CausalUse-CAL: Causal-Use Questions, Identification, and Realizability
About this pattern
This is a generated FPF pattern page projected from the published FPF source. It is canonical FPF content for this ID; it is not a FPF Reference product feature page.
How to use this pattern
Read the ID, status, type, and normativity first. Use the content for exact wording, the relations for adjacent concepts, and citations to keep active work grounded without pasting the whole specification.
Type: Calculus (C) Status: Stable Normativity: Normative unless explicitly marked informative
Plain-name. Causal-use calculus.
Intent. Help a practitioner decide what a causal-looking claim is supported to say, under which limits, and which narrower statement remains when the support is insufficient.
Primary EntityOfConcern. One exact causal-use question. The claim, estimand or contrast, evidence paths, identification result, estimate, sampling-realizability result, performed sampling and resulting data, simulation result, support result, and downstream decision remain separately identified.
Not a physical ontology. This pattern does not define causation in the world or replace domain science. It supplies a practical interface for using causal evidence and models without promoting association, simulation, or a graph label into a stronger causal claim.
Use C.28 when a result is offered as support for a causal effect, intervention, counterfactual comparison, causal fairness claim, causal policy, causal benchmark, or causal explanation. Common cues include:
Keywords
- causal-use question
- causality ladder
- association
- intervention
- counterfactual
- Pearl Causal Hierarchy
- Structural Causal Model
- causal diagram
- causal estimand
- identification
- counterfactual sampling realizability
- causal support components
- CausalUseSupportResult
- target trial
- causal fairness
- off-policy causal evaluation
- causal-RL evaluation.
Relations
Content
Use This When
Use C.28 when a result is offered as support for a causal effect, intervention, counterfactual comparison, causal fairness claim, causal policy, causal benchmark, or causal explanation. Common cues include:
- “method A improves the outcome”;
- “users who received X did better, so X works”;
- “this policy would have prevented the failure”;
- “the model shows what would have happened”;
- “this fairness metric proves the intervention is fair”; and
- “this benchmark shows that one causal method is better”.
The cue opens a question, not a verdict. Ask what claim is being supported and what use of the evidence depends on that support.
Not this pattern when. If no causal statement or causal evidential reliance is current, stay with the direct pattern: C.16 for measurement, C.27 for temporal change, A.10 for an evidence path, C.11 for choice, C.19 for live-pool policy, C.24 for call planning, D.5 for bias or fairness audit, or G.9 for ordinary parity.
Activation condition. C.28 is needed when causal support changes the statement relied on by a downstream publication, choice, deployment, audit, assurance, policy evaluation, or benchmark. C.28 decides only the causal-support boundary. The downstream pattern still decides whether to publish, choose, deploy, certify, assure, or abstain.
Simulation boundary at entry. A report that only describes simulator output and makes no causal use exits to ordinary model or simulation handling. Simulator output offered as support for an effect, counterfactual, policy, fairness, benchmark, or evidence claim stays in C.28 and must name the model, assumptions, validation, supported causal use, and unsupported use.
What Goes Wrong If Missed
- association becomes an intervention-effect claim;
- a changed metric becomes causal fairness;
- a simulated trace becomes realized counterfactual evidence;
- an estimated number is treated as proof that its estimand was identified;
- support for one population or environment is transported to another without an endpoint or assumption; or
- a support verdict is mistaken for permission to publish, deploy, or certify.
What This Buys
The cheap result states the question, rung, available support components, the common validity threat that matters now, the causal statement supported, the statement not supported, and the next useful step. Heavy profiles appear only when identification, estimation, counterfactual-sampling realizability, actual sampling evidence, transport, target-trial emulation, causal policy evaluation, representation learning, or fairness work is actually current.
First-Minute Questions
- What exact causal-use question is being asked, and which claim-bearing episteme states it?
- Is the intended statement observational, interventional, or counterfactual?
- What is actually available: an evidence path and empirical data regime, an identification or bound result, an estimate, a prospective counterfactual-sampling realizability result, dated sampling Work plus resulting data, or a simulation result?
- Which common validity problem could overturn the use: intervention definition or consistency, time order, confounding, overlap, interference, selection or missingness, measurement, or transport?
- What causal statement or evidential reliance is supported now, and what stronger statement is not?
- Does the downstream pattern have enough basis to make its own decision, or should it abstain, downgrade, or request more evidence?
First Output
The first output may be only this triage:
supportedUse means the causal statement or evidential reliance supported under the named limits. It is not a permission or command. unsupportedUse states the nearby stronger causal statement or reliance that the evidence does not support.
Triage may be the final result when it blocks the overclaim and names the narrower statement. Do not open a durable object merely because a causal word appears.
Adjacent simulation examples. “The simulator produced these traces” with no causal reliance returns keepNonCausalSimulationUse. “The simulated traces support what would happen under policy P” remains inside C.28 and needs simulationResultRef, model assumptions, validation, supported use, and unsupported use.
Problem Frame
FPF already has patterns for measurement, temporal claims, evidence, assurance, choice, exploration, call planning, fairness, parity, and mathematical lenses. Each keeps its own result. Causal support cuts across them, so a small shared interface is needed.
The practical question is “what does this evidence support us to say about this causal question, and what would overturn that conclusion?”
Problem
Three kinds of collapse can produce causal overclaim:
- Rung collapse: observation, intervention, and counterfactual comparison are treated as the same question.
- Support collapse: data regime, identification, estimation, direct sampling, and simulation are treated as one alternative-valued “basis”.
- Authority collapse: an evidential conclusion is treated as publication, choice, deployment, fairness, or assurance authority.
C.28 keeps those distinctions visible while allowing a cheap stop.
Forces
Solution
Use the smallest result that answers the current question:
- triage the claim;
- stabilize the question in a small card when it must be reused;
- run the common threat screen;
- add only the specialist result needed now; and
- issue a small
CausalUseSupportResultwhen another pattern must consume the conclusion.
Public contract and support components
C.28 uses references to actual objects. It introduces no universal kind for a causal-use question, estimand, or potential-outcome contrast.
CausalUseQuestionRefidentifies the exact question content, normally a C.2.1 episteme.CausalEstimandRefidentifies the mathematical target or the episteme that describes it under its direct pattern.PotentialOutcomeContrastRefidentifies the exact contrast or its description.CausalUseSupportResultRefidentifies one C.2.1 result episteme defined below.
Support is composable. A real result may use several components:
These fields answer different questions. Do not compress them into one exclusive value. Each optional specialist ref identifies the exact result or record actually used. Keep a component's own assumptions, uncertainty or sensitivity, supported and unsupported uses, and reopen information with that component when its defining contract requires them; do not copy fields merely to complete a standard list. The common CausalUseSupportResult still states its own supportedUse, unsupportedUse, limits, optional evidence window, and reopenCondition. Naming a specialist subject in prose does not make its result available to a consumer.
CausalEmpiricalDataRegime is a local classification used only when it helps distinguish evidence:
realizedCounterfactualSamplingData is used only when an A.10 evidence path cites dated sampling Work and the resulting sample or data. A realizability result or WorkPlan alone establishes no empirical regime. Model output is recorded separately through simulationResultRef, not as an empirical regime.
Causality-Ladder Rung
- observational: passive observation, natural behaviour, or association;
- interventional: action setting, experiment, policy change, or action effect;
- counterfactual: counter-to-fact, potential-outcome, or unit-history-conditioned comparison.
Lower-rung data may contribute to a higher-rung result only through a replayable identification, bound, or other specialist result. The rung label itself supplies no support.
Causal-Use Claim Kind
Choose the kind by the claim being supported, not by the tool or source. Simulation for a causal claim uses the appropriate claim kind plus simulationResultRef; it does not need a simulation-only claim kind.
Question cards and support result
Use a local card when the question must survive beyond the current sentence:
Use a durable card only for a reusable or consequential claim:
When another pattern needs a stable conclusion, issue this small C.2.1 result episteme:
Its identity and reference follow C.2.1. The result states causal support only. A downstream pattern may cite it as one basis and then make its own decision. undecided supplies no causal conclusion; the downstream pattern decides whether to abstain, seek evidence, or use a non-causal result.
Common causal-validity screen
Run only the questions relevant to the current claim. A live threat either points to an existing specialist field/result or lowers the support result; it does not trigger a mandatory dossier.
Ordinary effect case. A randomized treatment study records interventionWellDefinedOrConsistency=clear, temporalOrdering=clear, positivityOrOverlap=clear, interferenceOrSpillover=notApplicable, selectionCensoringOrMissingness=clear, and measurementErrorOrConstructShift=clear for its declared target and window. The screen points to the trial and estimate results; it does not repeat them.
Countercase. An observational cohort has the right rung label and a plausible estimand, but records exchangeabilityOrConfounding=liveThreat and positivityOrOverlap=liveThreat because severity is unmeasured and one treatment region has no comparator. The claimed effect remains unsupported until a suitable design or new evidence closes those threats; a suitable bound can instead support a correspondingly bounded claim. “Observational data” was classified correctly; that label does not establish validity.
Identification result
Identification answers whether the estimand can be expressed or bounded from the available data and assumptions. The conclusion must be replayable:
An identified label without an identifying expression or derivation is incomplete. A bounded result cites the bound. A nonidentified result exposes the obstruction or failure witness. Identification is neither a numerical estimate nor direct physical sampling.
Replayable identified case. For treatment_effect_in_population_P, AdjustmentSet_Z is justified as blocking the relevant back-door paths. backdoor_adjustment_derivation_7 states the identifying expression in ordinary terms: compare treated and untreated outcomes within each Z group, then average those differences using the target population's Z distribution. The result cites the data regime, assumptions, expression, and the confounding or overlap change that would reopen it.
Replayable nonidentified case. In a treatment cohort, unmeasured severity affects both treatment and outcome, and no valid adjustment set, instrument, proxy, or useful bound is available. unmeasured_severity_obstruction_3 is the failure witness. The result is nonidentified; reporting an adjusted number does not change that status.
Counterfactual sampling realizability
Use this result to answer whether a declared target distribution can be sampled under current constraints. It is prospective: it does not say that sampling was planned, performed, or yielded data.
A realizable result cites the sampling construction that the decision Method accepts. A bounded result cites its bound. A nonrealizable result exposes the obstruction or failure witness. unclear names what remains unresolved. “Realized counterfactual sampling” never means observing incompatible outcomes for one unit in one realized world.
If the team plans to draw samples, use a separate A.15.2 WorkPlan. If sampling occurs, recover every precise performer's A.13 core and independently admit the dated Work under A.15.1. Add F.6 only when the sampling claim also needs precise assignment-bound attribution. If the samples are used as evidence, cite the resulting data through an A.10 evidence path. Actual sampling support requires both the dated Work and resulting data or evidence ref; neither realizable nor a WorkPlan can stand in for them. Identification from those data, when claimed, is another CausalIdentificationResult.
Applied profiles
Target trial.
An observational emulation keeps the protocol and its mapping to available data as separate results:
Every mapping field identifies the actual mapping. protocolToDataGapAccountRef points to one account that lists the observed gaps or explicitly states that none was found within the declared source and window. The residual-confounding and sensitivity fields remain present even when their bounded result is favourable. Reporting completeness is not a risk-of-bias, identification, or estimate verdict.
Filled target-trial mapping. HypertensionEmulationMap-2025 maps HypertensionTargetTrial-1 to ClinicRecords-2022-2024: age and diagnosis fields implement eligibility; prescription records distinguish the two treatment strategies; the prescription date supplies assignment and time zero; encounter records map the twelve-month follow-up; and the recorded systolic-pressure field maps the outcome. GapRecord-17 states that adherence after prescription is not observed, ResidualConfoundingAssessment-17 retains unmeasured severity as a live threat, and SensitivityMap-17 points to the negative-control and quantitative-bias analyses. The result supports construction and review of this emulation. It does not by itself establish identification, low bias, or a transportable effect; new severity or adherence data reopens it.
Estimation.
At least one identification or explicit design-based basis is required before the estimate supports a causal use. Orthogonal scores, nuisance models, and cross-fitting belong in methodSpecificDetailRefs only when a DML Method is selected. estimationConsistencyResultRef points to the consistency result defined by the selected estimation Method or its direct evaluation pattern; C.28 introduces no universal consistency-result kind.
Counterfactual fairness. Before D.5 uses a counterfactual-fairness support result, its C.28 components cite the identification result and the extra assumptions needed to connect the available data to that counterfactual question. When the fairness conclusion depends on an estimate, they also cite the estimate and its estimationConsistencyResultRef. Without those conditions, return bounded or unsupported; more data, even an unlimited amount of the same data, does not repair missing counterfactual identification or an inconsistent estimator. Associative or interventional fairness claims use their own rung and do not inherit this stronger branch by label.
Non-DML estimate. A randomized trial cites random_assignment_identification_4, trial_data_8, DifferenceInMeansMethod_2, standard_error_result_5, and its attrition sensitivity check. It needs no orthogonal-score, nuisance-model, or cross-fitting fields. The estimate supports only the declared population, outcome, assignment, and follow-up window.
Transport.
Identify every endpoint dimension that differs in the current claim. Population, domain, environment, data-generating regime, and semantic scheme answer different questions. A shared label proves nothing; a semantic Bridge is added only when its F.9 relation independently obtains.
Off-policy evaluation.
Causal representation. Use this record only when variables are learned, selected, or abstracted rather than supplied by the domain:
The record states which interventions and queries the learned or abstracted variables preserve, not that they are causal variables for every query or domain.
Filled representation case. WardStateRepresentation-4 derives three state variables from monitor traces through WardStateAbstractionMethod-2. Its intervention-validity result covers dosage interventions, its invariance result covers the two hospitals in the training and hold-out comparison, and its query-preservation result passes the declared one-step counterfactual query but fails the long-horizon query. The record therefore supports the one-step policy comparison only; a new hospital, sensor scheme, intervention family, or long-horizon claim reopens it.
Graph and calculus names
Use specialist names only when the result depends on them:
These values classify the formal support form. Concrete refs point to the model, diagram, derivation, assumptions, or proof. A graph-class label is not a proof and does not replace the plain statement of what was identified or bounded.
Causal evidence design and Work
Use CausalUseEvidenceDesignRecord when additional evidence could change the support boundary:
The three optional specialist refs are included only when an existing target-trial mapping, off-policy evaluation, or causal-variable representation result shows what additional evidence could change the support boundary. Before execution, cite a MethodDescription or WorkPlan only when used. After execution, cite every precise performer's A.13 core and the independent A.15.1 Work admission; cite F.6 only when precise assignment-bound attribution is also current. If performed counterfactual sampling is used as evidence, also cite the resulting sample or data through evidencePathRefs; Work without output data and data without its Work and provenance path each remain incomplete for that claim. Do not copy performer-kind, assignment, or occurrence mechanics into this record unless one of those facts changes causal validity, safety, authorization, or supported use.
Additional evidence is worth planning only when it can change a material causal statement or downstream decision enough to justify cost, risk, and delay, or when safety or release rules independently require it.
Support is not authority
CausalUseSupportResult.verdict has four values:
supported: the named causal statement or evidential reliance is supported under the stated limits;bounded: only the narrower statement or reliance is supported;unsupported: the claimed causal statement or reliance is not supported;undecided: the available work does not establish a causal conclusion.
The result never authorizes publication, choice, deployment, certification, fairness approval, or assurance. The downstream pattern cites it as one basis, considers its own other conditions, and makes its own decision. Practical guidance such as “report association only” states the remaining evidence boundary; it is not a permission issued by C.28.
Causal action policy class
Use this classification only when policy use changes the causal question:
unknown is an unresolved classification, not a fifth member. Omit the field when the distinction changes no support, comparison, or downstream decision.
Naming and ontology settlement
The public ...Ref names above are local reference contracts, not newly admitted universal kinds. Recover the actual object before choosing a reference:
Do not create a universal object merely to preserve a familiar token. Do not replace plain practitioner sentences with a list of ontology fields when the shorter sentence carries the same distinction and stop condition.
Neighbor selection
Non-Goals
C.28 does not define physical causation, choose one causal school for every domain, certify a graph by naming it, replace domain intervention or outcome definitions, replace measurement/evidence/fairness/choice/assurance/parity patterns, or authorize a downstream action. It also does not require a durable card or specialist profile when triage already blocks the overclaim.
Cheap downgrade library
Payoff check
Keep a causal-use record only when it changes the supported causal statement, blocks a concrete overclaim, changes evidence work, or supplies a real basis to a downstream decision. Remove fields that do none of those things. Prefer triage when it preserves the same boundary.
Publication-unit boundary
When only wording inside one publication unit is unclear, use the publication and wording patterns. Open C.28 only when the wording is relied on causally. A publication decision remains with the publication pattern even after C.28 returns a support result.
Causal-laundering cases
Archetypal Grounding
System. A product team observes better outcomes among recipients of X. Triage returns association support. If the team needs an effect claim, it opens identification or evidence-design work; the deployment decision still needs its own downstream basis.
Fairness. A report claims counterfactual fairness after a policy change. C.28 identifies the rung and estimand, exposes the additional counterfactual-identifiability assumptions, and cites an estimate with its consistency result when the audit relies on that estimate. Missing identification or consistency lowers the support result even with more of the same data. D.5 carries the BiasAuditReport@Context and makes the audit conclusion.
Policy. Logged behaviour data are used to evaluate a new policy. The result names both policies, horizon, confounding and overlap checks, transport endpoints when changed, estimate and uncertainty, supported regime, and unsupported unqualified optimality. C.11 or another policy pattern makes the choice.
Causal RL. An online learner combines logged behaviour, interventions, and a counterfactual-data source. The sampling-realizability result explains whether that source can be produced; dated Work and the resulting data path show whether it was produced; a separate identification or estimate result says what follows from it. Replay reward does not become an optimal-action claim.
Evidence Work. A lab's CounterfactualSamplingRealizabilityResult cites its decision Method and positive construction. That result supports planning but claims no sample. The later WorkPlan remains prospective. After sampling, the lab cites independently admitted dated Work and the resulting data in an A.10 evidence path before using realizedCounterfactualSamplingData; it adds precise assignment-bound attribution only when the receiving support claim uses it. Identification from those data remains a separate result.
Simulation. A simulator supports rehearsal and sensitivity analysis under named assumptions and validation. The support result blocks realized-sample and intervention-effect wording. A pure simulator-output report exits C.28 earlier.
Transport. The population is unchanged but the care environment and measurement mechanism differ. The transport result names source and target environments and data-generating regimes, then states the assumptions and formula. A population ref alone would miss the shift.
Benchmark. G.9 compares an observational predictor, intervention optimizer, and counterfactual policy only after it checks rung, estimand, support components, window, and endpoints. The admissible result may be a selected set or abstention rather than a scalar winner.
Bias-Annotation
Watch for causal prestige, simulation laundering, metric proxy substitution, graph sufficiency, feasibility-as-performance, data-without-Work, support-label substitution, and benchmark scalarization. Recover the question, support components, live threats, supported statement, unsupported statement, and reopen condition in the shortest form that remains replayable.
Conformance Checklist
- One exact causal-use question remains identifiable from entry to result; question, claim, estimand, evidence, and records are not treated as one object.
- C.28 introduces no universal causal-use-question, estimand, or potential-outcome-contrast kind; it uses local refs to actual objects instead.
- Data regime, identification, estimate, sampling realizability, performed sampling evidence, simulation, and transport remain distinct and may be combined.
- A support result states evidence support only; every publication, choice, deployment, fairness, or assurance decision remains with its direct pattern.
- An identified result cites an expression or derivation; a bounded result cites a bound; a nonidentified result cites an obstruction or witness.
- A causal estimate cites an identification or explicit design-based result. Method-family details appear only when that Method is selected.
- The common threat screen routes every live ordinary threat or lowers the result; it is not a mandatory dossier.
- Non-causal simulator reporting and simulation-supported causal use take different routes at first entry.
- A sampling-realizability result cites its decision Method, any derivation used, and the construction, bound, or obstruction required by its status; it claims no Work or data.
- Performed counterfactual-sampling support cites independently admitted dated Work and resulting data or evidence; it cites exact assignment-bound attribution only when the receiving support claim uses it. A WorkPlan or
realizablelabel cannot satisfy this branch. - Evidence design cites each precise performer's A.13 core and the independent A.15.1 Work result; it cites F.6 only when exact assignment-bound attribution is current, rather than copying assignment mechanics.
- Transport identifies every changed population/domain/environment/data-generating-regime endpoint separately from semantic schemes.
- A counterfactual-fairness escalation exposes its additional identification assumptions and, when an estimate is used, estimation consistency before D.5 consumes it.
CausalActionPolicyClasshas the same four members in definition, examples, and consumers; unresolved classification is not a member.- Every specialist field changes support, a downstream decision basis, evidence work, or a reopen condition.
- A target-trial mapping result identifies the observational source and every protocol-to-data mapping, gap, residual-confounding assessment, and sensitivity mapping needed for its bounded use.
- Every retained specialist result that can independently change support can enter
CausalSupportComponentRefs; when it shapes further evidence, the evidence-design record can cite the same result without copying it. - The whole pattern remains understandable to a practitioner without requiring the formal graph vocabulary on the ordinary path.
Common Anti-Patterns and How to Avoid Them
Consequences
The pattern makes unsupported causal claims easier to lower while keeping ordinary triage cheap. Consequential claims become replayable across question, support components, threats, limits, and source window. The cost is additional specialist work only where a stronger causal statement or downstream decision genuinely depends on it.
Rationale
Temporal change, a higher metric, a convincing graph, or a plausible simulator can all be useful without supporting a causal effect. Conversely, observational data can support a causal estimate when an explicit identification result closes the inferential gap. C.28 therefore separates the question from the components that support it and separates that evidential conclusion from downstream authority.
The integrated contract is deliberately plural: SCM and graphical methods, potential outcomes, target-trial practice, design-based identification, causal ML, transport, causal representation learning, causal RL, and causal fairness may supply different specialist results. None is installed as the universal method.
SoTA and lineage
Qualification window. This comparison was reviewed through 2026-08-21. Reopen it when a current contribution is materially superseded, TARGET guidance changes, a specialist branch changes the minimum replay fields, or direct consumers need a different support-result contract.
Why this combination is retained. The 2025 realizability result answers whether samples can be produced; the 2026 completeness and bounding result answers what can be identified from realized Layer-3 data; the 2026 fairness result states extra conditions for one consequential downstream use. Keeping those results separate preserves their different questions while allowing explicit composition. The domain-specific 2026 lines improve selected Methods but do not dominate the small shared interface. This is the current non-dominated contract for a practitioner who needs a cheap causal stop plus replayable specialist results.
Synthesis boundary. No source above establishes the whole C.28 architecture. The orthogonal support components, common threat screen, small support-result interface, support/authority split, and cross-pattern consumer contract are a bounded FPF synthesis. Validate them through filled cases and consumer replay; reopen when they hide a real causal distinction, impose unused apparatus, or fail a practitioner.
Relations
C.16keeps measurements and scales;C.27keeps temporal-claim adequacy.A.10keeps evidence paths and provenance and may cite C.28 support components and result.A.2.4classifies how an episteme is used; it cannot promote simulation output or association into stronger causal evidence.A.15keeps Method, plan, Work, and attribution for interventions, target trials, and sampling.B.3may cite a C.28 result as one basis for a separate bounded assurance result.C.11,C.19, andC.24keep choice, pool treatment, and call planning and consume only the needed causal refs.D.5keeps bias/fairness audit and usesBiasAuditReport@Contextwhen a causal fairness question is consequential or reusable.G.5keeps method dispatch;G.9keeps parity and benchmark conclusions;G.11keeps refresh planning.C.26is used only for a residual quantum-like modelling issue after ordinary causal explanations are tried.
C.29 mathematical-lens relation
C.29 may describe a mapping as abstraction-like, quotient-like, coarse-graining-like, simulation-like, or macro-model-like. It does not decide causal support. When intervention, policy, counterfactual, causal explanation, or causal decision use is current, apply C.28; otherwise record no causal-use claim or the exact blocker.
C.28:End
Last Updated: 2026-09-10 — upstream FPF commit a87d0ef4 (github.com/ailev/FPF)