The Gold Standard of Experimental Design | Sample Size Re-Estimation, Information Fractions and Unblinded Variance Reviews

1. Begin with the uncertainty the design is meant to repair

Back to contents

List the assumptions that determined the original sample size: outcome variance, control event rate, dropout, intracluster correlation, treatment effect, type I error and target power. Re-estimation should address a prespecified uncertainty, not every unfavourable interim signal. Updating a nuisance parameter can protect power when the effect of interest remains fixed. Increasing the sample because the observed effect is smaller changes the procedure more deeply and can chase noise. State the maximum sample, permitted decrease or increase, timing and the estimand that remains unchanged.

2. Distinguish blinded and unblinded reviews

Back to contents

A blinded review uses information that does not reveal comparative treatment effects, such as pooled variance under suitable methods. It usually creates less operational bias, but pooled quantities can still depend on treatment effects and allocation ratios. An unblinded review uses treatment labels or effect estimates and requires protected access, prespecified rules and statistical adjustment where necessary. The labels describe information flow, not automatic validity. Ask exactly which statistics the reviewing body sees and which decisions they can trigger.

3. Define information fraction, not just calendar fraction

Back to contents

The information fraction is the proportion of final statistical information available at interim. It may differ from the fraction of participants enrolled. For a survival endpoint, events often determine information; for a continuous outcome, observed outcomes and variance matter; clustering and delayed measurement complicate both. A review at 50 percent enrolment may have far less than 50 percent information if outcomes are pending. Specify how the fraction is calculated and what happens if the planned information threshold is reached earlier or later than expected.

4. Separate nuisance-parameter updates from effect-based rescue

Back to contents

If the true variance is larger than assumed, a blinded variance update can increase size to restore planned power for the original effect. If the observed treatment difference is smaller, a rule based on conditional power uses effect information and may inflate false-positive risk or produce an unattractive zone in which promising but noisy results receive the largest increase. Distinguish sample size re-estimation under a fixed alternative from methods that adapt to the interim effect. Combination tests, conditional-error methods or other planned procedures can preserve inference when unblinded adaptation is allowed.

5. Understand information, type I error and estimator bias together

Back to contents

A design can control the chance of a false positive yet still yield biased effect estimates or confidence intervals after adaptation. Trials that expand when interim effects look moderately favourable select particular data paths. The final analysis must match the adaptive rule and report the actual information used. Simulate null, target and intermediate effects across nuisance values. Check type I error, power, expected sample size, estimation bias and interval coverage—not power alone. Operational behaviour should also be assessed because visible size changes can reveal interim trends.

6. Worked example: updating variance without chasing effect

Back to contents

A two-arm trial plans 200 participants for a continuous outcome assuming a standard deviation of 10. At 50 percent information, an independent group reviews pooled blinded data and estimates a standard deviation of 12. A prespecified rule raises the total to 288, capped at 300, while treatment effects remain hidden. The final analysis uses the original two-sided alpha and adjusted precision. Sensitivity checks show that a large treatment effect can slightly inflate pooled variance, so simulations cover several effects and allocation ratios. The conclusion is that nuisance uncertainty was managed—not that the intervention looked promising at interim.

7. Govern unblinded information as a separate system

Back to contents

An independent data monitoring committee or statistical centre should receive only the information needed for the rule. Investigators, recruiters, outcome assessors and sponsors should remain blinded where possible. Record data cut, code version, meeting materials, recommendation and implementation without disclosing comparative results. Even a total sample-size increase can invite speculation, affect recruitment or encourage endpoint changes. Use decision rules that limit discretion and communication plans that reveal the operational decision without revealing its statistical cause. Emergency safety review remains distinct from power adaptation.

8. Changed-context transfer check: a cluster education trial

Back to contents

A school trial planned with an intracluster correlation of 0.03 observes a blinded estimate near 0.07 after half the schools provide outcomes. Re-estimation may add clusters, but adding students within existing schools is not equivalent because design effect depends on both cluster size and correlation. The transferable method identifies the nuisance quantity, computes actual information, uses a capped rule and preserves masking. The context changes because school calendars limit recruitment and a visible addition of schools may signal assumptions. Feasibility and degrees of freedom must be simulated, not borrowed from an individually randomised trial.

9. Know the limits and common misconceptions

Back to contents

Re-estimation cannot repair a biased endpoint, broken randomisation, selective missingness or implausible effect target. Blinded does not mean assumption-free, and unblinded does not mean invalid if the method and governance are rigorous. A cap can leave the study underpowered when variance is extreme. Reducing sample size may sacrifice safety information or credibility. Conditional power calculated from a noisy interim effect can be unstable. More participants do not make a data-driven redesign equivalent to the original confirmatory study.

10. Audit operating characteristics and the decision trail

Back to contents

Publish the protocol, original inputs, interim information definition, nuisance estimator, adaptation algorithm, minimum and maximum size, multiplicity or conditional-error method, simulations and governance charter before access to comparative data. Preserve enrolment and outcome lags, the actual data cut, blinded or unblinded reports and implementation timing. Reproduce operating characteristics under skewness, unequal variances, dropout, clustering and treatment effects different from plan. Report the planned, recalculated and achieved information along with reasons for deviations.

11. Frequently asked questions

Back to contents

Is sample size re-estimation the same as continuing until significance? No; valid designs use a prespecified interim and bounded rule. Can a blinded variance review affect type I error? It can under some methods and settings, so the procedure needs justification and simulation. Why use information fraction rather than enrolment fraction? Outcomes and events, not headcount alone, determine precision. Must an unblinded rule use adjusted testing? Often yes, depending on the adaptation. Can the sample size decrease? Sometimes, but safety, precision and credibility must be considered.

12. The gold standard treats adaptation as part of the design

Back to contents

A credible re-estimation plan names the uncertain nuisance quantity, times the review by information, controls who sees what and matches the final analysis to the rule. It evaluates error, bias, coverage, power and sample-size distribution before the trial. When the sample changes, the report shows exactly why and what remained fixed. Adaptation then protects the question from bad planning assumptions without quietly replacing the question after outcomes begin to speak.