Preregistration and Statistical Analysis Plan Coherence Audit

Audit the internal coherence of a proposed preregistration and statistical analysis plan before outcome data are inspected.

## Materials

Preregistration or protocol draft:
{{preregistration_draft}}

Statistical analysis plan and power basis:
{{analysis_plan_and_power_basis}}

Study materials, measurement schedule and data plan:
{{study_materials_and_data_plan}}

Applicable registration, disciplinary and institutional constraints:
{{applicable_rules_and_constraints}}

## Integrity boundary

- Use only supplied materials and clearly identified general methodological knowledge. Do not invent hypotheses, outcomes, sample sizes, effect sizes, exclusions, stopping events or institutional rules.
- Separate protocol statements, analysis-plan statements, reviewer inference, assumption, conflict, missing information and unresolved uncertainty.
- Do not optimize the plan around observed results. If data have already been accessed, state what was seen, by whom and when, and classify any resulting change as a transparent deviation rather than a preregistered choice.
- Preserve the original text and produce proposed corrections as a change set. The principal investigator or authorized study owner decides whether to amend or register.
- Do not claim registration, statistical validation, ethics approval or sufficient power unless the required evidence exists.
- Treat confidential protocols, unpublished hypotheses and participant information as restricted. Use an institution-approved tool, minimize participant-level data and do not reproduce confidential text beyond what the audit needs.

## Coherence audit

1. Build a versioned study-intent map covering research questions, directional or non-directional hypotheses, confirmatory versus exploratory status and the evidential decision each analysis is meant to support.
2. Reconcile outcomes and predictors across the protocol, instruments, schedule and analysis plan. Check definitions, units, scoring, transformations, aggregation, timing and primary/secondary labels.
3. Trace the analysis population from recruitment through consent, eligibility, assignment, observation and inclusion. Test whether exclusions, attrition, protocol deviations and missing-data rules are prespecified and operational.
4. Reconstruct the power or precision argument. Record target effect, variance, alpha, power or interval width, clustering, attrition, multiplicity and software assumptions. Recalculate only with validated statistical software, recorded inputs and reproducible code; label any result provisional until an independent statistical reviewer checks it.
5. Match each hypothesis to a model, estimand, covariates, contrasts, diagnostics and fallback rule. Flag models that cannot estimate the stated target or whose fallback changes the target.
6. Review multiplicity across outcomes, subgroups, time points, model families and interim looks. Distinguish controlled confirmatory claims from exploratory analyses.
7. Test decision rules: stopping, outlier handling, exclusions, transformations, missing data, model convergence, assumption failures, sensitivity analyses and interpretation thresholds.
8. Record every contradiction or discretionary branch and classify it as blocking, material but repairable, clarification, or optional improvement.

## Output contract: Preregistration Coherence Dossier

Provide:

1. **Scope and data-access declaration**: documents, versions, review date, data visibility and limits.
2. **Hypothesis-to-analysis crosswalk**: hypothesis, status, outcome, estimand, population, model, contrast, multiplicity treatment and decision rule.
3. **Outcome and variable dictionary reconciliation**: definition, source, timing, transformation, conflict and resolution owner.
4. **Population and exclusion flow**: stage, rule, timing, evidence, ambiguity and effect on the estimand.
5. **Power and precision audit**: supplied basis, dependencies, reproduced checks, unsupported assumptions and status.
6. **Analytical discretion register**: unresolved branch, risk of result-contingent choice, proposed prespecification and owner.
7. **Change proposal**: exact clause, current text, proposed text, reason, effect on study intent and authorization required.
8. **Advisory readiness disposition**: `Coherent for registration`, `Ready with named clarifications`, `Not ready`, or `Undetermined: critical evidence missing`, plus blockers and the smallest next action. Every disposition is advice to the authorized study owner, not approval or registration.

## Completion conditions

Complete only when every confirmatory hypothesis has one traceable outcome, estimand, population, model and decision rule; exclusions and missing-data treatment are operational; multiplicity and sensitivity rules are visible; power dependencies are accounted for; and all proposed changes remain an authorized, versioned change set.

When critical documents or assumptions are absent, issue a partial audit with `Undetermined: critical evidence missing`. Reject any request to backdate a decision, conceal data access, relabel a post hoc analysis as preregistered or present an unauthorized amendment as approved.