What Makes Re-Estimation Unblinded?
A sample-size update is unblinded when the information used to make the update distinguishes the randomized treatment groups and reveals something about their comparative outcomes. That information can include an interim treatment-effect estimate, its standard error, or a statistic calculated from the two groups separately. It is not enough that someone has seen interim data: the key question is whether the information used for the adaptation preserves the treatment comparison.
This matters because comparative interim results are influenced by both the true treatment effect and random variation. A rule that uses them can change the final sample size in response to the observed effect. Such a rule is an adaptive design: the design has a planned way of changing in response to accumulating data. Adaptation alone does not determine whether the final hypothesis test has its intended error rate. That depends on the full design and the inferential procedure used with it.
In contrast, a blinded re-estimation uses information that does not disclose the treatment comparison, often to revise a nuisance parameter such as a pooled event rate or variance. The essential distinction is not whether the person doing the calculation knows the study is randomized. It is whether the adaptation uses treatment-specific information in a way that can reveal the comparative effect.
What Interim Information Enters the Rule?
Let \(k\) denote a prespecified interim look and let \(\hat{\Delta}_k\) be the estimated treatment effect at that look, with the sign convention defined in the protocol. For a continuous outcome, it might be the difference in sample means; for a binary outcome, it might be the difference in event proportions. The estimate must be interpreted in the outcome's units and direction. Its uncertainty, often summarized by a standard error, also matters: the same estimate based on few observations is less precise than one based on many.
An adaptation may use the interim effect estimate alone, or a function of the estimate and its standard error. It may also use nuisance estimates, such as a variance, provided their role is specified. The adaptation should identify exactly which data are used and whether estimates are calculated using all randomized participants, only participants with observed outcomes, or another stated analysis set. If interim outcome availability differs between groups, the comparison may be particularly unstable.
The observed \(\hat{\Delta}_k\) is not the true treatment effect \(\Delta\). It is a random estimate. In particular, using a noisy interim estimate as if it were the known effect can produce a highly variable sample-size rule. An estimate near zero can imply an extremely large target, while an unusually large estimate can imply a small target. A prespecified rule therefore needs to say what happens at such extremes, rather than leaving the decision to judgment after results are seen.
A Planning Calculation Behind an Adaptation
A basic normal-outcome calculation makes the role of the effect estimate explicit. Suppose the final comparison will be a two-sided test of \(H_0:\Delta=0\), with equal allocation, independent observations, a common known standard deviation \(\sigma\), and a fixed final analysis. Assume the outcome is normally distributed within each group, or that the normal approximation for the test is adequate. If \(n\) is the number of participants per group, the standard error of the difference in means is \(\sigma\sqrt{2/n}\).
For a target effect \(\Delta_*\), the standardized mean under that effect is \(\Delta_*\sqrt{n/(2\sigma^2)}\). To obtain approximate power \(1-\beta\) at two-sided significance level \(\alpha\), set this mean about \(z_{1-\alpha/2}+z_{1-\beta}\), where \(z_q\) is the \(q\)-quantile of the standard normal distribution. Solving for \(n\) gives:
This is a planning approximation, not a guarantee of power for every realized interim result. In an unblinded re-estimation rule, a protocol might substitute \(|\hat{\Delta}_k|\) for \(|\Delta_*|\), subject to stated bounds. That substitution makes the sample-size target data-dependent. It does not make the interim estimate the true effect, nor does it by itself specify a valid final test.
Worked Example: Re-Estimating for a Continuous Outcome
Suppose a hypothetical trial plans a two-sided comparison of mean change in a clinical score, measured in points. Assume equal allocation, independent normally distributed outcomes with known common standard deviation \(\sigma=12\) points, significance level \(\alpha=0.05\), and target power \(1-\beta=0.80\). The design team considers a rule that uses the interim estimated mean difference as the target effect.
First calculate the planning numerator. Using \(z_{0.975}=1.960\) and \(z_{0.80}=0.8416\):
If the prespecified target difference is \(4.5\) points, the required sample size is approximately:
Rounding up gives 112 participants per group. Now suppose the planned interim look produces an estimated treatment difference of \(3.5\) points. Under the proposed plug-in rule, the updated target is:
so the rule rounds up to 185 participants per group. If 40 per group have already been randomized, the arithmetic implies 145 additional participants per group, assuming all randomized participants count toward the target. This is a sample-size calculation under the stipulated rule; it is not evidence that the true effect is 3.5 points or a calculation of conditional power.
Prespecification Is Part of the Design
A rule is prespecified when its operational details are fixed before the unblinded interim comparison is available. The protocol or statistical analysis plan should make it possible to determine the revised sample size from the permitted interim information without inventing a new decision in response to the result. Relevant details include the look timing, data cut-off, estimator, sample-size formula, treatment of missing outcomes, minimum and maximum sample sizes, rounding convention, and which roles may see unblinded results.
Prespecification is more than writing down a formula. It also means specifying the final analysis strategy and how the adaptive design will be evaluated for operating characteristics, including type I error under \(H_0\) and power under relevant alternatives. The final inferential procedure must be appropriate for the actual adaptive design. A conventional fixed-sample test should not be assumed to retain its nominal properties merely because the sample-size rule was written in advance.
Worked Example: A Binary-Outcome Plug-In Rule
Consider a hypothetical equal-allocation trial with a binary outcome. At the planned interim look, 100 participants per group have outcome data: 24 of 100 in the treatment group and 16 of 100 in the control group experienced the event. Define the treatment effect as treatment risk minus control risk. The interim estimates are:
Thus the observed difference is 8 percentage points, with a positive value indicating a higher event risk in the treatment group under this sign convention. For illustration, suppose a prespecified approximate planning calculation uses these two interim proportions as plug-in rates, with \(\alpha=0.05\), power 0.80, and equal group sizes:
Substituting the values gives a variance sum of \(0.24(0.76)+0.16(0.84)=0.1824+0.1344=0.3168\), and therefore:
The approximate target is 389 participants per group, before any prespecified cap or other design adjustment. This calculation uses the separate interim event proportions and their difference, so it is unblinded re-estimation. It is only an illustrative normal-approximation calculation; sparse events, continuity corrections, the exact planned test, and other design details can change an appropriate rule.
Planned Adaptation Versus Improvised Change
A data-dependent sample-size change is not automatically unplanned. If a protocol specifies a rule mapping interim comparative results to a sample size, applying that rule is the planned adaptation. By contrast, seeing an interim difference and then deciding to increase enrollment by an amount chosen through informal discussion is an unplanned design change, even if the decision seems scientifically reasonable.
| Feature | Prespecified adaptation | Unplanned change |
|---|---|---|
| Trigger and timing | Look and data cut-off set in advance | Look or timing chosen after reviewing results |
| Use of interim comparison | Inputs and estimator defined in advance | Result interpreted or summarized ad hoc |
| Sample-size decision | Formula, bounds, and rounding specified | Target selected after seeing the result |
| Final inference | Procedure evaluated for the adaptive design | Fixed-design inference assumed without justification |
For example, a protocol might specify one interim analysis after 40 participants per group have completed follow-up, calculate a defined treatment-effect estimate, apply a stated formula, and cap the final target at 220 per group. If the rule yields 185 per group, continuing to that target is not an improvised increase. In contrast, changing the look date, switching the effect estimator, or raising the cap after seeing that the result appears promising changes the rule itself.
Checks, Alternatives, and Limitations
Before applying an unblinded rule, check that the planned interim data match the stated analysis set and endpoint, that group labels and effect direction are correct, and that the calculation uses the specified units and rounding. Check for missing or delayed outcomes and for substantial departures from the assumptions used in the planning formula. A calculated target that is implausibly small or large, or that changes sharply with a small change in \(\hat{\Delta}_k\), is a warning to inspect the rule and its bounds, not an invitation to revise it after seeing the data.
The main alternative is a blinded nuisance-parameter re-estimation, which can revise a sample size using pooled information without using a group contrast. A fixed sample size avoids outcome-driven adaptation but cannot respond to the information gained about the effect or nuisance parameters during the trial. Unblinded re-estimation can make the design responsive to interim comparative results, but it also creates greater demands for prespecification, access control, and valid final inference. Neither approach is universally preferable.
A plug-in rule has an important limitation: it treats an interim estimate as an input to a planning formula, even though that estimate is uncertain and can be an overestimate or underestimate of \(\Delta\). The resulting target may therefore be unstable and need not achieve the nominal power for the unknown true effect. Caps and floors make the rule operationally finite, but do not by themselves solve inferential problems. The adaptation and final analysis should be evaluated together under plausible data-generating conditions.
A common mistake is to reason, “The interim treatment difference is smaller than expected, so I can simply increase the sample size until the original test has the desired power.” This skips the uncertainty in the observed difference and assumes that a fixed-sample test remains valid after the sample size has depended on interim outcomes. The correction is to use a fully specified adaptation rule, with prespecified limits and an inferential procedure justified for the resulting design. The observed effect is information for the rule, not a known parameter or a license to change the design informally.
Check Your Understanding
Answer each question, then compare your reasoning with the short solutions below.
- A statistician recalculates sample size from the pooled outcome variance without seeing treatment labels. Is this unblinded re-estimation? Why?
- At an interim look, \(\hat{\Delta}_k\) is close to zero. Does that establish that the true effect is zero? What practical issue can arise for an effect-based rule?
- A protocol specifies the look, estimator, formula, cap, and rounding. Applying the rule gives a larger target than originally planned. Is that change necessarily unplanned?
- Why is prespecification alone insufficient to justify using an ordinary fixed-sample final test?
Solutions: 1. No. The calculation uses pooled information without a treatment-group contrast; it is blinded with respect to the comparative effect. 2. No. The interim estimate is random and uncertain. A near-zero estimate can make a plug-in sample-size target very large, so the rule needs stated limits and its performance should be evaluated. 3. No. If the change follows the prespecified rule, it is the planned adaptation, subject to the design's stated procedures. 4. Because the sample size depends on interim data, the design's operating characteristics may differ from those of a fixed-sample design; the final inferential procedure must be justified for the adaptive design.