Selecting Interim Timing For Nuisance Estimation

Why Interim Timing Is a Design Choice

A blinded sample-size re-estimation can use information about a nuisance parameter, such as outcome variance or a pooled event rate, without revealing treatment assignments. The review cannot be informative before enough usable data have accumulated. But waiting longer means recruiting more participants before the review, and may leave less time or capacity to complete any additional enrollment it indicates. Choosing an interim time is therefore a balance, not a search for one universally best fraction of the trial.

The relevant information is not always the number recruited. A variance estimate may require analyzable outcomes; a rate estimate requires outcomes observed and sufficiently mature to classify. Follow-up delays, missing measurements, or staggered outcome availability can make the usable sample smaller than the enrollment count at the review date.

Let \(n_I\) denote the number of independent, usable observations available at the interim review, and let \(N\) denote the current planned total enrollment. The fraction \(n_I/N\) is a convenient summary only when the interim observations are representative of the planned analysis population and are all available for estimation. The appropriate measure of information depends on the nuisance estimator.

Assumptions and a Precision Calculation for Variance

For a simple benchmark, suppose the usable observations are independent and normally distributed with a common variance \(\sigma^2\), and the variance estimator \(S^2\) is based on \(\nu\) residual degrees of freedom. Then \(\nu S^2/\sigma^2\) has a chi-square distribution with \(\nu\) degrees of freedom. This benchmark does not automatically describe every blinded pooled estimator; its degrees of freedom and sampling behavior must match the estimator actually planned.

The chi-square model gives \(E(S^2)=\sigma^2\) and \(\operatorname{Var}(S^2)=2\sigma^4/\nu\). Thus the standard deviation of \(S^2\), relative to the true variance, is \(\sqrt{2/\nu}\). This is a measure of the estimator's relative sampling variability, not a guarantee that one realized estimate will be close to \(\sigma^2\).

If a design wants this relative standard error to be no greater than a chosen tolerance \(r\), the benchmark calculation requires \(\sqrt{2/\nu}\le r\), or \(\nu\ge 2/r^2\). This supplies a transparent way to compare candidate timings. It does not choose the tolerance: that is a design judgment informed by the consequences of nuisance-estimation error and the operational cost of waiting.

Timing a Review for a Pooled Event Rate

For independent binary outcomes with a common pooled event probability \(p\), the sample proportion \(\hat p\) has standard error \(\sqrt{p(1-p)/n_I}\). Because \(p\) is unknown, a planning calculation can use a plausible value or a conservative bound. Since \(p(1-p)\le 0.25\), the standard error is at most \(0.5/\sqrt{n_I}\) under this simple independent Bernoulli model. Clustering, unequal observation weights or a different estimator would require a different precision calculation.

Information Availability, Burden and Remaining Enrollment

A precision target is only one part of timing. A review also has to occur early enough to be useful. If sample-size reassessment may indicate that more participants are needed, the design needs recruitment capacity and follow-up time after the estimate becomes available. Conversely, an excessively early review can impose a recruitment or administrative burden while producing an estimate too unstable to support a meaningful update.

For a fixed total plan \(N\), the simple remainder \(N-n_I\) indicates how much planned enrollment lies beyond the review. If the total could change after reassessment, this remainder is not a forecast of the final sample size; it is a measure of how much enrollment remains under the current plan. Reassessment procedures and any limits on the resulting sample size are separate design specifications.

A practical comparison of candidate timings can be organized around four questions:

Diagnostics, Alternatives and Limitations

The calculations are only as reliable as their assumptions. Before using a variance-based timing calculation, check whether observations are independent, whether the estimator has the assumed degrees of freedom, and whether its sampling distribution is reasonably represented by the normal-model benchmark. For a rate, check that the analysis unit is correct, outcomes are mature, and the assumed independence and pooling are defensible. Earlier material on blinded pooled-rate estimation and site heterogeneity addresses estimator construction and heterogeneity; timing calculations should use the estimator actually chosen, not replace those checks.

A fixed calendar-date review is easy to administer, but the number of usable outcomes by that date can vary. A target number of analyzable participants is more directly linked to variance precision, but may occur at different calendar times. A target number of events can be useful when event information drives the nuisance estimate, but it does not by itself ensure enough non-events or a stable rate estimate. A calendar date combined with a minimum information requirement can address some of these tradeoffs, though it adds operational rules that need to be clearly specified.

The standard-error calculations describe sampling variability under a model; they do not diagnose bias or guarantee that a blinded estimate represents the final analysis population. Changes in sites, case mix, follow-up, missingness, or event rates over time may undermine representativeness. In addition, a pooled blinded estimate can behave differently from a simple common-parameter model if treatment effects influence the pooled data. A precise estimate can still be a poor planning estimate when its model or target is wrong.

The main alternatives are therefore not simply “early” and “late.” One option is to use a prespecified nuisance value from external or earlier data, avoiding an interim review but accepting the risk that it is inaccurate for the current study. Another is to use a blinded interim estimate, trading additional information against recruitment and review burden. A later review usually has more information, but may leave less operational room to act. No approach removes uncertainty; the choice should reflect how sensitive the sample-size calculation is to the nuisance parameter and whether the study can respond in time.

Common Mistake

A second caution is to treat a precision target as a design criterion rather than a promise about the realized estimate. Even if the expected standard error meets a threshold, a particular blinded estimate can differ materially from the underlying nuisance parameter. Timing should be supported by sensitivity checks across plausible nuisance values and by a review of whether the assumed model remains credible.