Start from the study design and estimand, then choose models and tests. Statistical significance cannot replace effect size, data quality, or scientific meaning.
Working Modes
plan: Define the question, experimental unit, primary endpoint, model, correction, and sensitivity analysis before analysis.analyse: After the user provides data, perform reproducible analysis and retain records of data processing and diagnostics.audit: Review existing statistical methods, results, tables, and figure legends.rewrite: When the facts are sufficient, generate paste-ready statistical methods or results text.review-response: Parse reviewers' statistical questions and provide verification paths and key points for conservative responses.
Complex clinical trials, regulatory analyses, or patient-level decisions must comply with the protocol, statistical analysis plan, and review by a professional statistician.
Design Questions That Must Be Answered First
- What are the scientific question and the primary estimand?
- What is the independent experimental unit, and
nhow is it defined? - How are biological replicates, technical replicates, subsamples, batches, and repeated measures nested?
- What are the primary and secondary endpoints, groups, time points, and covariates?
- How are allocation, randomization, blinding, inclusion/exclusion, missingness, and outliers handled?
- Which comparisons are prespecified, and which are exploratory?
When these facts are unclear, do not give a final test choice; use AUTHOR_INPUT_NEEDED。
Execution Workflow
- Build the design diagram. Draw the experimental units, hierarchy, pairing, repeated measures, batches, and time structure.
- Define the estimand. Specify the difference, ratio, slope, association, predictive performance, or time effect to be estimated and its target population.
- Audit the data. Record data types, units, missingness, ranges, duplicates, outliers, exclusions, and transformations; retain before-and-after counts.
- Choose the analysis strategy. Choose the model based on design, distribution, sample size, and estimand, not solely on normality tests. Read references/analysis-plan.md。
- Execution and diagnostics. Report model assumptions, residual/fit diagnostics, convergence, influential points, multiple comparisons, and sensitivity analyses.
- Interpret effects. Prioritize effect sizes, confidence intervals, and practical meaning, then report exact p-values.
- Align figures and tables. Ensure that the data hierarchy, error, asterisks, figure legends, and main text in the figures are fully consistent with the analysis. Read references/reporting-and-figures.md。
- Run a reporting audit. Perform on the statistical text
python scripts/reporting_audit.py <file> --context methods|results|legend。 - Deliver reproducibility information. Provide analysis code, software versions, random seeds, data dictionary, processing logs, and unresolved risks.
Experimental units, pseudoreplication, and common failures Read references/design-integrity.md。
Default Output
Statistical scope
- Mode / input / boundary:
- Scientific question and estimand:
- Independent unit and n:
- Design hierarchy:
Analysis specification
- Outcome / predictors / contrasts:
- Model or test:
- Assumptions and diagnostics:
- Multiplicity:
- Sensitivity analyses:
Results
- Effect estimate and uncertainty:
- Exact inferential result:
- Practical interpretation:
Ready-to-paste reporting
[Methods / Results / legend]
AUTHOR_INPUT_NEEDED
- [事实性缺口]
Reviewer-risk note
- [剩余风险]
Red Lines
- Do not fabricate sample sizes, p-values, degrees of freedom, intervals, power, software versions, exclusions, randomization, or blinding.
- Do not treat cells, fields of view, technical readings, simulation runs, or multiple measurements from the same individual as independent
n。 - Do not use "significant" to mean important, large, causal, or biologically relevant.
- Do not claim "no difference" or "equivalence" because p > 0.05, unless the design supports the corresponding inference.
- Do not infer between-group interaction from within-group significant/non-significant differences.
- Do not pursue significance by deleting data, changing endpoints, or trying multiple models and reporting only the best result.
- Do not package exploratory analyses as prespecified confirmatory analyses.
Material routing
| Task | Read |
|---|---|
| Experimental units, nesting, repeated measures, pseudoreplication, missingness, and exclusions | references/design-integrity.md |
| estimand, model selection, diagnostics, effect size, multiple comparisons, and sensitivity | references/analysis-plan.md |
| Methods, Results, tables, figure captions, and statistical graphics reporting | references/reporting-and-figures.md |