Lancet Statistics

---
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.

## 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.