and rejoinder. This canon slot was supplied with a different title and DOI, but that record does not resolve in DOI registries, bibliographic indexes, UCLA’s 1995 bibliography or the cited journal volume. The exact phrase appears as a later section heading, not as a 1995 publication. Publishing it as a primary source would create false provenance.
Canon 0027 is therefore transparently repaired to the verified 1995 AISTATS proceedings paper “A causal calculus for statistical research, with applications to observational and experimental studies.” The proceedings paper overlaps substantially with the Biometrika programme. Repetition would add little, so this companion narrows its lens to the proceedings version’s explicit language of two conditioning operators, its operational reading of structural equations, the stepwise calculus, surrogate experiments and the boundary between linear instruments and nonparametric identification.
Seeing-versus-setting laboratory
This finite binary laboratory makes the semantic claim executable. A hidden severity variable U affects treatment X and outcome Y; treatment also affects outcome. Choose an outcome mechanism, then change only the assignment policy. The exact observational contrast generally moves because X carries different information about U. The exact interventional contrast remains fixed because setting X bypasses the assignment equation while preserving the outcome equation.
The “randomized assignment” option removes the U→X dependence. There, observation and intervention agree because treatment no longer diagnoses severity. The “no causal effect” option preserves confounding but removes X from the outcome mechanism; it shows how association can survive when the intervention effect is zero. All calculations enumerate the eight binary states exactly: there is no Monte Carlo error.
Memory studio
One sentence
A structural causal model earns its action semantics by specifying which equation an intervention replaces and which mechanisms are asserted to remain invariant.
Two bars
see conditions inside the assignment system.
set replaces that assignment system.
Three permissions
observe add or remove evidence.
exchange seeing with setting.
act add or remove an intervention.
Four obligations
Define the intervention. Defend the graph. Show the derivation. Audit estimation and transport.
The coefficient test
Change the assignment policy while holding the outcome mechanism fixed. A structural response should stay attached to the controlled change even when the observational regression moves.
The design lesson
When a target is not identified, ask what measurement, surrogate experiment or structural restriction would close the gap: and record its cost and credibility.
A working discipline for causal claims
Begin with the decision and intervention, not the dataset. State the units, timing, treatment versions and outcome horizon. Write whether the action fixes a value, changes a policy distribution, alters a mechanism or merely selects cases. Then draw the smallest graph that exposes the assumed causal routes, including plausible latent common causes. Invite domain experts to disagree about missing arrows and invariance before fitting a model.
Next, run identification as a symbolic stage. Name the criterion or list each calculus transformation with its modified graph and separation condition. End with an estimand written entirely in distributions the study can supply. If no reduction exists, say so. Decide whether to collect a blocking variable, run a direct or surrogate experiment, weaken the target to a bound, or add a functional restriction whose scientific meaning can be defended.
Only then estimate. Check support where the estimand averages conditional distributions, report sampling and model uncertainty, test sensitivity to unmeasured paths and selection, and distinguish in-sample identification from transport to the decision population. A causal workflow is not “fit, label and deploy.” It is target, model, identify, estimate, stress-test and decide.
The lasting achievement of the 1995 calculus is not that it turns every observational dataset into an experiment. It does something more disciplined: it makes the extra premises visible and supplies a formal procedure for learning what they buy. An equation may speak about action, but only after the model says why its mechanism survives action. A graph may license a formula, but only after its missing arrows are accepted. The value of the calculus lies as much in its refusals as in its answers.