Перейти к содержанию

Статистика · Средний

Template 4: Statistical analysis plan

Write this before examining the relevant outcomes whenever possible. If the plan is retrospective, state that fact and distinguish decisions influenced by observed results.

3 мин чтения · 26 окт. 2026 г.

Эта страница ещё не переведена, поэтому вы читаете английскую версию.

Главное

  • Adapt the prompts to the actual study.
  • Complete every statement from verified records.
  • Follow the target journal and institutional requirements.
Содержание

Write this before examining the relevant outcomes whenever possible. If the plan is retrospective, state that fact and distinguish decisions influenced by observed results.

Write this before examining the relevant outcomes whenever possible. If the plan is retrospective, state that fact and distinguish decisions influenced by observed results.

Question and data structure

Primary estimand: [Population, outcome, comparison and summary]

Independent unit: [Definition]

Pairing, clusters or repeated measures: [Structure and identifiers]

Primary outcome: [Variable, units, timing, derivation and direction]

Secondary outcomes: [Definitions and priority]

Analysis population: [Who or what is included, with rationale]

Preparation

Integrity checks: [Duplicates, impossible values, units and identifier checks]

Transformations: [Formula, rationale and interpretation]

Exclusions: [Prespecified criteria and decision process]

Missingness: [Patterns to report, assumptions, method and sensitivity checks]

Descriptive displays: [Plots and summaries appropriate to the outcome]

Primary analysis

Model or test: [Exact method and contrast]

Covariates: [Variables, coding and reason for inclusion]

Dependence handling: [Pairing, cluster or repeated measure method]

Assumptions and diagnostics: [What will be assessed and how]

Uncertainty: [Interval level and calculation method]

Hypothesis test, if used: [Null, alternative, direction and threshold]

Meaningful effect: [Scientific or practical criterion, if defensible]

Multiple comparisons

Family 1: [Complete set of outcomes or contrasts]

Error objective and method: [Family error control, FDR or another justified objective]

Additional families: [Scientific rationale for each]

Interval interpretation: [Individual or simultaneous intervals]

Exploratory analyses: [Clearly identified and reported as exploratory]

Robustness and reporting

Sensitivity analyses: [Alternative; reason; expected interpretation]

Stopping rule: [Fixed sample or justified sequential procedure]

Software and versions: [Actual tools]

Reproducibility record: [Inputs, script entry point, outputs and environment]

Reporting sentence: “We estimated [quantity] using [method] in [n and unit], with [interval]. We handled [dependence and missing data] by [method] and applied [multiplicity method] to [family].”

Deviation log: [Link or location for Template 16]

Do not choose the analysis by cycling through tests until one yields the desired conclusion. If the appropriate method is uncertain, record the uncertainty and seek methodological input before interpreting results.

Tutorial: complete the analysis plan in the right order

Read Articles 14 to 18 alongside the individual analysis guides. First define the quantity you want to estimate. Next describe the structure of the observations. Only then choose the model, uncertainty calculation and any hypothesis test. A distribution plot cannot reveal that several rows come from the same participant; the design record must supply that information.

Worked planning example using the supplied synthetic controller data: The outcome is positioning error in millimetres. The comparison is A minus B within each constructed layout case. The teaching estimand is the mean paired difference; a positive value means lower B error. The twelve case differences, not twenty four independent controller values, are the input to the paired calculation. A real protocol must additionally justify why its cases represent the intended evaluation setting. See Article 32.

Write the interpretation before seeing results: “We will discuss the estimated difference in millimetres and its uncertainty against a practically meaningful criterion that is justified independently of the observed effect.” If there is no defensible criterion, say that practical importance will remain uncertain. Do not invent a threshold after seeing which one the result exceeds.

Completion check: Explain what you would report if the interval includes zero, if it excludes zero but includes negligible effects, and if a sensitivity analysis changes the conclusion. Each is a reporting problem to plan for, not a reason to hide the result.