The Gold Standard of Experimental Design | Blinded Sample Size Re-Estimation, Variance Misspecification and Type I Error

1. Identify which assumption is being repaired

Back to contents

Write the original sample-size formula and distinguish the target treatment effect from nuisance parameters. For a continuous outcome, the common standard deviation may be uncertain; for a binary outcome, the pooled event rate can affect variance; for time-to-event designs, information depends on event accrual. Re-estimating an effect size is different from updating variance. State which parameter may change, why existing evidence is weak and which design features—alpha, power, allocation and estimand—must remain fixed.

2. Plan the internal pilot before enrolment

Back to contents

Specify the interim information time, eligible observations, blinding method, estimator, decision rule, minimum and maximum total sample and who will perform the calculation. The protocol and statistical analysis plan should explain how late or missing outcomes are handled and whether recruitment pauses. A flexible promise to “reassess if necessary” creates operational and inferential discretion. Prespecification turns adaptation into part of the design rather than a reaction to an inconvenient interim pattern.

3. Understand what blinded data conceal and retain

Back to contents

A blinded assessment pools treatment groups or uses masked labels so the analyst does not deliberately inspect the treatment contrast. It can still reveal overall outcome levels, variability, missingness and recruitment. Pooled variance may include between-group differences as well as within-group noise, particularly when a real effect exists. Blinding reduces routes to outcome-driven adaptation, but it does not make every estimator unbiased or every operational decision harmless. Document the masking and data-access boundary.

4. Estimate nuisance parameters with the design in mind

Back to contents

For normally distributed continuous outcomes, several blinded variance estimators exist, with different small-sample behaviour and treatment-effect contamination. Binary, cluster and survival designs need methods aligned to their information structure. Do not transplant a formula from a parallel continuous trial into a clustered or repeated-measures design. Account for covariate adjustment, allocation ratio and correlation exactly as the final analysis will. Simulation is often the clearest way to evaluate bias and achieved power under plausible parameter ranges.

5. Protect type I error as an operating characteristic

Back to contents

A sample-size change based only on appropriate blinded nuisance information can often be designed with little or controlled effect on the false-positive rate, but that is not a universal theorem. Small samples, discrete outcomes, non-inferiority settings, estimator choice and data-dependent caps can alter operating characteristics. Unblinded conditional-power or effect-size adaptations require methods that explicitly control alpha. Evaluate the whole rule by simulation and report the achieved type I error, not merely the word “blinded.”

6. Worked example: variance higher than planned

Back to contents

A two-arm trial targets a mean difference of 5, assumes a standard deviation of 10, uses equal allocation, two-sided alpha 0.05 and 80% power. The simple normal approximation gives about 63 participants per arm: two times (1.96 plus 0.84) squared times 100 divided by 25. A prespecified blinded internal pilot suggests SD 13, raising the approximation to about 106 per arm. The increase addresses variance misspecification within the cap; it does not revise the target effect or guarantee success. Because pooled variance can include a treatment difference, the protocol also evaluates estimator bias in simulation.

7. Keep operations from revealing the answer

Back to contents

Only a designated independent statistician or committee should access data needed for the adaptation, and the investigators should receive the permitted decision—such as continue to a revised total—without treatment-specific summaries. Access logs, code versioning and decision minutes preserve the firewall. A dramatic increase can itself hint at outcomes, so communications should be standardised. Recruitment, endpoint assessment and care should not change because staff infer which arm is performing better.

8. Changed-context transfer check: a binary endpoint

Back to contents

Move from a continuous measurement to a binary response. The nuisance quantity is tied to response probabilities and the chosen comparison, not a single SD. A blinded pooled event rate may update information, yet its behaviour depends on the true arm rates and allocation. The transferable method is to prespecify the masked estimator, decision rule and simulations. The numerical formula and risk of conservatism change with the endpoint, so the continuous-outcome rule cannot simply be renamed.

9. Align the adaptation with the final analysis

Back to contents

If the final model adjusts for baseline covariates, handles repeated measures or estimates a cluster-level effect, the information update should reflect that structure. Retain internal-pilot participants in the final analysis unless the design justifies otherwise; discarding them wastes information and can introduce selection. Do not change the endpoint, estimand or non-inferiority margin under the cover of sample-size re-estimation. Any other adaptation needs its own rationale and multiplicity control.

10. Know when re-estimation does not rescue the study

Back to contents

A maximum sample cap may leave power below target when variance is much larger than expected. Delayed outcomes may make the decision arrive after most recruitment. High variability caused by data errors or inconsistent measurement calls for process repair, not merely more participants. Re-estimation cannot fix a biased endpoint, poor randomisation, weak adherence or an implausible target effect. Report when the rule reaches its cap and the conditional assurance remains limited.

11. Audit the calculation and the firewall

Back to contents

Archive the original assumptions, protocol rule, eligible interim dataset, masking method, estimator, code, cap, decision record and access list. Reproduce the revised total from frozen data. Simulate type I error and power across plausible effects, variances, recruitment patterns and missingness, including no adaptation and cap-binding cases. Confirm that the final analysis and estimand match the plan. The audit should show both statistical validity and protection against operational bias.

12. The gold standard adapts information without chasing effects

Back to contents

Blinded sample-size re-estimation is useful when uncertainty lies in a nuisance parameter and the adaptation is planned as part of the trial. The gold standard identifies the repair, limits access, uses a design-aligned estimator, caps the decision and verifies operating characteristics. It neither treats “blinded” as a magic safeguard nor promises to restore power under every scenario. The result is a transparent adjustment to information, not an interim search for a favourable treatment story.