WHY ENGLISH?
Decide how evidence will be analysed before seeing the answer
Use five routes to align the study question, define analysis data, specify methods, test robustness, and produce traceable outputs under qualified statistical and clinical governance.
Open the full contents · See the How English Works hub
Full contents
Align question
Define data
Specify method
Test robustness
Produce evidence
Practice and next steps
Practice and next steps
Practice and next steps
Practice and next steps
Practice and next steps
Practice and next steps
A clinical study statistical analysis plan, or SAP, explains how study data will be turned into estimates, uncertainty and conclusions. English matters because the objective, estimand, endpoint, analysis population, model, missing-data assumption, comparison, confidence interval and sensitivity analysis must describe one coherent clinical question before results influence choices.
People searching for statistical analysis plan, clinical trial SAP, prespecified analysis, estimand, missing data or clinical study tables figures and listings need more than a list of software commands. They need a decision document connecting protocol intent to reproducible analysis and transparent reporting.
The official ICH E9 Statistical Principles for Clinical Trials is the current Step 4 guideline dated 5 February 1998. It stresses prespecification, analysis sets, missing values, estimation, multiplicity, subgroups, software validity and reporting. The ICH E9(R1) addendum on estimands and sensitivity analysis reached Step 4 in 2019. Singapore's Health Sciences Authority clinical-trials overview explains the local regulatory framework. The protocol, current guidance, approvals and qualified statistician govern a real study.
Use the SAP map study, protocol, version, objective, estimand, hypothesis, endpoint, analysis set, treatment group, baseline, visit window, data cut, randomisation, blinding, sample size, model, contrast, interval, multiplicity, covariate, subgroup, missing data, intercurrent event, deviation, outlier, sensitivity, interim analysis, safety, output, programming, quality control, change and archive.
Prespecification does not mean pretending every future detail is known. It means documenting the important decision rules, assumptions and alternatives early enough that the data do not quietly choose the most flattering method. When change is necessary, the record should show timing, reason, approval and impact.
Identify the study
Align protocol number title phase sponsor and investigational product with the clinical question and protocol.
The analysis answers a different question when meaning can drift across versions, roles or evidence.
Define the decision and preserve the source, boundary, owner and decision path.
Question example. Study SG-26-014 remains consistent across protocol data and outputs.
Did You Know? ICH E9 treats statistics as part of trial design, conduct, analysis and evaluation, not a calculation added only after data collection. That is why protocol number title phase sponsor and investigational product should stay connected to its exact source and decision boundary.
Identify the SAP version
Turn version date status authors reviewers and approvals into an unambiguous analysis variable or population.
Reproducibility fails when meaning can drift across versions, roles or evidence.
Write an executable rule and preserve the source, boundary, owner and decision path.
Data example. Version 1.0 is final before unblinding.
A reviewer should be able to derive the same analysis set.
State the purpose
Specify how the SAP expands protocol statistical principles into executable detail with model, comparison and uncertainty.
Analytical discretion becomes hidden when meaning can drift across versions, roles or evidence.
Prespecify the method and preserve the source, boundary, owner and decision path.
Method example. The SAP does not silently change the clinical question.
Technical detail matters because small wording changes can change inference.
Map source documents
Test protocol amendments charter data plan and programming standards through a named robustness question.
Sensitivity analysis becomes decorative when meaning can drift across versions, roles or evidence.
Connect assumption to consequence and preserve the source, boundary, owner and decision path.
Robustness example. Protocol amendment 3 is the governing design source.
Alternative analyses should explain what uncertainty they probe.
Define objectives
Carry primary secondary exploratory and safety questions into validated outputs and reporting.
The evidence chain breaks when meaning can drift across versions, roles or evidence.
Trace code, data, decision and preserve the source, boundary, owner and decision path.
Output example. Each analysis traces to an objective rather than output habit.
A final table is credible when its origin and planned role remain visible.
Define estimands
Align population variable treatment condition intercurrent-event strategy and summary with the clinical question and protocol.
The analysis answers a different question when meaning can drift across versions, roles or evidence.
Define the decision and preserve the source, boundary, owner and decision path.
Question example. The treatment effect question stays explicit after discontinuation.
The SAP should make the intended treatment effect readable before coding starts.
Define hypotheses
Turn null alternative direction and testing framework into an unambiguous analysis variable or population.
Reproducibility fails when meaning can drift across versions, roles or evidence.
Write an executable rule and preserve the source, boundary, owner and decision path.
Data example. The superiority hypothesis is written before results are viewed.
A reviewer should be able to derive the same analysis set.
Define endpoints
Specify measure derivation timing window and clinical meaning with model, comparison and uncertainty.
Analytical discretion becomes hidden when meaning can drift across versions, roles or evidence.
Prespecify the method and preserve the source, boundary, owner and decision path.
Method example. Day-84 change uses the prespecified baseline and visit window.
Did You Know? ICH E9 treats statistics as part of trial design, conduct, analysis and evaluation, not a calculation added only after data collection. That is why measure derivation timing window and clinical meaning should stay connected to its exact source and decision boundary.
Define primary endpoint
Test the one central variable tied to the primary objective through a named robustness question.
Sensitivity analysis becomes decorative when meaning can drift across versions, roles or evidence.
Connect assumption to consequence and preserve the source, boundary, owner and decision path.
Robustness example. Primary outcome matches the sample-size assumption.
Alternative analyses should explain what uncertainty they probe.
Define secondary endpoints
Carry supportive variables and inferential role into validated outputs and reporting.
The evidence chain breaks when meaning can drift across versions, roles or evidence.
Trace code, data, decision and preserve the source, boundary, owner and decision path.
Output example. Key secondary endpoints have a stated hierarchy.
A final table is credible when its origin and planned role remain visible.
Define exploratory endpoints
Align hypothesis-generating analyses kept separate from confirmation with the clinical question and protocol.
The analysis answers a different question when meaning can drift across versions, roles or evidence.
Define the decision and preserve the source, boundary, owner and decision path.
Question example. Biomarker associations are labelled exploratory.
The SAP should make the intended treatment effect readable before coding starts.
Define safety variables
Turn events laboratories vital signs exposure and other measures into an unambiguous analysis variable or population.
Reproducibility fails when meaning can drift across versions, roles or evidence.
Write an executable rule and preserve the source, boundary, owner and decision path.
Data example. Treatment-emergent adverse events use an exact time rule.
A reviewer should be able to derive the same analysis set.
Define analysis populations
Specify randomised treated full analysis per-protocol and safety sets with model, comparison and uncertainty.
Analytical discretion becomes hidden when meaning can drift across versions, roles or evidence.
Prespecify the method and preserve the source, boundary, owner and decision path.
Method example. Every population has an inclusion rule and reason log.
Technical detail matters because small wording changes can change inference.
Define treatment groups
Test randomised actual received and rescue treatment handling through a named robustness question.
Sensitivity analysis becomes decorative when meaning can drift across versions, roles or evidence.
Connect assumption to consequence and preserve the source, boundary, owner and decision path.
Robustness example. Safety summaries use treatment received under the stated rule.
Alternative analyses should explain what uncertainty they probe.
Define baseline
Carry last qualifying value and handling of repeated measures into validated outputs and reporting.
The evidence chain breaks when meaning can drift across versions, roles or evidence.
Trace code, data, decision and preserve the source, boundary, owner and decision path.
Output example. Baseline is not chosen after seeing which value is favourable.
Did You Know? ICH E9 treats statistics as part of trial design, conduct, analysis and evaluation, not a calculation added only after data collection. That is why last qualifying value and handling of repeated measures should stay connected to its exact source and decision boundary.
Define visit windows
Align target dates allowable windows and tie-breaking with the clinical question and protocol.
The analysis answers a different question when meaning can drift across versions, roles or evidence.
Define the decision and preserve the source, boundary, owner and decision path.
Question example. An assessment between visits is assigned by the stated algorithm.
The SAP should make the intended treatment effect readable before coding starts.
Define data cut
Turn database snapshot and included follow-up into an unambiguous analysis variable or population.
Reproducibility fails when meaning can drift across versions, roles or evidence.
Write an executable rule and preserve the source, boundary, owner and decision path.
Data example. The analysis identifies the extract date and data version.
A reviewer should be able to derive the same analysis set.
Define randomisation
Specify strata blocks allocation information and analysis use with model, comparison and uncertainty.
Analytical discretion becomes hidden when meaning can drift across versions, roles or evidence.
Prespecify the method and preserve the source, boundary, owner and decision path.
Method example. Randomisation stratum is distinguished from recorded site data.
Technical detail matters because small wording changes can change inference.
Define blinding
Test who remains blinded and when access may occur through a named robustness question.
Sensitivity analysis becomes decorative when meaning can drift across versions, roles or evidence.
Connect assumption to consequence and preserve the source, boundary, owner and decision path.
Robustness example. Finalisation is documented before treatment codes are released.
Alternative analyses should explain what uncertainty they probe.
Define sample size
Carry effect assumption variability power alpha and allowance into validated outputs and reporting.
The evidence chain breaks when meaning can drift across versions, roles or evidence.
Trace code, data, decision and preserve the source, boundary, owner and decision path.
Output example. The SAP cites rather than reinvents the protocol calculation.
A final table is credible when its origin and planned role remain visible.
Define descriptive statistics
Align n mean SD median quartiles range counts and percentages with the clinical question and protocol.
The analysis answers a different question when meaning can drift across versions, roles or evidence.
Define the decision and preserve the source, boundary, owner and decision path.
Question example. Decimal precision is set before table production.
The SAP should make the intended treatment effect readable before coding starts.
Define primary model
Turn response distribution effects covariance and estimation method into an unambiguous analysis variable or population.
Reproducibility fails when meaning can drift across versions, roles or evidence.
Write an executable rule and preserve the source, boundary, owner and decision path.
Data example. The model formula names treatment baseline and stratification factors.
Did You Know? ICH E9 treats statistics as part of trial design, conduct, analysis and evaluation, not a calculation added only after data collection. That is why response distribution effects covariance and estimation method should stay connected to its exact source and decision boundary.
Define contrasts
Specify the exact comparisons and reference group with model, comparison and uncertainty.
Analytical discretion becomes hidden when meaning can drift across versions, roles or evidence.
Prespecify the method and preserve the source, boundary, owner and decision path.
Method example. Treatment minus control keeps one sign convention.
Technical detail matters because small wording changes can change inference.
Define confidence intervals
Test level method and interpretation through a named robustness question.
Sensitivity analysis becomes decorative when meaning can drift across versions, roles or evidence.
Connect assumption to consequence and preserve the source, boundary, owner and decision path.
Robustness example. A 95% interval is linked to the estimand and scale.
Alternative analyses should explain what uncertainty they probe.
Define multiplicity
Carry family hierarchy gatekeeping or adjustment method into validated outputs and reporting.
The evidence chain breaks when meaning can drift across versions, roles or evidence.
Trace code, data, decision and preserve the source, boundary, owner and decision path.
Output example. Key secondary testing begins only under the stated rule.
A final table is credible when its origin and planned role remain visible.
Define covariates
Align prespecified variables coding and rationale with the clinical question and protocol.
The analysis answers a different question when meaning can drift across versions, roles or evidence.
Define the decision and preserve the source, boundary, owner and decision path.
Question example. Baseline severity categories are frozen before analysis.
The SAP should make the intended treatment effect readable before coding starts.
Define interactions
Turn terms tests and cautious interpretation into an unambiguous analysis variable or population.
Reproducibility fails when meaning can drift across versions, roles or evidence.
Write an executable rule and preserve the source, boundary, owner and decision path.
Data example. Subgroup interaction differs from within-group significance.
A reviewer should be able to derive the same analysis set.
Define subgroups
Specify categories cut-points and purpose with model, comparison and uncertainty.
Analytical discretion becomes hidden when meaning can drift across versions, roles or evidence.
Prespecify the method and preserve the source, boundary, owner and decision path.
Method example. Age categories are clinically justified rather than data-mined.
Technical detail matters because small wording changes can change inference.
Define missing data
Test mechanism assumptions primary method and sensitivity through a named robustness question.
Sensitivity analysis becomes decorative when meaning can drift across versions, roles or evidence.
Connect assumption to consequence and preserve the source, boundary, owner and decision path.
Robustness example. Missing outcomes are not simply replaced without rationale.
Did You Know? ICH E9 treats statistics as part of trial design, conduct, analysis and evaluation, not a calculation added only after data collection. That is why mechanism assumptions primary method and sensitivity should stay connected to its exact source and decision boundary.
Define intercurrent events
Carry discontinuation rescue death switching and other post-randomisation events into validated outputs and reporting.
The evidence chain breaks when meaning can drift across versions, roles or evidence.
Trace code, data, decision and preserve the source, boundary, owner and decision path.
Output example. Each event follows the estimand strategy.
A final table is credible when its origin and planned role remain visible.
Define protocol deviations
Align categories importance review and effect on populations with the clinical question and protocol.
The analysis answers a different question when meaning can drift across versions, roles or evidence.
Define the decision and preserve the source, boundary, owner and decision path.
Question example. Major deviations are classified without treatment knowledge.
The SAP should make the intended treatment effect readable before coding starts.
Define outliers
Turn detection review and analysis treatment into an unambiguous analysis variable or population.
Reproducibility fails when meaning can drift across versions, roles or evidence.
Write an executable rule and preserve the source, boundary, owner and decision path.
Data example. An extreme value remains unless a prespecified reason excludes it.
A reviewer should be able to derive the same analysis set.
Define transformations
Specify scale formula back-transformation and presentation with model, comparison and uncertainty.
Analytical discretion becomes hidden when meaning can drift across versions, roles or evidence.
Prespecify the method and preserve the source, boundary, owner and decision path.
Method example. Log estimates are reported with interpretable ratios.
Technical detail matters because small wording changes can change inference.
Define repeated measures
Test time structure covariance and degrees-of-freedom approach through a named robustness question.
Sensitivity analysis becomes decorative when meaning can drift across versions, roles or evidence.
Connect assumption to consequence and preserve the source, boundary, owner and decision path.
Robustness example. The repeated model uses all specified visits under its assumptions.
Alternative analyses should explain what uncertainty they probe.
Define time-to-event analysis
Carry origin event censoring competing events and method into validated outputs and reporting.
The evidence chain breaks when meaning can drift across versions, roles or evidence.
Trace code, data, decision and preserve the source, boundary, owner and decision path.
Output example. Progression-free time uses one censoring convention.
A final table is credible when its origin and planned role remain visible.
Define binary analysis
Align responder rule missing handling model and measure with the clinical question and protocol.
The analysis answers a different question when meaning can drift across versions, roles or evidence.
Define the decision and preserve the source, boundary, owner and decision path.
Question example. Response requires every component in the prespecified window.
Did You Know? ICH E9 treats statistics as part of trial design, conduct, analysis and evaluation, not a calculation added only after data collection. That is why responder rule missing handling model and measure should stay connected to its exact source and decision boundary.
Define count analysis
Turn exposure time recurrent events and overdispersion into an unambiguous analysis variable or population.
Reproducibility fails when meaning can drift across versions, roles or evidence.
Write an executable rule and preserve the source, boundary, owner and decision path.
Data example. Event rate uses participant-time rather than raw counts alone.
A reviewer should be able to derive the same analysis set.
Define non-inferiority or equivalence
Specify margin direction analysis sets and interpretation with model, comparison and uncertainty.
Analytical discretion becomes hidden when meaning can drift across versions, roles or evidence.
Prespecify the method and preserve the source, boundary, owner and decision path.
Method example. The margin is justified clinically and statistically.
Technical detail matters because small wording changes can change inference.
Define sensitivity analyses
Test alternative assumptions testing robustness of the primary conclusion through a named robustness question.
Sensitivity analysis becomes decorative when meaning can drift across versions, roles or evidence.
Connect assumption to consequence and preserve the source, boundary, owner and decision path.
Robustness example. A tipping-point analysis explores departures from missing-at-random.
Alternative analyses should explain what uncertainty they probe.
Define supplementary analyses
Carry different questions from the main estimand into validated outputs and reporting.
The evidence chain breaks when meaning can drift across versions, roles or evidence.
Trace code, data, decision and preserve the source, boundary, owner and decision path.
Output example. Treatment-policy and hypothetical strategies are not conflated.
A final table is credible when its origin and planned role remain visible.
Define interim analysis
Align timing boundaries decision roles and confidentiality with the clinical question and protocol.
The analysis answers a different question when meaning can drift across versions, roles or evidence.
Define the decision and preserve the source, boundary, owner and decision path.
Question example. An independent committee receives only the approved outputs.
The SAP should make the intended treatment effect readable before coding starts.
Define early stopping
Turn efficacy futility safety and operational rules into an unambiguous analysis variable or population.
Reproducibility fails when meaning can drift across versions, roles or evidence.
Write an executable rule and preserve the source, boundary, owner and decision path.
Data example. A boundary crossing triggers governance rather than an automatic claim.
A reviewer should be able to derive the same analysis set.
Define data monitoring outputs
Specify closed and open reports recipients and access with model, comparison and uncertainty.
Analytical discretion becomes hidden when meaning can drift across versions, roles or evidence.
Prespecify the method and preserve the source, boundary, owner and decision path.
Method example. Unblinded tables stay within the authorised group.
Did You Know? ICH E9 treats statistics as part of trial design, conduct, analysis and evaluation, not a calculation added only after data collection. That is why closed and open reports recipients and access should stay connected to its exact source and decision boundary.
Define safety summaries
Test incidence severity relatedness seriousness and exposure through a named robustness question.
Sensitivity analysis becomes decorative when meaning can drift across versions, roles or evidence.
Connect assumption to consequence and preserve the source, boundary, owner and decision path.
Robustness example. Participant counts are not confused with event counts.
Alternative analyses should explain what uncertainty they probe.
Define laboratory summaries
Carry units reference ranges shifts and potentially important values into validated outputs and reporting.
The evidence chain breaks when meaning can drift across versions, roles or evidence.
Trace code, data, decision and preserve the source, boundary, owner and decision path.
Output example. Original units are converted through a validated rule.
A final table is credible when its origin and planned role remain visible.
Define tables figures listings
Align shell numbering titles footnotes denominators and sources with the clinical question and protocol.
The analysis answers a different question when meaning can drift across versions, roles or evidence.
Define the decision and preserve the source, boundary, owner and decision path.
Question example. Each output states population treatment and cut-off.
The SAP should make the intended treatment effect readable before coding starts.
Define programming standards
Turn software version code review validation and reproducibility into an unambiguous analysis variable or population.
Reproducibility fails when meaning can drift across versions, roles or evidence.
Write an executable rule and preserve the source, boundary, owner and decision path.
Data example. Independent review confirms the primary output.
A reviewer should be able to derive the same analysis set.
Define deviations from plan
Specify documentation timing rationale and impact with model, comparison and uncertainty.
Analytical discretion becomes hidden when meaning can drift across versions, roles or evidence.
Prespecify the method and preserve the source, boundary, owner and decision path.
Method example. A post-unblinding change is labelled and explained in the report.
Technical detail matters because small wording changes can change inference.
Define quality control
Test traceability review reconciliation and sign-off through a named robustness question.
Sensitivity analysis becomes decorative when meaning can drift across versions, roles or evidence.
Connect assumption to consequence and preserve the source, boundary, owner and decision path.
Robustness example. One endpoint value can be traced from source dataset to table.
Alternative analyses should explain what uncertainty they probe.
Define archiving
Carry approved SAP code datasets outputs logs and decisions into validated outputs and reporting.
The evidence chain breaks when meaning can drift across versions, roles or evidence.
Trace code, data, decision and preserve the source, boundary, owner and decision path.
Output example. The analysis package preserves the exact run environment.
Did You Know? ICH E9 treats statistics as part of trial design, conduct, analysis and evaluation, not a calculation added only after data collection. That is why approved SAP code datasets outputs logs and decisions should stay connected to its exact source and decision boundary.
Explain results responsibly
Align estimate uncertainty clinical relevance and limits with the clinical question and protocol.
The analysis answers a different question when meaning can drift across versions, roles or evidence.
Define the decision and preserve the source, boundary, owner and decision path.
Question example. A small p-value is not written as proof of practical benefit.
The SAP should make the intended treatment effect readable before coding starts.
Reconcile with the clinical study report
Turn planned analyses actual conduct results and deviations into an unambiguous analysis variable or population.
Reproducibility fails when meaning can drift across versions, roles or evidence.
Write an executable rule and preserve the source, boundary, owner and decision path.
Data example. The report cites the SAP version used for every primary result.
A reviewer should be able to derive the same analysis set.
Protect participant privacy
Specify analysis datasets outputs small cells and controlled access with model, comparison and uncertainty.
Analytical discretion becomes hidden when meaning can drift across versions, roles or evidence.
Prespecify the method and preserve the source, boundary, owner and decision path.
Method example. Listings omit unnecessary direct identifiers.
Technical detail matters because small wording changes can change inference.
Plan reproducible reruns
Test frozen inputs code dependencies seeds and execution logs through a named robustness question.
Sensitivity analysis becomes decorative when meaning can drift across versions, roles or evidence.
Connect assumption to consequence and preserve the source, boundary, owner and decision path.
Robustness example. A qualified reviewer can recreate the primary table.
Alternative analyses should explain what uncertainty they probe.
Record reviewer decisions
Carry statistical clinical programming and data-management resolutions into validated outputs and reporting.
The evidence chain breaks when meaning can drift across versions, roles or evidence.
Trace code, data, decision and preserve the source, boundary, owner and decision path.
Output example. A query log explains why one derivation changed before finalisation.
A final table is credible when its origin and planned role remain visible.
Teach the transferable skill
Align planning analysis before seeing answers with the clinical question and protocol.
The analysis answers a different question when meaning can drift across versions, roles or evidence.
Define the decision and preserve the source, boundary, owner and decision path.
Question example. Students can preregister how they will compare two study methods.
The SAP should make the intended treatment effect readable before coding starts.
A worked example
A protocol names a primary endpoint but leaves baseline, analysis population, missing data, model, contrasts and multiplicity open. After unblinding, several reasonable choices produce different results.
A strong SAP finalised under appropriate blinding expands the protocol into executable definitions, models, algorithms and output shells. It links every choice to the study question and plans sensitivity analyses for key assumptions.
Validated programming and traceability then connect source data to derived variables, results and tables. Any deviation from plan is dated and explained so readers can distinguish prespecified confirmation from later exploration.
A practical checklist
- Study protocol SAP version authors reviewers and approvals exact
- Objectives estimands hypotheses and endpoints aligned
- Analysis populations treatment groups baseline and visits defined
- Data cut randomisation blinding and sample-size sources stated
- Models contrasts intervals covariates and coding executable
- Multiplicity subgroup and interaction roles prespecified
- Missing data intercurrent events and deviations handled coherently
- Sensitivity and supplementary analyses answer named questions
- Interim access stopping and monitoring governance protected
- Safety summaries and output shells define denominators and units
- Software code validation traceability changes and archive controlled
- Qualified statisticians and current regulatory guidance lead real work
Advice for students, parents and young adults
Students can preregister a simple study-method comparison by choosing the outcome, time point, groups and missing-response rule before collecting results.
Parents can use the topic to explain why honest analysis begins before numbers appear: planning reduces the temptation to choose only the most favourable comparison.
Clinical SAPs require qualified statisticians, clinicians, data managers, programmers and regulatory specialists. This educational framework cannot authorise a study or replace current guidance.
Frequently asked questions
What is a statistical analysis plan?
It is a controlled document that expands the protocol’s statistical principles into detailed, executable analysis methods and outputs.
When should an SAP be finalised?
Timing depends on governance, but important analyses should be prespecified before unblinded results can influence choices.
What is an estimand?
It is a precise description of the treatment effect of interest, including population, variable, treatment conditions, handling of intercurrent events and summary measure.
Why plan sensitivity analyses?
They test how robust the primary conclusion is to important assumptions or alternative analytical choices.
Can the SAP change?
Necessary changes can occur, but version, timing, rationale, approval and impact should be documented transparently.
Does a SAP guarantee a valid conclusion?
No. Validity also depends on design, conduct, data quality, assumptions, implementation and interpretation.
The deeper English lesson
SAP English is question-to-inference language. It freezes the analytical grammar of a study early enough to protect credibility while still making assumptions, uncertainty and justified change visible.
Useful next reading
Continue with writing a clinical trial protocol deviation report, writing a research ethics application, writing a research data management plan, writing a laboratory method validation report, and the How English Works.