What Goes in a Statistical Analysis Plan, and What Reviewers Flag

A statistical analysis plan (SAP) translates protocol commitments into analyses that can be reproduced before unblinding. The protocol describes objectives, design, eligibility and endpoints; the SAP specifies estimands, analysis populations, models, missing-data procedures and reporting rules. That is the core of SAP vs protocol: protocol amendments change how the trial is run, while SAP amendments change how it is analysed, and both need documented approval.

Throughout this article, a hypothetical antihypertensive trial, with change in systolic blood pressure from baseline to Week 12 as the primary endpoint, illustrates each decision. The protocol names the outcome and time point; the SAP states the change score, population, covariates and contrast.

The statistical analysis plan components below are those reviewers check first.

Objectives, hypotheses and populations. State the objectives, endpoint hierarchy, treatment comparison, significance level, power and assumptions, and align the hypothesis with the primary estimand. Define populations operationally: under intention-to-treat (ITT), participants are analysed as randomised; a full analysis set (FAS) may approximate ITT with justified exclusions; the per-protocol set applies predefined adherence criteria; the safety set usually includes dosed participants, analysed by treatment received. Define protocol deviations before unblinding.

Endpoint analyses. Specify each endpoint’s method, treatment effect, covariates, contrast, assumptions and reporting (Table 1). For the Week 12 example, ANCOVA models baseline-adjusted change and reports the adjusted mean difference with its confidence interval. A mixed model for repeated measures (MMRM) includes treatment, visit, their interaction and baseline; prespecify the covariance structure and a fallback if the model fails to converge, rather than choosing a structure after seeing results.

An SAP in clinical trials must say how post-randomisation events affect the treatment effect. ICH E9(R1) describes an estimand through five attributes: population, treatment, variable (endpoint), handling of intercurrent events and population-level summary [3].

Suppose a participant stops treatment because of an adverse event before Week 12. Under a treatment-policy strategy, the Week 12 value is used regardless of discontinuation. Under a hypothetical strategy, the analysis estimates the outcome as if discontinuation had not occurred (Figure 1). The two answer different clinical questions, so the SAP should justify its choice.

Discontinuation is an intercurrent event; a missing Week 12 measurement is a missing-data problem. For repeated measures, MMRM is a common primary analysis under a missing-at-random assumption. If multiple imputation is used, document the imputation model, variables and pooling. Sensitivity analyses, such as delta adjustment or a tipping-point analysis, should test departures from the primary assumption [1, 3].

Align the sample-size calculation with the primary estimand and analysis. For continuous endpoints, state the target difference, standard deviation, significance level, power, allocation ratio and expected dropout; for time-to-event endpoints, the target hazard ratio, required events, accrual and follow-up. For interim analyses, define timing, information fraction, stopping boundaries and how alpha is spent, plus data monitoring committee access to unblinded data [2]. For safety, define treatment-emergent adverse events, the coding dictionary and version, the observation period and denominators [2].

Define analysis datasets and derivations. ADSL holds subject-level information such as treatment and population flags; Basic Data Structure (BDS) datasets support parameter- and visit-level analyses through variables such as PARAMCD, AVISIT, BASE and CHG. For the Week 12 change, trace the source assessment, baseline selection and CHG derivation to the model and the output table. Our guide to SDTM vs ADaM works through a baseline example, and TLFs in clinical trials shows how outputs map back to the SAP.

Vague or post-hoc specifications. “Appropriate tests will be applied” cannot be executed. Name the model, covariates, contrast and diagnostics, and define subgroups before comparative results are seen (Table 2) [1].

Protocol–SAP discrepancies. Reconcile endpoints, assessment times, populations or sample-size assumptions that differ, decide whether a protocol amendment is needed, and keep dated approvals.

Weak missing-data, multiplicity or sensitivity strategies. Excluding participants with missing outcomes can bias results. Where there are several confirmatory endpoints, define the testing families and control the overall Type I error rate (family-wise error rate) with a prespecified method, such as Bonferroni, Holm, fixed-sequence (hierarchical) testing or gatekeeping [4].

Incomplete programming and QC rules. Unclear derivations, visit windows, censoring or denominators produce inconsistent results. Specify independent programming and checks across the SAP, ADaM, TLF shells and clinical study report.

ICH E9 expects the SAP to be finalised before the blind is broken, with the dates of finalisation and unblinding recorded [2]. Gamble and colleagues recommend writing it early, before data are collected, with the blind review as the last chance to amend it [1]. A blind review may reveal data-quality issues, but it must never be used to choose methods based on treatment differences. Record every amendment with its rationale and date.

A statistical analysis plan template is only useful when every decision is explicit. Use Table 3 as a final check. Gamble and colleagues’ 55-item recommendations give a content framework rather than a regulatory checklist [1]; SPIRIT 2025 [5] and CONSORT 2025 [6] guide the protocol and the report, and neither replaces a full SAP.

A well-specified statistical analysis plan makes a trial’s analysis transparent and reproducible. Prespecifying estimands, models, missing-data strategies, multiplicity control and programming rules answers most reviewer questions before unblinding, rather than after.