PaperlayerView exact revision on arXiv

Source-linked AI summary

Information Set Emulation: Causal Certificates for AI Derived EHR Features

Takes Fujita, Nobutaka Hattori

arXiv:2609.17777v1stat.MEcs.AIq-bio.QM

TL;DR

AI-derived EHR features can recover clinically meaningful information, but predictive usefulness alone does not establish that they are admissible for causal inference. The paper introduces a typed, certificate-based framework for reconstructing a prospective information set and measuring residual ambiguity in a locked estimand. Its conclusion is that typed evidence can support a point claim when ambiguity is sufficiently reduced, while unresolved roles call for compatible reporting.

  • Problem

    AI can recover clinically meaningful EHR features, but predictive usefulness does not establish whether reconstructed information is admissible for causal inference.

    Source: §1 · p. 2

  • Method

    The AI typed lift attaches source, timing, decision-time availability, representation version, causal-role, and ambiguity evidence to features, then routes unresolved downstream roles away from point estimation.

    Source: §1 · p. 2, Certificate workflow · p. 8

  • Results

    Typed evidence narrows the compatible estimand set, and its squared radius measures the residual minimax information-structure risk for the locked scalar target.

    Source: §13 · p. 36, §14 · p. 40

  • Takeaways & Limitations

    The framework uses the remaining information ambiguity to distinguish when reconstructed EHR information can support a point claim from when compatible reporting is needed.

    Source: Certificate workflow · p. 8, §13 · p. 36

    The paper uses synthetic data and text vignettes only, so it does not provide patient-data validation or establish treatment effectiveness.

    Source: §12 · p. 35, Downstream domination is a… · p. 39

Abstract

from arXiv · show

AI and large language models can recover clinically meaningful features from electronic health records (EHRs), but predictive usefulness does not establish admissibility for causal inference. We introduce information set emulation: an AI typed lift attaches source evidence, clinical and recording times, decision-time availability, representation version, proposed causal roles, and unresolved ambiguity to extracted features under a locked target trial. Causal certificates record auditable evidence for those roles. Features with unresolved downstream roles are routed to compatible reporting or separate analyses. Typed evidence defines an observational fiber of causal worlds consistent with the observed law. The locked scalar estimand maps this fiber to a compatible image whose squared Chebyshev radius equals the residual minimax mean squared error when the image is nonempty and compact. This classical identity provides a target-specific measure of information ambiguity. The contribution is its integration with a joint EHR observation map and an auditable certificate architecture. Under explicit exchangeability, positivity, and nuisance-consistency conditions, we give identification and cross-fitted augmented inverse probability weighted estimation, distinguishing empirical and population targets. An EHR compression-drift identity separates the roles of frame presence, treatment assignment, and outcome observation. Artificial simulations and a common-law finite-world example illustrate estimation failures and information-radius reduction. Synthetic Phase 0 notes demonstrate audit diagnostics; a separate role-specific analysis spread illustrates routing and is not an exact fiber radius. All experiments are synthetic. The framework specifies when reconstructed information can support a point claim and when compatible reporting is required.

1 Introduction

Information set emulation asks whether AI-reconstructed EHR evidence recovers the pre-treatment information needed for a locked causal analysis, not merely whether extracted features predict outcomes. It uses typed evidence and certificates to assess which causal worlds remain compatible with the record and whether a point claim is supportable.

2 EHR event streams and prospective information

  • A retrospective study may match a prospective trial's treatment strategies and timing yet differ in the information available for confounding control, follow-up, and interpretation.The information set must cover treatment assignment, eligibility, observation, and outcome comparison; outcome prediction alone is insufficient.Source: §1 · p. 2, §2 · p. 5

1 Introduction

  • Information set emulation reconstructs a prospective pre-treatment information set by attaching source evidence, clocks, availability, proposed causal roles, and ambiguity to AI-derived features.The typed lift's value is measured by how much it shrinks the compatible estimand set.Source: §1 · p. 2, §2 · p. 4
  • Without sufficient evidence about source time, decision-time availability, and causal role, compatible prospective worlds can imply different values of the locked estimand.The resulting nonidentification cannot be resolved by prediction accuracy alone.Source: §1 · p. 2

2 EHR event streams and prospective information

  • Features with unresolved downstream roles are kept out of the primary point estimator and routed to compatible reporting or separate analyses.This avoids admitting possible mediators, outcome proxies, or mixed-window features merely because a baseline role is also plausible.Source: §1 · p. 3, §2 · pp. 4–5

1 Introduction

  • The locked scalar estimand maps the observational fiber of compatible causal worlds to an estimand set whose squared Chebyshev radius is the exact minimax mean squared error.Typing and certification are useful when they shrink that radius without excluding plausible worlds.Source: §1 · p. 3, §2 · p. 5
  • The paper is a theory and simulation study, not a patient-data analysis or a claim about a clinical treatment effect.Its contribution is the certificate architecture, joint EHR observation map, and role-aware compatible-estimand construction.Source: §1 · p. 3

3 AI typed EHR lift and causal certificates

The AI typed lift sorts EHR features into proposed causal roles and attaches auditable evidence to support their admissibility under a locked target trial. A role-specific certificate and workflow determine whether a feature enters point estimation or is routed elsewhere.

3 AI typed EHR lift and causal certificates

  • The typed lift performs a first causal sorting of raw EHR material into baseline, design, history, effect-modifier, observation, mediator, outcome-proxy, intercurrent-event, or excluded roles.It is not an assertion that AI is an oracle; the lift must return evidence along with types.Source: §3 · pp. 5–6
  • A causal certificate records auditable evidence that a feature is admissible for a proposed role under the locked target trial; it does not certify that the feature is true.The evidence includes source, timing windows, representation version, local role claim, and unresolved information.Source: §3 · pp. 5–6

Certificate workflow

  • A primary causal estimator admits a feature only through a role whose required predicates pass under the locked target trial.If roles remain uncertain, the target is a compatible estimand set; features that fail point-estimator predicates can be routed to compatible reporting or other analyses.Source: Table 1 · p. 7, §3 · p. 7, Certificate workflow · p. 8

3 AI typed EHR lift and causal certificates

  • Clinical time before treatment is distinct from decision-time availability: a post-index note may describe a pre-treatment fact without establishing that it informed treatment choice.Source separation can support baseline content, while availability certification is a stronger claim.Source: §3 · p. 7

Architecture neutral extraction schema

  • The extraction workflow locks the trial, records source and clocks, stores the representation version, proposes evidence-backed roles, and applies role predicates before outcome modeling.The framework does not require a specific NLP architecture, but a feature value alone is insufficient for confirmatory causal estimation.Source: Certificate workflow · p. 8, Architecture neutral extraction schema · pp. 8–9

4 Prospective approximation as comparison of information experiments

The paper compares the information experiment of a prospective study with that of an AI-typed EHR study, asking whether the latter can reproduce the former closely enough for a causal decision. It distinguishes this global benchmark from target-specific information ambiguity and cautions that the global distance is usually impractical to estimate directly in an applied EHR study.

4 Prospective approximation as comparison of information experiments

  • If the AI-typed EHR experiment is close to the prospective information experiment, prospective causal decision rules can be approximated with a controlled increase in risk.The risk-transfer result assumes bounded loss and a small one-sided information deficit.Source: §4 · p. 10
  • The experiment distance is a global benchmark for how well typed EHR information approximates prospective information, not a statistic routinely estimated in real EHR studies.The paper states that estimating it directly is usually too ambitious and that no general KL or f-divergence lower bound for the target-specific radius is established.Source: §4 · p. 10
  • The prospective emulation gap is target-specific: it measures remaining exact minimax risk for the locked scalar estimand rather than approximation of the whole information experiment.When the prospective structure point-identifies the target, its radius is zero and the gap equals the retrospective squared radius.Source: §4 · p. 11
  • Trial reports and protocols can anchor a prospective reference information structure, but they do not determine the EHR information radius.The retrospective radius also depends on source, clock, availability, role evidence, sensitivity envelopes, and the observed law.Source: §4 · p. 11

5 Observational fibers and compatible estimands

The compatible estimand set captures the causal effects still possible under the observed EHR law and typed evidence; its Chebyshev radius quantifies residual information risk. The section develops exact and near-fiber results, the value of refining evidence, and a validation-calibrated approach to reporting compatible estimates.

5 Observational fibers and compatible estimands

  • The observational fiber contains causal worlds consistent with the observed law and typed evidence, conditional on the analyst’s declared identification and sensitivity model.A richer model that allows more worlds can only enlarge the compatible set and its radius.Source: §5 · p. 11
  • When the compatible estimand set is nonempty and compact, its squared Chebyshev radius equals the exact minimax mean squared error for the locked scalar target.A Chebyshev center attains this bound; the radius measures information-structure risk, not sampling variance.Source: §5 · pp. 12–13
  • At finite sample sizes, the approximate information modulus measures how far the locked estimand can differ across worlds that are difficult to distinguish under the typed retrospective law.It is a near-fiber analogue, not a general formula linking global experiment distance to the exact radius.Source: §5 · p. 13
  • A flat table without source-time, availability, or causal-role evidence can leave pre-treatment, post-treatment, and outcome-proxy interpretations compatible, preventing identification when their estimands differ.In that case the information radius is positive.Source: §5 · p. 14
  • Refining typed evidence can only shrink the compatible estimand set and its radius when the refined evidence can reproduce the coarser observation.Certificates are useful insofar as they refine evidence and remove worlds from the fiber; they do not create trust by assertion.Source: §5 · pp. 12, 14

Value of typing and value of certification

  • The value of AI typing or additional certification is assessed through reductions in radius, but a capped squared-radius score equals minimax risk reduction only when the cap is inactive at both evidence states.Additional evidence can be evaluated by its reduction in radius, and a wide radius can help identify which evidence to collect next.Source: Value of typing and value of… · p. 15

Estimating a compatible reporting set

  • Compatible reporting fixes a finite set of role assignments for the same locked estimand and combines their nominal intervals with validation or sensitivity envelopes.Direct-effect and total-effect targets require separate locks; they are not alternative values of one locked estimand.Source: Estimating a compatible reporting set · pp. 15–16
  • Coverage requires estimand-scale bounds linked to validated discrepancies; an extraction or availability error rate alone does not supply such a bound.The compatible set may be disconnected, and its reporting radius combines sampling and information uncertainty.Source: Estimating a compatible reporting set · pp. 16–17

6 Causal effect identification and estimation

The section makes causal estimation conditional on a typed, locked information set and credible identification assumptions; uncertainty about feature roles instead calls for compatible estimands. Once the information set and assumptions are fixed, cross-fitted augmented inverse probability weighting can estimate population or empirical-covariate targets.

6 Causal effect identification and estimation

  • The typed lift determines which features may define the estimand and nuisance functions, but does not establish exchangeability or eliminate unmeasured confounding.Residual clinician judgment, site practice, severity, or observation mechanisms may remain unmeasured.Source: §6 · p. 18
  • Point identification requires certified pre-treatment measurability, consistency, information-set and observation exchangeability, and positivity for treatment and outcome observation.The estimator targets the effect identified by the certified information set, which coincides with the prospective estimand only when sufficiently close for the scientific purpose.Source: §6 · p. 19
  • When observation exchangeability is doubtful, the analysis should use an observation-sensitivity family; larger allowed departures yield a larger compatible set and radius.The radius curve makes departures explicit rather than hiding them in a complete-case point estimate.Source: §6 · p. 19
  • Cross-fitted augmented inverse probability weighting distinguishes the population effect from the empirical-covariate target, with variance depending on which target is reported.The stated empirical-target result concerns marginal error around the random covariate-average target; conditional coverage needs an additional conditional central limit theorem.Source: §6 · pp. 20–21
  • For repeated episodes, folds must split patients and scores must use declared within-patient weights; arbitrary dependent episodes cannot be treated as independent contributions.Longitudinal treatment or observation requires corresponding sequential identification assumptions; product-rate nuisance conditions alone do not establish the influence function without individual consistency or fitted-score convergence.Source: §6 · p. 21

AI extraction regimes

  • Value instability across seeds or versions differs from role uncertainty, which changes the estimand and belongs in the compatible set rather than an ordinary standard error.Outcome-tuned extraction is not confirmatory without an independent split, external validation, or justified selection adjustment; unacceptable false-admissible rates preclude a single confirmatory point estimate.Source: AI extraction regimes · pp. 22–23

7 EHR compression drift

The EHR compression-drift identity relates the observed outcome mean after compressing prospective information to the potential-outcome mean in the compressed stratum. Its drift depends on conditional covariance between the outcome mean and joint selection into the EHR frame, treatment, and observed outcome.

7 EHR compression drift

  • Under consistency, the stated conditional independence assumptions, and a positive denominator, compression drift equals the conditional covariance of the potential-outcome mean and joint observation mechanism divided by its conditional mean.The joint mechanism combines EHR frame presence, treatment assignment, and outcome observation.Source: §7 · p. 23
  • The identity reduces when the target is restricted to the EHR frame or all outcomes are observed, removing the corresponding frame-presence or observation factor.With randomized treatment and fixed follow-up, only assignment remains, recovering the simpler generated-covariate compression result.Source: §7 · pp. 23–24
  • Preserving the joint selection probability or the potential-outcome mean within the compressed representation is sufficient for zero drift; accurate prediction of observed outcomes alone guarantees neither condition.Either measurability condition is sufficient, not necessary: the two residual components may also be conditionally orthogonal.Source: §7 · p. 24
  • Zero arm-specific drift can arise through conditional covariance cancellation even when both the outcome mean and observation mechanism vary within a compressed stratum.The section gives a finite example with both functions nonconstant but covariance zero.Source: §7 · p. 24

8 Recovery error and drift bounds

A recovery-error bound provides one sufficient route from reconstruction of prospective information to small compression drift. It controls residual variation in both the potential-outcome mean and the joint observation mechanism.

8 Recovery error and drift bounds

  • With Lipschitz outcome and observation functions and a positive conditional mean observation mechanism, integrated absolute drift is bounded by their Lipschitz constants times squared reconstruction error.This is a sufficient bound, not a necessary condition: exact measurability of either function or conditional covariance cancellation can also yield zero drift.Source: §8 · p. 25

9 Post treatment text and source separation

Post-index notes can mix pre-treatment history with treatment response, so clinical time and decision-time availability must be distinguished when reconstructing baseline information. Source separation can support baseline use, while decision reconstruction additionally requires evidence of availability by the treatment decision.

9 Post treatment text and source separation

  • A post-index note may combine pre-treatment history with post-treatment response, making exclusion risk loss of baseline information and use of the full note risk leakage.A source-separated feature may be baseline clinical information, but decision information requires the stronger availability condition.Source: §9 · pp. 25–26
  • For a feature from a post-index note, baseline admissibility requires invariance to relevant post-treatment content or source separation; decision or design use also requires certified availability by decision time.If source separation fails, conditioning can compare treatment-specific strata; if availability fails, the feature may describe baseline state without reconstructing the information guiding treatment choice.Source: §9 · p. 26
  • The paper proposes masking, span-restricted extraction, and adversarial review as auditable source-separation checks, while noting they do not prove invariance for all notes.The checks can inform the separation allowance and exclusion of features that cannot be separated.Source: §9 · p. 26

Synthetic certificate vignette

10 Certificate conditional approximation frontier

The certificate conditional approximation frontier compares information ambiguity with center-estimation error to assess whether an AI-augmented EHR study meets a prospective-study error criterion. The criterion is conditional: the information radius and estimation bias must be sufficiently small.

10 Certificate conditional approximation frontier

  • For a nonempty compact compatible image, the theorem bounds worst-world error using the information radius and Chebyshev-center estimation error.The bound separates residual ambiguity from sampling error in estimating the center.Source: §10 · pp. 27–28
  • The prospective-study approximation criterion compares the resulting EHR error bound with prospective mean squared error plus a declared nonnegative tolerance.The passage distinguishes the mean-squared-error criterion from the weaker root-mean-squared-error inequality.Source: §10 · p. 28
  • The frontier is a decision criterion, not a claim that retrospective EHR studies equal prospective studies.A large EHR study approximates the mean squared error of a smaller prospective study only when its information radius and estimation bias are small enough.Source: §10 · p. 28

11 Simulation study

The artificial simulations test whether reconstructed information preserves a locked causal target and whether certificates reduce ambiguity enough for point estimation. Results show that accurate, correctly certified pre-treatment information can approach the oracle benchmark, while unresolved or misclassified information can leave the target ambiguous or distort estimation.

11 Simulation study

  • The simulation is a controlled artificial failure-mode experiment, not a clinical treatment-effect demonstration or external validation of EHR feasibility.The mechanisms and prospective information set are known; the design uses 1000 replications with 1000 generated patients per replication.Source: §11 · pp. 28–29

11.1 Data generating mechanism

  • The ISE causal estimator recovered nearly oracle performance with small bias and approximately nominal coverage; structured EHR retained large residual bias, while naive AI, design erasure, and post-index leakage moved toward the wrong target.The post-index leakage estimator was precise but wrong.Source: §11.1 · pp. 29–30

11.2 Flat table nonidentification and certificate shrinkage

  • The same flat observed table supports baseline, mediator, or outcome-adjacent roles for Z with different estimands; typed evidence shrinks the radius from 0.315 to 0.105, and source-and-availability auditing reduces it to zero.A post-index certificate removes the baseline world without shrinking the radius because the two extremal targets remain.Source: §11.2 · pp. 30–31, Table 9 · p. 30, Table 10 · p. 31

11.3 Compression drift identity

  • Exact enumeration checked the compression-drift identity across 3 compressed versus 12 richer strata, with discrepancies attributable only to floating-point arithmetic.The table reports population-stratum-probability-weighted absolute drift and the conditional-covariance summary; there is no Monte Carlo or binning error.Source: §11.3 · p. 31, Table 11 · p. 31

11.4 Prospective approximation frontier

  • Three worlds with the same observed law have estimands ΨB = 0.250, ΨM = 0.460, and ΨY = −0.170.These values belong to explicitly compatible causal worlds, not estimator-bias proxies for the information radius.Source: §11.4 · p. 32
  • 0.068 is the prospective reference RMSE, and zero tolerance is met only when the source-and-availability audit certifies the pre-treatment world.The typed lift alone narrows the fiber but leaves baseline-versus-mediator ambiguity.Source: Table 12 · p. 32, Figure 1 · p. 33, §11.4 · p. 33

11.5 Certificate and extraction stress tests

  • Stress tests show that extraction error, leakage, and availability error can worsen estimator performance, while positivity stress increases variability; failures act through different parts of the information structure.The settings are stylized controlled experiments, not external validation; when certificate errors cannot be bounded, the analysis should report a compatible set or remain exploratory.Source: §11.5 · pp. 33–34, Table 13 · p. 34

12 Phase 0 validation of the typed lift

Phase 0 tests whether an AI typed lift can reliably produce an information ledger, not whether treatment is effective. Synthetic note validation illustrates audit diagnostics, while a separate routing example does not estimate exact fiber radius or minimax risk.

12 Phase 0 validation of the typed lift

  • Phase 0 samples notes or event streams for typed-lift review and reports source, timing, availability, role, disagreement, and extractor-stability diagnostics.When a common-law world model and validated exclusion rules are available, reporting can include information-radius change.Source: §12 · p. 34, Table 14 · p. 35
  • A successful Phase 0 shows that the record can be lifted with quantified residual ambiguity, but does not establish exchangeability or treatment effectiveness.Persistent reviewer disagreement remains role or availability uncertainty and is carried into compatible reporting; the validation may use tens to hundreds of records.Source: §12 · p. 35

12.1 Synthetic Phase 0 note validation

  • The synthetic note exercise generates source, time, availability, and role labels, then tests recovery under prespecified errors and a simulated audit.Its role-only point admission omits the complete certificate predicates, availability gate, and outcome-analysis workflow; the false-admissible proportion is conditional on that rule.Source: §12.1 · p. 35
  • The note validation and separate routing illustration use declared artificial mechanisms; stipulated analysis lists are not inferred from measured audit errors.The validation measures event-stream lifting, while the routing illustration has no patient data and makes no claim of external EHR feasibility.Source: §12.1 · pp. 35–36
  • Audit mainly improves clock, availability, and source-separation evidence, fields that can determine whether a feature enters the point information set or compatible set.The synthetic validation uses 120 generated patients and 600 candidate feature records; reported values are proportions.Source: §12.1 · p. 36, Table 15 · p. 36
  • Table 16’s decreasing half-range is descriptive only, not evidence that audit performance shrinks a common-law fiber or exact minimax risk.The role-specific values answer different questions, and the lists are not calibrated from Table 15.Source: §12.1 · p. 36, Table 16 · p. 36

13 Discussion

The discussion frames typed evidence as a way to connect retrospective records to a prospective information set and determine whether a locked causal claim is point-identified or requires compatible reporting. Certificates, audit radii, and strict routing rules make unresolved information visible and guide validation.

13 Discussion

  • Typed evidence defines the observational fiber and compatible estimand image; the image’s squared radius is the exact minimax information-structure risk for the locked scalar target.The construction extends beyond EHRs to retrospective event streams, though clinical facts, timing, availability, observation, and outcome-adjacent narratives make EHRs especially entangled.Source: §13 · pp. 36–37

Relation to adjacent methods

  • The contribution is an auditable certificate joining source, timing, availability, observation status, and causal role to a joint fiber and estimand-scale validation rules.The radius and Chebyshev identity are not presented as a new standard-error correction or general theorem; the contribution is their integration in an EHR information structure.Source: Relation to adjacent methods · p. 37

Auditing published studies and reverse trial reports

  • A published-study audit measures ambiguity in reported information, not the probability that a conclusion is true or whether it is false.Its radius is conditional on documented evidence, observed-data laws, and declared identification and sensitivity models; this paper defines the audit object but does not conduct a full applied audit.Source: Auditing published studies and reverse… · p. 38

How to shrink the radius in practice

  • A wide radius can direct the next validation step, such as checking source timing, adjudicating availability, testing extractor stability, or modeling observation processes.When clinician judgment is unrecorded, chart review, a prospective substudy, or an explicit sensitivity parameter are options; unresolved components require compatible estimand reporting.Source: How to shrink the radius in practice · pp. 38–39

Downstream domination is a primary-analysis rule

  • Features that may be mediators, outcome proxies, post-treatment descriptors, or observation processes are excluded from the primary point estimator absent evidence resolving their roles.They remain available for compatible estimand sets, separate analyses, or sensitivity analysis; a favorable sensitivity parameter does not justify primary-estimator inclusion.Source: Downstream domination is a… · p. 39
  • The framework uses synthetic data and text vignettes, so it does not establish feasibility in a real hospital system; a small deidentified Phase 0 study is the proposed empirical next step.Generated-regressor uncertainty is only partly addressed, and radius coverage depends on role enumeration, sensitivity bounds, and simultaneous sampling and uniform validation conditions.Source: Downstream domination is a… · p. 39

14 Conclusion

Information set emulation asks whether an EHR study reconstructs the pre-treatment information needed for its target trial, not merely the trial design itself. Typed evidence narrows compatible causal estimands and quantifies residual ambiguity, while large ambiguity calls for compatible reporting rather than an overconfident point claim.

14 Conclusion

  • Information set emulation reconstructs the prospective pre-treatment information set from EHR events through an AI typed lift before causal estimation.The lift organizes notes, orders, laboratory trajectories, encounters, and reports into a typed causal information ledger.Source: §14 · p. 40
  • Without sufficient evidence about source time, availability, and causal role, compatible prospective estimands can remain nonidentified; the squared information radius is the residual exact minimax mean squared error.Typing, chart review, availability adjudication, source separation, repeated extraction, or a prospective substudy can be assessed by how much they reduce ambiguity and corresponding risk.Source: §14 · p. 40
  • When the radius is small enough for scientific tolerance, standard causal estimation can use certified information; when it remains large, report a compatible set and missing-information diagnosis.The promise of approaching prospective-study reasoning is conditional on the remaining radius being small enough for the target claim.Source: §14 · p. 40

Data and code availability

  • The paper uses theory and artificial simulations, with scripts and synthetic Phase 0 materials supplied for reproducibility rather than real-patient validation.A real EHR or clinical NLP case study would require separate data access, ethics approval or exemption, and site-specific annotation review.Source: Data and code availability · pp. 40–41

A Artificial data generating mechanisms

  • Artificial stress tests vary extraction error, post-index leakage, certificate role misclassification, availability error, and treatment positivity.The simulations define the target using latent clinician judgment among patients with S_i = 1, and the generated EHR frame omits that judgment.Source: A Artificial data generating mechanisms · pp. 41–42

B Certificate and reporting templates

The appendices specify certificate fields and a reporting-set calculator, then give an operational bound for observation sensitivity. The calculator preserves disconnected sets but relies on calibration conditions; the observation bound requires an explicit bounded-outcome or clipped-target assumption.

B Certificate and reporting templates

  • The certificate template records feature identity, source, value, clinical and recording times, availability, treatment and outcome relations, roles, version, seed rule, and reviewer status.It also records compatible and certified roles and failed predicates.Source: B Certificate and reporting templates · p. 42
  • The reporting calculator combines role-specific estimates, standard errors, and validation-calibrated sensitivity envelopes for one locked scalar target.Its simultaneous image-coverage guarantee requires separate sampling and uniform validation conditions; the calculator does not establish them.Source: B Certificate and reporting templates · p. 42
  • The calculator uses cS = z1−αS/(2m), with m = |C(K)| and αS + αV ≤ α, and preserves disconnected interval unions.Every nonzero bound should be linked to evidence such as chart review, repeated extraction audit, overlap diagnostics, or observation-process analysis.Source: B Certificate and reporting templates · p. 42

C A conservative MNAR observation bound

  • The observation sensitivity bound assumes a bounded outcome scale, or a prespecified clipped outcome that defines the reported target.It bounds arm-specific mean change when an unmeasured observation mechanism tilts the observed-outcome distribution by a density ratio between exp(−Γa) and exp(Γa).Source: C A conservative MNAR observation bound · pp. 42–43
  • For a treatment contrast, the arm-specific bounds can be added to the observation component of the reporting bound.This bound is not a replacement for a problem-specific missingness analysis.Source: C A conservative MNAR observation bound · pp. 42–43
  • For unbounded outcomes, a plausible range alone does not establish the observation bound; a valid support assumption, clipped target, or separate tail or moment bound is needed.The purpose is to keep the unverified observation assumption visible on the estimand scale, not to make MNAR harmless.Source: C A conservative MNAR observation bound · p. 43
Loading 2609.17777v1…