A participant stops the assigned treatment at week 4 and starts rescue medication at week 6. At week 12, they come back and have their blood pressure measured. Does that reading count? The statistical analysis plan (SAP) settles that question before the trial team sees how the groups compare. Write it so two analysts working from the same raw data would make the same call.
Identify the trial, registration number, protocol version, allocation ratio, randomization unit, and any stratification factors. State whether the primary comparison tests superiority, noninferiority, or equivalence. For noninferiority or equivalence, give the prespecified margin and its justification or point to the relevant protocol section. Record the SAP version, authors, reviewers, and sign-off date.
Then define the primary estimand, the treatment effect the trial aims to estimate. Write down the population, treatments being compared, outcome, strategy for events that affect the interpretation or existence of that outcome, and summary measure. Discontinuation, rescue medication, treatment switching, and death can raise different questions. Give each relevant event an explicit rule. ICH E9(R1) guidance on estimands.
In an illustrative two-arm trial, the primary estimand could be the difference in mean week-12 systolic blood pressure between the new treatment and usual care among eligible adults represented by the trial, regardless of discontinuation or rescue medication. Under that treatment-policy strategy, the week-12 reading in the opening example counts. The follow-up schedule should therefore call for collecting that reading even after assigned treatment stops. An event such as death would require its own strategy because the week-12 measurement would be undefined.
In this trial, week 12 could mean days 77 through 91 after randomization. Use the visit closest to day 84. If two visits are equally close, take the earlier one. With that rule, the day-83 visit wins over the day-85 visit. Give the units and explain how to derive or score the outcome, including how to combine several readings at one visit and when to discard an invalid measurement.
Write executable inclusion rules for each analysis population. For an intention-to-treat efficacy analysis, specify how all randomized participants enter the analysis and retain their assigned groups. State separately how the analysis uses participants with missing outcomes. If you plan a per-protocol analysis, define major deviations and adherence thresholds up front and explain how selection after randomization may affect interpretation. Define the safety population and the grouping used for safety summaries, such as by treatment received.
For each confirmatory outcome, specify the model, covariates and coding, effect measure, confidence interval, significance level, and decision rule. Account for clustering or repeated measurements when the design calls for it. In the blood pressure example, a prespecified linear regression could use week-12 pressure as the outcome, with assigned treatment, baseline pressure, and randomization strata as predictors. Report the adjusted mean difference in mm Hg, calculated as new treatment minus usual care, with a 95% confidence interval. If lower pressure favors the new treatment, a superiority rule for this single primary comparison could require that the entire interval lie below zero. Specify variance estimation and any fallback rules before inspecting treatment results.
Document the sample-size rationale or reference the exact protocol section. Record the target effect, variance or event-rate assumptions, allocation, power, significance level, and allowance for missing outcomes. Check that the calculation matches the primary comparison and design. If the design uses a model whose power depends on baseline adjustment or clustering, account for those assumptions in the calculation.
The plan already says to use the participant's week-12 reading after rescue medication. If they miss the visit, that reading is missing under the treatment-policy strategy. Plan to show how often this happened in each arm, at which visits, and for what reasons. State what the primary analysis assumes about the readings you never got.
For example, if the main analysis uses multiple imputation under a missing-at-random assumption, explain why the observed information could account for differences between participants with and without a week-12 reading. Name the variables in the imputation model, including treatment group and relevant earlier measurements, and specify the number of imputations and how estimates will be combined. Then plan a sensitivity analysis for the same estimand. A prespecified range of shifts to imputed blood pressure values could show how much worse the missing outcomes would need to be to change the primary conclusion. Set the range using clinical judgment before reviewing the comparison. ICH E9(R1) guidance on sensitivity analysis.
List the hypotheses that can support confirmatory claims. When several endpoints or comparisons share a false-positive error budget, specify the family and the adjustment or testing order. For subgroup analyses, name the subgroup definitions and the treatment-by-subgroup interaction you will estimate. Label exploratory analyses in the plan and in the report. If there are interim analyses, specify when they occur, who sees the results, what can trigger early stopping, and how repeated looks affect the final decision rule.
For safety, define the observation window, what counts as a treatment-emergent event, how events will be grouped and graded, and the denominators for each summary. Distinguish the number of participants with at least one event from the total event count. Prepare mock tables for participant flow, the primary and secondary results, missing outcomes, and harms. The headings and footnotes should make each denominator and comparison explicit. Gamble and colleagues' clinical trial SAP checklist.
Ask a second analyst to read the primary analysis section and populate a mock results table using synthetic data. Include a missed visit, an out-of-window reading, and a participant who stopped treatment. Record the software, analysis dataset version, and code needed to reproduce the outputs. Set a sign-off deadline before anyone making analysis decisions sees unblinded comparative results. If an amendment is needed, record what changed, why, when, and what data were visible at the time. Disclose departures from the plan in the trial report.
Before signing off, rerun the sample-size assumptions for the comparison you actually plan to report. Use the StatPower power planner to explore how the target effect changes recruitment needs for supported designs. Use a design-specific calculation for features the calculator does not represent, such as clustering or repeated measures.
One email whenever we publish something new. No spam, unsubscribe anytime.
{{ subscribeErrorMsg }}