stats-reporting-audit

Nature Statistics Reporting Skill

Use this skill to make manuscript statistics transparent, reproducible, and appropriately bounded. It is a reporting and review skill, not a substitute for a statistician reanalysing raw data unless the user supplies the data and explicitly asks for computation.

Default stance

• Prioritize design transparency over decorative statistical language.
• Separate three questions: what was measured, what unit was analysed, and what inference was claimed.
• Treat the independent experimental unit as the default n; do not silently treat cells, fields of view, repeated readings, spectra, model runs, or technical replicates as independent biological or experimental samples.
• Prefer effect sizes, uncertainty intervals, sample sizes, and exact test definitions over significance-only phrasing.
• State missing information as AUTHOR_INPUT_NEEDED instead of inventing sample sizes, tests, software, corrections, exclusion rules, randomization, or blinding.
• If a journal-specific instruction, study-type guideline, or field standard conflicts with this skill, follow the more specific source and mark the source used.

Accepted inputs

The skill may receive:

• a Statistical analysis / Methods subsection
• Results paragraphs containing test statistics or p values
• figure panels, legends, captions, or source-data notes
• reviewer comments about statistics
• author notes in Chinese or English
• tables of reported comparisons
• raw or summary data, only when the user wants a concrete reanalysis or figure-statistics check

If the input is partial, run a bounded audit and state which parts cannot be assessed.

Workflow

1. Classify the task. Decide whether the user wants audit, rewrite, draft, reviewer-response support, figure-statistics alignment, or data-backed reanalysis.
2. Extract the design. Identify groups, treatments, time points, endpoints, blocking factors, repeated measures, randomization, blinding, exclusions, and missing-data handling.
3. Define n and replication. Separate independent experimental units, biological replicates, technical replicates, repeated measures, cells/fields/subsamples, simulations, and pooled observations.
4. Map claims to analyses, both ways. For each result claim, record the comparison/model, test family, assumptions, correction strategy, effect estimate, uncertainty, and exact p-value policy. Then map the other way: every test, statistic or interval that Statistical analysis declares must correspond to a result the paper reports. A declared-but-unused statistic (for example a paired t declared for every comparison while the text reports only mean differences and sign consistency across seeds) is removed from the Methods, not added to the Results.
5. Check common failure modes. Use references/common-failure-modes.md when the text involves nested data, many comparisons, cell-level measurements, interaction claims, correlations, regression, outliers, small samples, or significance-only reasoning.
6. Check reporting completeness. Use references/statistical-reporting.md to verify that Methods and Results give enough information for readers and reviewers to understand the analysis.
7. Align figure statistics. Use references/figure-statistics.md when figure legends, panel labels, stars, error bars, box plots, violin plots, source data, or supplementary figure notes are involved.
8. Draft or revise. Produce ready-to-paste text whose claims stay within the supplied design and evidence. Do not upgrade statistical association into mechanism or causality. Report precision as numbers (interval, n), not as a caveat sentence; each reporting statement (unit and n, P-value policy, multiplicity, blinding, exclusions) appears once, in Statistical analysis. For descriptive designs without inferential tests, use the descriptive variant in references/statistical-reporting.md.
9. Run final QA. Use references/reviewer-checklist.md before final delivery for severity labels, unresolved author questions, and reviewer-facing risk.