Recalculating Continuous-Outcome Sample Size

Why Recalculate Sample Size?

A continuous-outcome trial is usually planned using an assumed outcome variance. If an interim blinded estimate suggests that the variance is larger than assumed, the original sample size may no longer deliver the planned power for the target difference. Recalculation replaces the planning variance with the interim estimate while keeping the target difference, type I error, power, and allocation ratio fixed.

This is a nuisance-parameter update, not a new estimate of the treatment effect. In a blinded internal-pilot design, the variance estimate must be obtained using the protocol's prespecified method for masked data, including any correction for treatment-effect contamination where appropriate. The earlier material on pooled variance estimation and its correction provides that context; here the estimate is treated as an input to the sample-size calculation.

Definition: A variance-based sample-size re-estimation substitutes an interim estimate of the common outcome variance into the sample-size formula, while retaining the prespecified target difference \(\Delta\), significance level \(\alpha\), target power \(1-\beta\), and allocation ratio. The update is based on the variance estimate, not on an unblinded interim treatment-effect estimate.

The Two-Sample Normal-Theory Calculation

Suppose the final comparison is between two independent groups with normally distributed outcomes and a common variance \(\sigma^2\). Let \(n_C\) be the control-group sample size and \(n_T\) the treatment-group sample size. Define the allocation ratio as \(r=n_T/n_C\), so \(n_T=r n_C\). The target difference \(\Delta\) is the difference in population means that the design is intended to detect; it is not necessarily the effect observed at the interim analysis.

For a two-sided fixed-sample test with significance level \(\alpha\), the normal-theory approximation uses the standard-normal quantiles \(z_{1-\alpha/2}\) and \(z_{1-\beta}\). With a planning variance \(\sigma^2\), the required control-group size is

n_C = \frac{(1+1/r)\sigma^2 \left(z_{1-\alpha/2}+z_{1-\beta}\right)^2} {\Delta^2}, \qquad n_T=r n_C.

In re-estimation, substitute the interim variance estimate \(\hat{\sigma}_{\mathrm{int}}^2\) for \(\sigma^2\). The resulting calculation is a plug-in approximation: it treats the interim estimate as the variance for planning, although that estimate itself is uncertain.

Assumptions for this formula: The two groups are independent; outcomes are approximately normal with a common variance; the allocation ratio is fixed; the target is a difference in means; and the test is the stated two-sided normal-theory test. The calculation uses a fixed-sample normal approximation. If the trial will use group sequential boundaries, this formula alone does not establish the power or type I error of that sequential design; the final design calculation must account for its monitoring plan.

Why the Variance Update Scales the Sample Size

With \(\Delta\), \(\alpha\), power, and \(r\) held fixed, every factor in the formula other than the variance is constant. Therefore, if the original planning variance was \(\sigma_{\mathrm{plan}}^2\), the new continuous-valued control sample size is the old continuous-valued size multiplied by the ratio of the interim variance estimate to the planning variance.

\frac{n_{C,\mathrm{new}}}{n_{C,\mathrm{old}}} = \frac{\hat{\sigma}_{\mathrm{int}}^2}{\sigma_{\mathrm{plan}}^2}, \qquad \frac{n_{T,\mathrm{new}}}{n_{T,\mathrm{old}}} = \frac{\hat{\sigma}_{\mathrm{int}}^2}{\sigma_{\mathrm{plan}}^2}.

The same proportional change applies to both groups because the allocation ratio is unchanged. In particular, the variance ratio—not the ratio of standard deviations—is the multiplier. For example, a 25% increase in standard deviation corresponds to a variance increase of \(1.25^2=1.5625\), and therefore a 56.25% increase in the unrounded sample size.

Confirm the design inputs.
Use the prespecified \(\Delta\), \(\alpha\), power, and allocation ratio. Confirm that the intended final analysis matches the formula being used.

Obtain the interim variance estimate.
Use the prespecified blinded method and record the estimate in squared outcome units.

Recalculate and round.
Substitute the interim estimate, calculate each group size, and round up so rounding does not reduce the calculated target.

Apply design constraints.
Check the protocol's minimum and maximum sample sizes and any rules for handling an update. Document the final targets and the inputs used.

Worked Calculations

Worked Example: Updating a Two-to-One Allocation

Hypothetical example. A trial plans a two-sided comparison with \(\alpha=0.05\), power \(0.90\), target difference \(\Delta=4\) outcome units, and treatment-to-control allocation ratio \(r=2\). The prespecified blinded interim procedure yields a corrected pooled variance estimate of \(64\) outcome-units squared. Find the recalculated group sizes.

For these inputs, use \(z_{0.975}=1.960\) and \(z_{0.90}=1.282\). Substitution gives

n_C = \frac{(1+1/2)(64)(1.960+1.282)^2}{4^2} = \frac{96(10.510564)}{16} = 63.063384.

Rounding up gives a control-group target of \(64\). With \(r=2\), the corresponding treatment-group target is \(2(64)=128\), for a total of \(192\) participants. These targets retain the stated difference, significance level, power, and allocation ratio under the formula's assumptions.

Worked Example: Comparing the Update With the Original Plan

Hypothetical example. Use the same \(\alpha=0.05\), power \(0.90\), \(\Delta=4\), and \(r=2\), but suppose the original planning variance was \(49\) outcome-units squared. The original unrounded control target was

n_{C,\mathrm{old}} = \frac{(1+1/2)(49)(3.242)^2}{16} = \frac{73.5(10.510564)}{16} \approx 48.28.

The original rounded targets would be \(49\) control and \(98\) treatment. Replacing variance \(49\) with the interim estimate \(64\) multiplies the unrounded sizes by \(64/49\approx1.306\). Thus the control target changes from about \(48.28\) to \(63.06\) before rounding, consistent with the recalculation above. The update reflects a 30.6% increase in estimated variance, not a change in the target effect.

Worked Example: Equal Allocation and a Different Power Target

Hypothetical example. A separate trial uses equal allocation, \(\Delta=3\) units, \(\alpha=0.05\), and power \(0.80\). Its interim variance estimate is \(49\) units squared. Here \(r=1\), \(z_{0.975}=1.960\), and \(z_{0.80}=0.8416\). The required size per group is

n_C=n_T = \frac{(1+1)(49)(1.960+0.8416)^2}{3^2} = \frac{98(7.84896256)}{9} = 85.46648.

Rounding up gives \(86\) participants per group, or \(172\) in total. The variance is expressed in squared outcome units, while \(\Delta^2\) is also in squared units, so their ratio is dimensionless; the resulting sample size is a count.

Worked Example: Rounding With Unequal Allocation

Hypothetical example. Suppose \(\alpha=0.05\), power \(0.80\), \(\Delta=2.5\) units, and \(r=1.5\). The interim variance estimate is \(25\) units squared. Using \(z_{0.975}=1.960\) and \(z_{0.80}=0.8416\),

n_C = \frac{(1+1/1.5)(25)(2.8016)^2}{2.5^2} = \frac{(1.6667)(25)(7.84896256)}{6.25} \approx 52.3264.

Round the control size up to \(53\), then apply the allocation ratio: \(1.5(53)=79.5\), which must be rounded up to \(80\) treatment participants. The resulting targets are \(53\) control and \(80\) treatment, total \(133\). Rounding each group up preserves at least the calculated continuous-valued targets and keeps the realized allocation close to the planned ratio.

Checks, Alternatives, and Limitations

Before adopting an updated target, check that the variance estimate is in the correct units and corresponds to the variance convention used in the formula. Also verify that the same target difference, two-sided or one-sided test, power, and allocation ratio have been used consistently. A very large or very small update deserves a careful review of the masked variance calculation and the protocol's sample-size limits; the plug-in formula does not establish that an unusual estimate is reliable.

A common mistake is to multiply the old sample size by the ratio of standard deviations. If the assumed standard deviation rises from \(8\) to \(10\), this mistake uses \(10/8=1.25\). The correct variance ratio is \(10^2/8^2=1.5625\), so the unrounded sample size rises by 56.25%, not 25%. Another mistake is to round to the nearest integer or round down. Because the formula gives a minimum continuous-valued requirement, round the group targets up.

The principal alternative is to retain the original sample size regardless of the interim variance. That avoids a variance-driven change in enrollment but may leave the trial underpowered if the true variance is substantially larger than planned, or larger than necessary if it is smaller. A more exact fixed-sample calculation can use the noncentral \(t\) distribution and iterate over integer sample sizes rather than rely on the normal approximation. For a group sequential design, a design-specific power calculation must incorporate the monitoring boundaries and planned looks rather than treat the fixed-sample formula as a complete sequential calculation.

Finally, a recalculated size is not a guarantee of the realized power: the interim variance estimate is uncertain, and the true variance may differ from it. Nor does using a blinded estimate alone prove that the trial's type I error is unchanged. The adaptation and final analysis must follow a prespecified, statistically appropriate procedure for the design. The calculation here isolates the variance-to-sample-size step; it does not replace those design safeguards.

Key takeaway: For fixed \(\Delta\), \(\alpha\), power, and allocation ratio, the normal-theory sample size is proportional to the variance. Substitute the prespecified interim variance estimate, calculate each group target, and round up; then verify that the formula matches the trial's actual final analysis plan.

Check Your Understanding

Answer each question before reading the solutions.

With all other design inputs fixed, an interim variance estimate is 1.44 times the planning variance. By what factor do the unrounded group sizes change?
For equal allocation, \(\Delta=3\), \(\alpha=0.05\), power \(0.80\), and variance \(36\), calculate the unrounded and rounded size per group. Use \(z_{0.975}=1.960\) and \(z_{0.80}=0.8416\).
Why is it incorrect to use the ratio of standard deviations as the sample-size multiplier?
A fixed-sample calculation is used for a trial that will have group sequential monitoring. Does the calculation alone establish the sequential design's power and type I error?

Solutions. 1. The unrounded sizes increase by a factor of \(1.44\), because sample size is proportional to variance when the other inputs are fixed. 2. The size per group is \(\{2(36)(2.8016)^2\}/9=565.1253/9\approx62.79\), so round up to \(63\) per group. 3. The formula contains \(\sigma^2\), not \(\sigma\); the multiplier must be the variance ratio, which is the square of the standard-deviation ratio. 4. No. The fixed-sample calculation does not incorporate sequential monitoring boundaries, so a design-specific calculation is needed to establish the sequential design's power and type I error.