Statistical Analysis Plan: What Goes In It, a Worked Entry, and the Mismatches Reviewers Catch

Here is how one entry reads when it is finished. The trial is hypothetical: two groups, a symptom score at 12 weeks as the primary outcome, sized at 63 per group and 75 recruited to allow for 15% loss.

Primary outcome. Symptom score (0 to 100, higher is worse) at 12 weeks after randomisation.

Estimand. The difference in mean 12-week score between groups among all randomised participants, regardless of adherence to the assigned treatment.

Population. All randomised participants, analysed in the group to which they were assigned.

Model. Linear regression of the 12-week score on treatment group, adjusted for baseline score and recruiting site (the stratification factor). We will report the adjusted mean difference with its 95% confidence interval and two-sided p-value.

Missing data. Multiple imputation by chained equations under a missing-at-random assumption, 50 imputed datasets, with imputation models including treatment group, site, baseline score and the 4- and 8-week scores. Sensitivity analysis: a delta-adjustment tipping-point analysis, shifting imputed values in the intervention group by 2, 4, 6 and 8 points.

Secondary outcomes. Tested in a fixed hierarchy (function score, then quality of life, then rescue medication use), each only if the previous comparison is significant at 0.05.

Subgroups. Two, pre-specified: baseline score above or below the median, and sex. Each assessed with a treatment-by-subgroup interaction term and reported as exploratory.

Every sentence closes a door. With the model named, nobody can quietly drop the site adjustment. With the hierarchy written down, a significant result for the third secondary outcome means nothing if the first one failed. With only two subgroups declared, a third cannot appear in the abstract as a finding.

Write it before anyone sees outcome data split by group. The Trials editorial notes that for many simple trials the full plan can be settled when the protocol is written, while adaptive designs or new methods may take longer. Our view is stricter on one point: whatever the design, the plan has to be final before unblinding. Its job, as the editorial describes it, is to reduce the occurrence of bias from selective analysis and reporting and to make it easier to detect.

Trials encourages SAPs in any of four forms: a section inside the protocol, an appendix to the protocol, an addendum to an already published protocol, or a standalone article with its own DOI. Whichever you use, keep the dated version you finalised before unblinding. When a reviewer asks to see the plan, the date on it is the first thing that makes it credible.

Two extensions exist or are coming. Early phase trials (phase I and non-randomised phase II) got their own version in 2022, which modified 25 of the original items and added 11. A version for cluster randomised trials is in development, and its published protocol lists the complications individually randomised guidance misses: accounting for clustering, imbalance from recruiting participants after clusters are randomised, and small-sample corrections when there are few clusters.

A published protocol or a registry entry gives reviewers something to compare your paper against. These are the gaps that turn a methods question into a credibility problem.

The primary outcome changed. The registry says 12-week score; the paper leads with the 8-week score because it was significant. CONSORT 2010 item 6b asks for "any changes to trial outcomes after the trial commenced, with reasons." A change with a reason is survivable. A change with no mention is the one reviewers remember.

Exploratory analyses dressed as planned ones. CONSORT item 18 asks authors to report other analyses "distinguishing pre-specified from exploratory." A subgroup result in the abstract that appears nowhere in the plan reads as a search for significance, even when it was not.

The model drifted. The plan adjusts for site and baseline; the paper adds three covariates and drops site. Report the planned model first, then the alternative with the reason.

Missing data handled differently from the plan. Planned multiple imputation, reported complete-case, with no sensitivity analysis. Reviewers who read methods for a living check this line.

Multiplicity promised, then forgotten. The plan names a testing hierarchy; the results table reports every secondary outcome with an unadjusted p-value and the discussion leans on the one that crossed 0.05.

Manusights cannot see your SAP unless it is in the manuscript, so it cannot tell you whether the analysis matches a plan it has not read. What it can do is check the paper for the symptoms above. The $99 Submission-Ready Dossier includes a statistical audit that takes every reported t, F, chi-square, correlation or z statistic with its degrees of freedom, recomputes the p-value, and flags any that disagree with the one printed; a second pass reviews the methodology for missing sample-size justification, uncorrected multiple comparisons, missing effect sizes or confidence intervals, a test unsuited to the data, and signs of selective reporting. It returns in about one to two hours.

It does not replace a trial statistician. Choosing an estimand, a missing-data assumption or a hierarchy for secondary outcomes is a design decision for your team, and it has to be made before the data exist.