fpf.tools
Library / First Principles Framework (FPF) - Core Conceptual Specification
Jump to passage
In this reading
Source · lines 63960–64096

Link to current text

Published source confirmed at last check

Source changed 2026-10-03 05:29:54 UTC · snapshot created 2026-10-03 05:30:57 UTC · last check 2026-10-03 07:10:10 UTC

C.28:4.6 - Applied profiles

Target trial. The TARGET Statement, 2025 supplies reporting guidance for an observational study explicitly emulating a target trial. Use it for the protocol-to-data account below, with causal validity checked separately.

TargetTrialProtocolRecord:
  causalUseQuestionRef: CausalUseQuestionRef
  targetPopulationRef
  eligibilityCriteriaRef
  treatmentStrategyRefs
  assignmentProcedureRef?
  timeZeroRef
  followUpWindowRef
  outcomeMeasureRef
  potentialOutcomeContrastRef?: PotentialOutcomeContrastRef
  causalEstimandRef: CausalEstimandRef
  analysisPlanRef

An observational emulation keeps the protocol and its mapping to available data as separate results:

TargetTrialMappingResult:
  causalUseQuestionRef: CausalUseQuestionRef
  targetTrialProtocolRef
  observationalDataSourceRef
  eligibilityCriteriaMappingRef
  treatmentStrategyMappingRefs
  assignmentAndTimeZeroMappingRef
  followUpWindowMappingRef
  outcomeMeasureMappingRef
  identifyingAssumptionRefs
  protocolToDataGapAccountRef
  residualConfoundingAssessmentRef
  sensitivityMappingRefs
  supportedUse
  unsupportedUse
  reopenCondition

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.

CausalEstimateResult:
  causalEstimandRef: CausalEstimandRef
  identificationResultRef?: CausalIdentificationResultRef
  designBasedIdentificationResultRef?
  dataRef
  estimatorMethodRef
  diagnosticRefs?
  uncertaintyResultRef
  sensitivityAnalysisRef?
  estimationConsistencyResultRef?  # when consistency is a live support condition
  methodSpecificDetailRefs?        # only for the selected estimator family
  supportedUse
  unsupportedUse

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.

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. For the first-moment population-measure problem under covariate shift, use Boughdiri, Berenfeld, Josse and Scornet, A Unified Framework for the Transportability of Population-Level Causal Measures, 2025, §§2–4 with its internal trial-validity, overlap and selected exchangeability conditions. Other endpoint changes need their own identifying result.

CausalTransportabilityResult:
  causalUseQuestionRef: CausalUseQuestionRef
  sourcePopulationRef?
  targetPopulationRef?
  sourceDomainRef?
  targetDomainRef?
  sourceEnvironmentRef?
  targetEnvironmentRef?
  sourceDataGeneratingRegimeRef?
  targetDataGeneratingRegimeRef?
  selectionAssumptionRefs?
  domainShiftAssumptionRefs?
  sourceWindowRef?
  targetWindowRef?
  overlapEvidenceRef?
  transportComparatorOrFormulaRef
  semanticBridgeRef?          # only when interpretation differs
  supportedUse
  unsupportedUse
  unresolvedAssumptionRefs?

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.

OffPolicyCausalEvaluationResult:
  causalUseQuestionRef: CausalUseQuestionRef
  evaluationPolicyRef
  behaviorPolicyRef
  sequentialHorizonRef?
  unitHistoryConditioningRef?
  confoundingAssumptionRefs?
  overlapOrSupportCheckRef
  policyTransportabilityResultRef?
  estimatorRef?
  uncertaintyResultRef?
  supportedUse
  unsupportedUse
  reopenCondition

Causal representation. Use this record only when variables are learned, selected, or abstracted rather than supplied by the domain:

CausalVariableRepresentationRecord:
  causalUseQuestionRef: CausalUseQuestionRef
  sourceRepresentationRef
  selectionOrAbstractionMethodRef
  representationAssumptionRefs
  interventionValidityResultRef
  invarianceResultRefs?
  abstractionFidelityResultRef?
  counterfactualQueryPreservationResultRef?
  uncertaintyResultRef?
  shiftLimitRefs?
  supportedUse
  unsupportedUse
  reopenCondition

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.