Pooled Variance Estimation Under Blinding

Why Blinded Pooling Needs Care

In a blinded sample size re-estimation, interim outcomes may be available while treatment assignments remain concealed. For a continuous endpoint, the pooled sample variance is an appealing estimate of the variability that drives the planned sample size: it can be calculated without revealing which participants received which treatment. But it measures spread around the overall interim mean. If the treatments have different true means, that spread includes both ordinary outcome variability and separation between the treatment groups.

The key distinction is between a variance estimate calculated without labels and an estimate of the within-treatment variance needed for planning. They coincide in expectation under a common-mean model, but generally not under a treatment effect. Blinded pooling can therefore support re-estimation when its assumptions and purpose are clear; it should not automatically be treated as a pure estimate of within-arm variance.

Definition: Suppose $N=n_1+n_2$ participants have continuous outcomes $Y_1,\ldots,Y_N$, and treatment labels are concealed. The blinded pooled sample variance is $$ S_{\mathrm{blind}}^2=\frac{1}{N-1}\sum_{i=1}^{N}(Y_i-\bar{Y})^2, \qquad \bar{Y}=\frac{1}{N}\sum_{i=1}^{N}Y_i. $$ It is the sample variance of all interim outcomes around their overall mean, without separating the outcomes by treatment group.

Model and Assumptions

For the derivations below, assume a two-arm trial with independent observations and fixed interim group sizes $n_1$ and $n_2$. In group $j$, outcomes follow a normal distribution with mean $\mu_j$ and a common within-group variance $\sigma^2$. Write the treatment-effect difference as $\delta=\mu_1-\mu_2$. The normal model and common variance make the usual pooled within-group variance an appropriate estimate of $\sigma^2$; the effect of unequal means on the blinded variance follows from the sum-of-squares decomposition.

The standard planning calculation for a balanced, two-sided comparison of means uses a target difference $\delta^*$, significance level $\alpha$, and power $1-\beta$. A normal-quantile approximation to the required number per arm is

n_{\mathrm{per\ arm}}\approx \frac{2\sigma^2\left(z_{1-\alpha/2}+z_{1-\beta}\right)^2}{(\delta^*)^2}.

Here $z_q$ is the $q$-quantile of the standard normal distribution. The formula shows why a variance estimate matters: with the other design quantities fixed, required sample size is proportional to $\sigma^2$. It is an approximation to planning for a two-sample mean comparison; a final design may use a more exact calculation, such as one based on the $t$ distribution.

How a Treatment Effect Enters the Blinded Variance

The total sum of squares around the overall mean separates into within-group and between-group components:

\sum_{i=1}^{N}(Y_i-\bar{Y})^2 = \sum_{j=1}^{2}\sum_{i\in j}(Y_i-\bar{Y}_j)^2 + \sum_{j=1}^{2}n_j(\bar{Y}_j-\bar{Y})^2.

The first term measures variation around each group mean. The second measures the extra spread caused by the group means' distance from the overall mean. In the assumed model, the expected first term is $(N-2)\sigma^2$, because each of the two groups contributes $n_j-1$ within-group degrees of freedom.

For the second term, the group means' difference has expectation $\delta$ and variance $\sigma^2(1/n_1+1/n_2)$. The between-group sum of squares equals $n_1n_2/N$ times the squared difference of the group means. Therefore its expectation is

\frac{n_1n_2}{N} \left[\delta^2+\sigma^2\left(\frac{1}{n_1}+\frac{1}{n_2}\right)\right] = \frac{n_1n_2}{N}\delta^2+\sigma^2.

Adding the expected components and dividing by $N-1$ gives the expected blinded sample variance:

E(S_{\mathrm{blind}}^2) = \sigma^2+ \frac{n_1n_2}{N(N-1)}\delta^2.

The second term is nonnegative and is zero when the group means are equal. Thus, under these assumptions, the blinded pooled variance is unbiased for $\sigma^2$ under the null hypothesis of equal means, but its expectation exceeds $\sigma^2$ when the true treatment effect is nonzero. For equal group sizes $n_1=n_2=n$, the coefficient of $\delta^2$ is $n/[2(2n-1)]$, which approaches $1/4$ as $n$ grows. This is an expectation over repeated trials, not a guarantee that any one interim estimate will exceed the true variance.

Worked Example: Quantifying Effect Contamination

Suppose a hypothetical interim analysis has 20 participants per arm, a common within-arm variance of $100$ points squared, and a true difference in group means of $8$ points. The assignments remain concealed. Calculate the expected blinded sample variance under the model.

Here $n_1=n_2=20$, $N=40$, $\sigma^2=100$, and $\delta=8$. Substitution into the derived expression gives

E(S_{\mathrm{blind}}^2) =100+\frac{20(20)}{40(39)}(8^2) =100+\frac{400}{1560}(64) =100+16.4103 =116.4103\ \text{points}^2.

The expected blinded estimate is about $116.4$ points squared, rather than $100$ points squared, because the pooled outcomes include between-group separation as well as within-group variation. This does not mean every interim estimate will equal $116.4$; it is the model-based expectation across repetitions of the trial.

Using the Estimate in a Re-Estimation Rule

A blinded re-estimation rule can substitute $S_{\mathrm{blind}}^2$ for the planning variance in the sample size formula, then apply prespecified operational constraints such as a minimum or maximum sample size. The rule should specify in advance which interim outcomes are used, how missing observations are handled, the calculation method, and how the resulting sample size is rounded. The calculation should not require access to treatment assignments.

The following example illustrates the direct, unadjusted calculation. The interim variance is hypothetical and is not presented as a finding from a real trial.

Worked Example: Re-Estimating Sample Size From a Blinded Variance

Suppose a hypothetical balanced trial has 40 interim participants, pooled across concealed assignments. Their overall sum of squared deviations from the pooled mean is $5616$ points squared. The target difference is $10$ points, the planned two-sided significance level is $\alpha=0.05$, and the target power is $0.80$. Calculate the blinded variance estimate and the approximate required number per arm if it is used without adjustment.

The blinded sample variance uses $N-1=39$ degrees of freedom:

S_{\mathrm{blind}}^2=\frac{5616}{39}=144\ \text{points}^2.

For the stated design, $z_{1-\alpha/2}=z_{0.975}\approx1.960$ and $z_{1-\beta}=z_{0.80}\approx0.8416$. The squared sum of quantiles is

$$ (1.960+0.8416)^2=(2.8016)^2=7.84896. $$

Using the estimate directly in the planning formula gives

n_{\mathrm{per\ arm}} \approx \frac{2(144)(7.84896)}{10^2} = \frac{2260.50048}{100} = 22.6050.

Rounding up gives a requirement of 23 participants per arm, or 46 total, before any prespecified cap or other design adjustment. For comparison, using a planning variance of $100$ points squared would give $2(100)(7.84896)/100=15.6979$, requiring 16 per arm. The difference illustrates how strongly sample size can respond to the variance input; it does not establish that the larger interim estimate represents within-arm variance alone.

Adjustment and Alternative Estimates

One possible response to effect contamination is to use a prespecified effect assumption to adjust the blinded variance. From the expectation formula, subtracting the estimated between-group contribution yields

\widehat{\sigma}_{\mathrm{adj}}^2 = S_{\mathrm{blind}}^2- \frac{n_1n_2}{N(N-1)}(\delta_0)^2,

where $\delta_0$ is a prespecified assumed treatment difference. Under the model, this adjustment is unbiased for $\sigma^2$ if $\delta_0^2=\delta^2$. It is not a way to discover the concealed effect: if the assumed effect is wrong, the adjustment may over-correct or under-correct. A protocol must also specify how to handle a negative adjusted value, which is not a plausible variance estimate. For example, a prespecified lower bound can prevent an unusable result, but the bound itself affects the re-estimation rule.

Worked Example: Effect-Adjusted Planning Value

Use the previous hypothetical interim variance of $144$ points squared and group sizes of 20 per arm. Suppose the design had prespecified an assumed treatment difference $\delta_0=10$ points for this adjustment. Calculate the adjusted estimate and its sample size implication.

The finite-sample coefficient is $400/[40(39)]=400/1560\approx0.2564103$. The assumed between-group contribution is therefore

\frac{400}{1560}(10^2)=25.6410\ \text{points}^2.

Subtracting it from the blinded variance gives

\widehat{\sigma}_{\mathrm{adj}}^2=144-25.6410=118.3590\ \text{points}^2.

With the same target difference, significance level, and power as before, the approximate sample size is

n_{\mathrm{per\ arm}} \approx \frac{2(118.3590)(7.84896)}{100} = 18.580\ \text{(approximately)}.

Rounding up requires 19 participants per arm. The adjusted value is not automatically preferable to the unadjusted value: its usefulness depends on whether the prespecified effect assumption is defensible and whether the adjustment rule was set before interim data were examined.

A different alternative is an unblinded pooled within-group variance, calculated after treatment labels are available. Under the common-variance normal model, it is

S_p^2= \frac{(n_1-1)s_1^2+(n_2-1)s_2^2}{N-2}.

This estimates within-group variance without the between-means contribution, but it requires treatment-specific interim summaries and thus is not blinded. Operational safeguards and the implications of using treatment information are separate design considerations. The alternatives can be summarized as follows.

Approach	What it estimates or uses	Main trade-off
Unadjusted blinded pooling	Overall interim spread	Maintains concealment; can include treatment-effect spread
Effect-adjusted blinded pooling	Overall spread minus a prespecified effect contribution	Maintains concealment; depends on the assumed effect and adjustment rule
Unblinded within-group pooling	Within-treatment variance	Targets the variance directly under the model; requires treatment labels

Diagnostics, Limitations, and a Common Pitfall

The derivation relies on independent outcomes, a continuous endpoint, fixed interim group sizes, and a common within-arm variance. Departures can make a single pooled variance a poor planning input. Before applying a rule, check for data errors, unusual outliers, strong skewness, and unequal variability across groups when such information can be assessed under the planned access controls. Review whether the analyzed participants and endpoint match the protocol's re-estimation definition, and whether missing outcomes or unequal allocations change the intended calculation. A blinded pooled estimate alone cannot diagnose every model failure, because it does not reveal group-specific patterns.

The normal-theory sample size expression is also an approximation, and a variance estimated from a modest interim sample is itself uncertain. Re-estimation based on one noisy estimate can produce an unstable sample size. A prespecified rule may address this with bounds, a minimum information requirement, or a more conservative planning choice; these choices should be justified rather than selected after seeing the interim result. The calculation addresses variance input, not every issue concerning error rates or inference after adaptation.

Common mistake: “The interim assignments are concealed, so the pooled variance must estimate only the within-arm variance.” Concealment prevents direct use of treatment labels; it does not remove the spread caused by different group means.
Correction: Under the stated model, $E(S_{\mathrm{blind}}^2)=\sigma^2+[n_1n_2/\{N(N-1)\}]\delta^2$. The extra term vanishes under equal means, but not generally under a treatment effect.
Another pitfall: Treating the expected contamination term as the exact amount in one trial. It is an expectation; the observed variance also varies from trial to trial.

Key takeaway: Blinded pooling can provide a practical variance input when the endpoint and model support it, and it preserves treatment concealment. Its interpretation depends on the treatment means: under unequal means, the pooled variance includes both within-arm variation and between-arm separation. Any adjustment should be prespecified and tied to an explicit assumption.

Check Your Understanding

Answer each question using the model and definitions in this tutorial. Full solutions follow the questions.

In a balanced two-arm trial with a nonzero treatment effect, why can the blinded pooled sample variance exceed the common within-arm variance in expectation?
For $n_1=15$, $n_2=25$, and $N=40$, write the coefficient multiplying $\delta^2$ in the expected blinded variance.
If an unadjusted variance estimate is $121$ points squared and the planning variance was $100$ points squared, what happens to the approximate sample size when all other planning quantities stay fixed?
What assumption is needed for the effect-adjusted estimate to be unbiased for $\sigma^2$?

Solutions. (1) The overall pooled sum of squares includes the between-group term $n_1n_2(\bar{Y}_1-\bar{Y}_2)^2/N$; when the true means differ, this contributes a nonzero treatment-separation component in expectation. (2) The coefficient is $15(25)/[40(39)]=375/1560\approx0.2404$, so the expectation is $\sigma^2+0.2404\delta^2$. (3) The sample size is proportional to the variance, so its unrounded value is multiplied by $121/100=1.21$, a 21% increase; the final integer depends on the original calculation and rounding rule. (4) The prespecified assumed difference must satisfy $(\delta_0)^2=\delta^2$ under the model, and the other model assumptions must hold. If that effect assumption is wrong, the adjustment need not be unbiased.