Agent Skills
› brycewang-stanford/Awesome-Journal-Skills
› lancet-statistics
lancet-statistics
GitHub用于规范《柳叶刀》临床统计报告，确保提供置信区间而非仅P值，明确ITT主要分析人群，控制多重性，谨慎处理亚组交互作用，并规范缺失数据处理及效应量报告。
Trigger Scenarios
结果报告P值但缺少置信区间或效应量
主要分析人群（ITT与符合方案集）不明确
多个次要终点或亚组未处理多重性或交互作用
涉及统计分析员审查且分析未预先指定
Install
npx skills add brycewang-stanford/Awesome-Journal-Skills --skill lancet-statistics -g -y
SKILL.md
Frontmatter
{
"name": "lancet-statistics",
"description": "Use to enforce The Lancet's clinical-statistics reporting — confidence intervals over bare P values, a pre-specified primary analysis, intention-to-treat with per-protocol sensitivity, multiplicity control for secondary endpoints, cautious pre-specified subgroup analyses with interaction tests, missing-data handling, and absolute plus relative effects."
}
Clinical Statistics (lancet-statistics)
When to trigger
- Results report P values but not confidence intervals or effect sizes.
- The primary analysis population (ITT vs per-protocol) is unclear.
- Many secondary endpoints or subgroups are reported with no multiplicity/interaction handling.
- A statistical reviewer is likely (The Lancet uses them) and the analysis is not pre-specified.
The reporting backbone (every clinical estimate)
Each estimate needs: effect measure + 95% CI + the analysis population + the pre-specified test.
- Confidence intervals for every key estimate — CIs over bare P values; The Lancet de-emphasises isolated significance.
- Exact P values (e.g., p=0·013), not "p<0·05," reported alongside CIs, not instead of them.
- Absolute and relative effects: report the absolute risk/rate difference and the relative measure (RR/OR/HR); add NNT/NNH where clinically relevant.
- Analysis population stated: ITT as the primary population for superiority RCTs.
- The test/model named and matched to outcome type (survival → Cox/log-rank; binary → logistic/risk; count → Poisson/negative binomial; clustered → mixed/GEE).
Pre-specification and the primary analysis
- The primary analysis is pre-specified in the SAP (see
lancet-study-design) before unblinding. - Report the pre-specified primary analysis first; clearly label any post-hoc analysis as post-hoc.
- For non-inferiority/equivalence trials: state the margin, justify it, and analyse both ITT and per-protocol (per-protocol matters here).
ITT and sensitivity analyses
- ITT primary; per-protocol as a pre-specified sensitivity analysis (especially for non-inferiority).
- Report missing data explicitly: amount, pattern, and the handling method (e.g., multiple imputation), with sensitivity analyses under different assumptions. Do not silently use complete-case only.
Multiplicity and secondary endpoints
- Control for multiplicity across multiple primary/secondary endpoints (hierarchical testing, alpha-splitting, or correction) and state the strategy.
- Secondary endpoints are supportive, not confirmatory, unless pre-specified within a multiplicity-controlled hierarchy — say so.
Subgroups — pre-specified and cautious
- Report only pre-specified subgroups; label any post-hoc subgroup as exploratory.
- Test the interaction (treatment × subgroup), not separate within-subgroup P values — and interpret cautiously; subgroup claims are a classic over-interpretation trap.
- Present subgroups in a forest plot with interaction P values (see
lancet-figures-tables).
Survival and time-to-event
- Kaplan–Meier estimates with numbers at risk; hazard ratios with 95% CI from Cox models; state the proportional-hazards assumption check.
What the Lancet statistical reviewer expects
The Lancet routinely assigns an independent statistical reviewer whose report is a distinct, high-priority track. They read for pre-specification first (primary analysis matching the registered outcome and the SAP filed before unblinding), expect confidence intervals to carry the inference with exact P beside them, and want the absolute effect reported beside the relative one. Within-subgroup P values and per-protocol-as-primary for a superiority trial draw the sharpest scrutiny. Confirm current expectations against the journal's author guidelines.
Worked micro-example (illustrative numbers — not real data)
A hypothetical superiority RCT, binary primary outcome (response at 12 weeks), ITT population.
Primary (illustrative, ITT): response 612/1 020 (60.0%) vs 510/1 016 (50.2%)
Absolute risk difference 9.8 pp (95% CI 5.4-14.2); RR 1.20 (95% CI 1.10-1.30)
Exact p=0.0003; NNT ~10 (illustrative)
Sensitivity (per-protocol): RD 12.1 pp (95% CI 7.3-16.9) -> consistent with ITT.
Subgroups: 5 pre-specified; forest plot, interaction p=0.28 (no heterogeneity).
The CI and absolute risk difference carry the inference, the NNT makes the magnitude clinical, and the single interaction P keeps the subgroups honest.
Reviewer-pushback patterns and the venue-specific fix
- "P values but no effect sizes with CIs." → Add the effect with a 95% CI for every key estimate; keep exact P beside, not instead of, the interval.
- "Primary analysis differs from the registered outcome." → Reconcile to the registered outcome and SAP, or explain the dated change; label anything post-hoc.
- "Subgroup claims rest on within-subgroup P values." → Replace with interaction tests on pre-specified subgroups; interpret cautiously.
Output format
【Per-estimate backbone】 effect measure + 95% CI / analysis population / test → list gaps
【CIs over bare P?】 yes/no 【Exact P reported?】 yes/no
【Absolute + relative effects (NNT where relevant)?】 yes/no
【Primary analysis pre-specified + ITT?】 yes/no
【Missing data handled + sensitivity?】 yes/no
【Multiplicity strategy for secondary endpoints?】 stated / missing
【Subgroups pre-specified + interaction tested + cautious?】 yes/no
【Survival: KM + numbers at risk + HR(CI)?】 yes/no/NA
【Next】 lancet-figures-tables
Anti-patterns
- Do not report P values without confidence intervals or effect sizes.
- Do not report only a relative effect (HR/OR/RR) when the absolute effect changes the clinical message.
- Do not present per-protocol as the primary analysis for a superiority trial.
- Do not mine subgroups: separate within-subgroup P values without an interaction test invite over-interpretation.
- Do not infer "no effect" from a non-significant result in an underpowered analysis.
Version History
- 1839142 Current 2026-07-05 14:01
