Unblinded Sample Size Re-Estimation

What an Unblinded Re-estimation Uses

A sample size re-estimation changes the planned enrollment using information observed during the trial. In an unblinded re-estimation, the interim information includes outcomes identified by treatment arm, so the calculation can use an estimated treatment effect. That can make the recalculation responsive to the effect seen so far, but it also means someone must be able to see treatment-identified results.

The distinction matters operationally. A statistician may calculate an interim contrast from arm-specific data, while sites need to know only how many more participants to enroll. Giving every operational role the contrast or arm-specific summaries is not required just because the final sample size depends on them. The permitted access should follow the prespecified procedure, including who calculates the update and who, if anyone, reviews or authorizes it.

Definition: In unblinded sample size re-estimation, an interim rule uses a treatment-identified estimate of the effect to determine a revised target sample size. The estimate may be a difference between treatment arms, or another prespecified effect measure appropriate to the endpoint. The revised target is a total sample size, including participants already enrolled; it is not automatically the number still to enroll.

A Standard Two-Arm Calculation

For a concrete calculation, suppose the primary endpoint is continuous, higher values are better, and the planned analysis compares two independent treatment groups with equal allocation. Assume the endpoint is approximately normal in each group, with a common, known standard deviation \(\sigma\); observations are independent; and the target is a two-sided test of equal means at significance level \(\alpha\). Let \(\delta\) be the treatment difference the trial is designed to detect, and let \(1-\beta\) be the target power.

Under these assumptions, if each group has \(n\) participants, the standard error of the difference in sample means is \(\sigma\sqrt{2/n}\). The usual normal-approximation planning formula is:

n \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, and \(n\) is the number per group. This formula is an approximation: it uses the upper-tail critical value plus the target-power quantile and neglects the usually small opposite-tail probability under the planning alternative. The exact normal-theory power can instead be computed from both tails.

The formula requires a prespecified planning value for \(\sigma\), a target effect on the endpoint's scale, equal allocation, and the stated model and test. If the endpoint, allocation, variance model, or analysis differs, the sample size calculation must be changed accordingly. In an unblinded re-estimation, the planned \(\delta\) may be replaced by an interim estimate, but doing so does not make that interim estimate precise or unbiased as a predictor of the final effect.

Worked Example: Updating the Target From an Interim Difference

Suppose a hypothetical trial plans a two-sided \(\alpha=0.05\) test with 90% power to detect a 5-point mean difference. The common standard deviation is taken to be 10 points. At an interim analysis, 50 participants per group have complete outcomes: the treatment-group mean is 102 points and the control-group mean is 99 points. Assume the endpoint and analysis meet the normal, independence, common-variance, and equal-allocation conditions above.

The interim estimated treatment effect is \(\hat{\delta}=102-99=3\) points. Using \(z_{0.975}=1.960\) and \(z_{0.90}=1.282\), the plug-in target is:

n \approx \frac{2(10^2)(1.960+1.282)^2}{3^2} = \frac{200(3.242)^2}{9} = \frac{2102.1}{9} = 233.6

Rounding up gives a target of 234 participants per group, or 468 total. Since 50 per group have already contributed interim outcomes, this target would mean enrolling 184 more per group, assuming those participants count toward the final analysis and no other adjustments are required. The update does not establish that the true treatment effect is 3 points; it is the result of applying the prespecified calculation to a noisy interim estimate.

Why the Formula Has This Form

The sample size expression follows from the sampling distribution of the difference in means. Under the planning alternative with a true difference \(\delta>0\), the standardized statistic has approximately a normal distribution with variance 1 and mean \(\lambda=\delta\sqrt{n/(2\sigma^2)}\). A two-sided test rejects when the statistic exceeds \(z_{1-\alpha/2}\) or is below \(-z_{1-\alpha/2}\). For the upper-tail part of power to be approximately \(1-\beta\), its mean should be at least \(z_{1-\alpha/2}+z_{1-\beta}\). Thus:

\delta\sqrt{\frac{n}{2\sigma^2}} \ge z_{1-\alpha/2}+z_{1-\beta} \quad\Longrightarrow\quad n \ge \frac{2\sigma^2\left(z_{1-\alpha/2}+z_{1-\beta}\right)^2}{\delta^2}

This is a planning justification, not a guarantee about realized power. The calculation is especially sensitive to \(\delta\): it appears squared in the denominator. A smaller effect estimate can therefore imply a much larger target, while a larger estimate can imply a smaller one.

Prespecify Floors, Caps, and What the Result Means

A trial should specify how the calculated target becomes an operational decision. Common elements include a minimum and maximum total sample size, a rule for rounding, and whether the trial can reduce its original target. The protocol should also state what happens if the estimated effect is zero or points in the opposite direction: the sample size formula uses the squared effect magnitude, but a large absolute estimate in the unexpected direction should not mechanically be treated as evidence supporting the intended benefit.

A cap may prevent an impractically large recalculation, but it does not preserve the original target power when the uncapped requirement exceeds the cap. Likewise, a minimum can prevent a calculated target below that floor from reducing enrollment too far. The rule should specify whether a target may fall below the initial planned size and how already enrolled participants count.

Worked Example: Applying a Prespecified Floor

In another hypothetical trial, let \(\sigma=10\), \(\alpha=0.05\), and target power be 90%, as before. At interim, the treatment mean is 107 points and the control mean is 100 points, giving \(\hat{\delta}=7\) points. The protocol allows a reduction from the initial target but requires at least 80 participants per group. Suppose 50 per group have already been enrolled and are included in the final analysis.

The unconstrained calculation is:

n \approx \frac{2(10^2)(1.960+1.282)^2}{7^2} = \frac{2102.1}{49} = 42.90

Rounding up gives 43 per group, but the prespecified floor is 80 per group. The operational target is therefore 80 per group, not 43. With 50 already enrolled per group, the trial would enroll 30 more per group. This decision follows the stated floor; it does not mean that the interim estimate proves the true effect is 7 points or that the floor is statistically necessary in every design.

An upper cap works in the opposite direction. If the formula calls for 234 per group but the protocol caps enrollment at 120 per group, the target is 120 per group. That cap limits feasibility and exposure; it does not deliver the nominal 90% power calculated for 234 per group under the planning assumptions.

Who Needs Treatment-Identified Results?

To calculate a treatment-effect estimate, a designated person or group must be able to connect outcomes to treatment assignments. A masked analyst who sees only pooled outcomes cannot calculate the arm-specific difference in the example. In practice, this means that at least the assigned statistician or data analyst performing the calculation needs the treatment-identified interim data or an equivalent treatment-identified summary.

Whether an independent reviewer or decision-making body also needs the estimate depends on the prespecified process. If that role must review the calculation or make a decision that explicitly depends on the treatment contrast, it needs access to the information required for that responsibility. If its task is only to confirm the revised enrollment target after the independent calculation, it may receive the target and its operational implications without receiving arm-specific outcomes or the estimated effect. The same distinction applies to sponsor personnel and sites: they need the revised enrollment instruction to implement it, but do not automatically need the treatment contrast.

Worked Example: Separating Calculation From Implementation

Consider a hypothetical interim dataset with 30 participants per group, a treatment mean of 104.2 points, a control mean of 100.7 points, and a prespecified \(\sigma=10\). The trial uses the same two-sided \(\alpha=0.05\) and 90% power calculation. Assume the outcomes are independent and the normal, common-variance model is appropriate for this illustrative calculation.

The identified interim difference is \(104.2-100.7=3.5\) points. The calculated target is:

n \approx \frac{2(10^2)(1.960+1.282)^2}{3.5^2} = \frac{2102.1}{12.25} = 171.6

Rounding up gives 172 participants per group, or 344 total. With 30 per group already enrolled, the operational instruction is to enroll 142 more per group. The statistician or analyst assigned to compute this update needs the treatment identities and arm-specific means. If a designated reviewer must verify the effect-based calculation, that reviewer also needs the treatment-identified information specified for the review. Sites can receive an instruction to enroll 142 more participants per group without receiving the means, the 3.5-point contrast, or the treatment identities.

The principle is minimum necessary access for the assigned task, not that every role connected to a sample size decision must see the interim effect. Conversely, if a role is expected to make an independent judgment about the effect itself, providing only the revised number would not supply the information needed for that judgment.

Inference, Diagnostics, and Limitations

An interim estimate is variable, especially when few outcomes are available. Replacing a target effect with a noisy interim estimate can produce a very large target after an unexpectedly small estimate, or a smaller target after an unusually large estimate. The update should therefore be checked against the prespecified bounds, the accrued sample size, the direction and scale of the effect, and the assumptions used for the calculation. Unexpectedly different arm sizes, missing outcomes, changing outcome definitions, or an interim estimate on the wrong scale are signals to pause and verify the data and rule before acting.

Unblinded re-estimation also raises an inferential issue. The final test is no longer necessarily a fixed-sample test whose sample size was chosen independently of observed treatment differences. Depending on the re-estimation rule and final analysis, using an ordinary fixed-sample test without accounting for the adaptation can fail to control the intended type I error rate. This is not a claim that every unblinded update inflates error: validity depends on the full design and analysis. A prespecified, appropriate inferential procedure is needed to justify the final error control; recalculating sample size alone does not provide that justification.

Blinded nuisance-parameter re-estimation is a different approach: it can update quantities such as variability without using an identified treatment contrast. It may avoid the same direct exposure to treatment-effect information, but it cannot use an unblinded interim effect estimate in the way illustrated here. Fixed sample size avoids interim sample-size decisions altogether, but may be inefficient if planning assumptions prove inaccurate. Neither alternative removes the need to match the design, access plan, and final analysis.

Common Pitfalls

Treating the interim estimate as the true effect. The observed contrast is uncertain. Use it only as the input specified by the recalculation rule, not as a guarantee of the final effect.
Confusing the target with additional enrollment. A target of 172 per group with 30 already enrolled means 142 more per group, not 172 more.
Assuming a cap preserves power. A cap below the calculated requirement limits enrollment; it does not achieve the power associated with the uncapped sample size.
Giving the effect estimate to everyone who needs an operational decision. Sites may need the enrollment target without needing treatment identities or arm-specific summaries.
Assuming the final fixed-sample test is automatically valid. The adaptation and final inferential method must be assessed together under the prespecified design.

Key takeaway: An unblinded sample size update substitutes an interim treatment-effect estimate into a prespecified sample size rule, subject to its assumptions and bounds. Only roles whose assigned calculations or reviews require treatment-identified information need access to that information; implementation roles can often receive the revised target alone.

Check Your Understanding

Answer each question, then compare with the brief solutions.

With \(\sigma=8\), two-sided \(\alpha=0.05\), 80% power, and an interim estimated difference of 4 points, calculate the approximate target per group. Use \(z_{0.975}=1.960\) and \(z_{0.80}=0.842\).
A recalculated target is 150 per group, and 45 per group are already enrolled. How many additional participants per group are required if all accrued participants count toward the target?
Why does an upper sample size cap not guarantee the power calculated from the uncapped formula?
Which role needs treatment-identified results if the prespecified procedure assigns it to calculate the arm-specific interim effect? Must every site receive that effect estimate?
Why can a final ordinary fixed-sample test require special justification after an unblinded sample size update?

Solutions. 1. \(n\approx 2(8^2)(1.960+0.842)^2/4^2=128(2.802)^2/16=62.8\), so round up to 63 per group. 2. \(150-45=105\) additional participants per group. 3. The cap can be smaller than the sample size required under the planning assumptions, so it limits enrollment without delivering the uncapped target power. 4. The assigned statistician or analyst needs treatment-identified information sufficient for the calculation; sites can receive the enrollment instruction without the treatment contrast. 5. Because sample size depends on observed treatment differences, the adaptation and final test together may not have the error properties of a fixed-sample design; validity must be established for the prespecified procedure.