Blinded Variance Re-Estimation

Why Re-Estimate Variance While Blinded?

A trial’s planned sample size depends on its assumed outcome variability. If the true within-group standard deviation is larger than planned, a trial may have less power than intended; if it is smaller, the planned sample size may be larger than necessary. An interim estimate of variability can therefore inform a sample-size update without requiring access to treatment-specific efficacy results.

For a continuous outcome, a blinded interim analyst can calculate the variance of the pooled outcomes without using treatment assignments or comparing arm-specific means. The important qualification is that pooled variance is not always identical to within-arm variance: when treatment means differ, their separation contributes to the variation in the combined data. A blinded calculation can avoid revealing treatment-specific results, but it does not guarantee that the pooled variance is unaffected by the treatment effect.

Model and Sample-Size Formula

Suppose a randomized trial compares a treatment group with a control group on a continuous outcome. Assume independent participants, a common within-group variance \(\sigma^2\), and normally distributed outcomes within each group. Let \(\delta^*\) be the prespecified treatment difference for which the design aims to have power \(1-\beta\). For a two-sided test with significance level \(\alpha\) and equal allocation, the usual normal-approximation target is

where \(n_{\mathrm{req}}\) is the required number of participants per group, and \(z_q\) is the \(q\)-quantile of the standard normal distribution. This approximation follows from the standard error of a difference in sample means: with \(n\) participants per group, it is \(\sigma\sqrt{2/n}\). The target difference divided by that standard error must be large enough to give the specified power, yielding the formula above. Exact power calculations may instead use a noncentral \(t\) distribution, particularly when sample sizes are small.

A blinded update replaces the planned variance with an interim pooled estimate \(s_{\mathrm{blind}}^2\):

The result is rounded up to a whole participant per group. The target \(\delta^*\), significance level, power, test, and allocation rule should be prespecified. Replacing \(\delta^*\) with an interim treatment-effect estimate would no longer be a variance-only blinded update.

What the Pooled Variance Estimates

Let the treatment allocation fraction be \(\pi\), so the control fraction is \(1-\pi\). Suppose the within-group means are \(\mu_T\) and \(\mu_C\), with common variance \(\sigma^2\). The variance of the combined outcome distribution follows from the law of total variance: variability within groups contributes \(\sigma^2\), while variability between the two group means contributes \(\pi(1-\pi)(\mu_T-\mu_C)^2\). Thus,

This population-mixture expression explains the key limitation of a raw blinded pooled variance. If the treatment and control means differ, then the pooled variance includes a treatment-effect contribution. With equal allocation, that contribution is one quarter of the squared mean difference. In a finite interim sample, the usual sample variance fluctuates around the corresponding pooled variability; it is not a precise reading of the population quantity.

Under the null hypothesis of equal means, the mixture contribution is zero. Under an alternative, it is positive. Consequently, using raw pooled variance as though it were a within-group variance can produce a larger sample-size target than a calculation based on the true within-group variance. That may be a conservative increase in some settings, but neither the degree of conservatism nor error-rate control should be assumed without checking the full design.

Interpreting and Checking the Update

A variance-based update is useful only if the interim estimate is sufficiently informative and appropriate for the planned analysis. The pooled data should reflect the same outcome definition, measurement schedule, and relevant population as the final analysis. Review the interim snapshot for data errors, missing outcomes, extreme values, changes in measurement practice, and substantial differences in follow-up. A pooled standard deviation can be distorted by any of these, and a small interim sample can yield a noisy estimate even when the model is correct.

The update should also follow a prespecified rule: when the re-estimation occurs, how the pooled variance is calculated, the allowable sample-size range, and how the final target is rounded and allocated. Caps can limit the effect of an unstable estimate or operational constraints, but a cap does not correct a misspecified variance model. If the trial uses covariate adjustment, repeated measurements, clustering, or unequal variances, the simple two-group formula may not represent the variance of the planned estimator; the update must match the actual design and analysis.

Blinding, Alternatives, and Limitations

The advantage of the pooled approach is informational: the calculation does not require arm-specific means, an estimated treatment contrast, or treatment identity. This can help preserve masking for people who should not see treatment-specific efficacy results. It is not a guarantee that the pooled data reveal nothing about the treatment effect. As the mixture-variance derivation shows, pooled variability can reflect a difference between group means, although the blinded calculation does not identify that difference or its direction.

An unblinded variance calculation can estimate within-group variability directly from arm-specific data, but requires access to treatment assignments and may expose treatment-specific interim information. A raw blinded pooled estimate avoids that access but may include treatment-effect mixing. An adjustment that subtracts a prespecified mixture contribution is possible only by making additional assumptions about the treatment difference and allocation; it should not be presented as a direct empirical estimate of within-group variance. The choice among these approaches is a design decision, not an arithmetic detail.

Most importantly, keeping the variance calculation blinded does not, by itself, prove that the adaptive design preserves the nominal type I error rate. The sample-size rule, timing, final test, and any caps must be evaluated together under the null hypothesis using a design-specific argument or appropriate simulation. The adaptation should be written into the protocol and statistical analysis plan before interim results are reviewed.

Common Mistakes

• Calling pooled variance the within-group variance. With unequal group means, pooled variance includes a between-group component. State which variance is being estimated and recognize the possible inflation.
• Using interim efficacy to choose the target effect. A blinded variance update retains the prespecified \(\delta^*\); substituting an interim treatment-effect estimate changes the information used for adaptation.
• Rounding down. A sample-size target is rounded up, then allocated according to the planned randomization ratio.
• Assuming blinding guarantees error control. Masking limits access to treatment-specific results, but type I error control depends on the complete adaptive design and final analysis.
• Applying the simple formula to a different analysis model. Correlation, clustering, unequal variances, or a different estimator can change the relevant sampling variance.