Why the Information Set Matters
A sample size re-estimation rule uses interim trial information to revise the planned number of participants. A central design choice is whether the rule can use treatment assignments and arm-specific outcomes. With blinded re-estimation, the people applying the rule do not see which treatment produced each outcome; the rule uses treatment-combined data. With unblinded re-estimation, an authorized person or committee can see treatment assignments and arm-specific interim results.
These are not merely different ways to calculate the same quantity. They expose the rule to different information. Blinded rules are often designed to update a nuisance parameter, such as outcome variability, while limiting access to the emerging treatment contrast. Unblinded rules can use the observed contrast directly, but that contrast is noisy, and revealing it creates statistical and operational risks. Either approach must be specified in advance: who sees which data, what calculation is performed, and how the result changes the maximum or target sample size.
A Planning Formula to Make the Difference Concrete
Consider a hypothetical two-arm trial with a continuous outcome. Assume independent participants, equal allocation, normally distributed outcomes with a common standard deviation \(\sigma\), a prespecified clinically relevant mean difference \(\delta>0\), and a two-sided test of \(H_0:\mu_1-\mu_0=0\) at significance level \(\alpha\). For a normal-approximation calculation with target power \(1-\beta\), the approximate number of participants required per arm is
where \(z_q\) is the \(q\)-quantile of the standard normal distribution. The formula assumes the stated allocation and variance model; in an actual design, the planned test and any small-sample or other design adjustments must also be reflected in the calculation. Rounding \(n\) upward gives a whole-participant target. The formula shows why an estimate of \(\sigma\) can matter under either kind of rule, and why a rule that substitutes an interim treatment difference for \(\delta\) can change the target sharply.
Worked Example: The Initial Fixed-Design Target
Suppose a hypothetical trial plans a two-sided test with \(\alpha=0.05\), 90% power, a target difference of 5 outcome units, and a planning standard deviation of 10 outcome units. Use \(z_{0.975}=1.960\) and \(z_{0.90}=1.282\).
The assumptions for the formula are equal allocation, independent outcomes, and the common-variance normal model described above. Substituting the planning values gives
Rounding up gives a target of 85 participants per arm, or 170 in total. This is a planning benchmark under the stated model and assumptions, not a guarantee that the trial will attain 90% power for every possible realized data set.
Blinded Re-Estimation: Update Without the Arm Contrast
A blinded rule can use a prespecified summary of treatment-combined interim outcomes to update a nuisance parameter. For example, the rule might produce an estimate of outcome variability and insert it into the sample size formula while leaving the target difference \(\delta\) unchanged. The procedure must be defined carefully: treatment labels remain unavailable to the re-estimation process, and the calculation must not silently use arm-specific results.
Blinding limits direct access to the observed treatment difference, but it does not make every treatment-combined summary independent of that difference. If the treatment has an effect, the combined outcome distribution can reflect both within-arm variability and separation between the arms. Therefore, a blinded summary intended to estimate within-arm variability needs a method whose assumptions and behavior are appropriate for the design. Simply treating any overall spread as the common within-arm standard deviation can be misleading.
Worked Example: A Blinded Update to the Variability Estimate
Continue the hypothetical design above. Suppose a prespecified blinded procedure, applied to interim treatment-combined data, returns an estimated standard deviation of 11 outcome units. Assume for this calculation that the procedure supplies a suitable estimate of the common within-arm standard deviation. The rule keeps \(\delta=5\), \(\alpha=0.05\), and 90% power unchanged.
The update is to the nuisance-parameter estimate, not to the assumed treatment difference. Substituting \(\sigma=11\) gives
Rounding up gives 102 participants per arm, or 204 in total. This exceeds the initial target of 85 per arm because the updated variability estimate is larger. The calculation is justified only if the blinded procedure's output estimates the variance parameter required by the planning model; a treatment-combined spread that includes treatment separation cannot automatically be interpreted that way.
Unblinded Re-Estimation: Arm-Specific Information
An unblinded rule can access treatment-specific summaries, including the interim mean difference. One possible rule is a plug-in calculation that replaces the planning difference \(\delta\) with the observed interim difference. Because the sample size formula is inversely proportional to the squared difference, a small observed difference can produce a much larger target, while a large observed difference can produce a smaller one. This arithmetic is straightforward; deciding whether such a rule is statistically and operationally acceptable is not.
The interim difference is an estimate based on a subset of the trial, not a known treatment effect. It can be highly variable, particularly when few participants have contributed outcomes. A rule that reacts strongly to that estimate can therefore produce large changes driven by random fluctuation. Moreover, people who learn the interim contrast may change their behavior, consciously or not, or may communicate information that compromises blinding. Unblinded access needs a prespecified governance plan, restricted data access, and a clear record of the decision process.
Worked Example: How an Interim Difference Changes a Plug-In Target
Suppose, for illustration, that 40 participants per arm have completed outcome assessment in the hypothetical trial. An authorized unblinded analyst sees interim means of 102.0 and 100.8 outcome units, so the observed difference is 1.2 units. For this example only, consider a rule that plugs the absolute observed difference into the original formula, keeps \(\sigma=10\), \(\alpha=0.05\), and 90% power, and does not allow the target to fall below the 40 participants already observed per arm.
Using the observed difference as the plug-in value gives
The rounded target is 1,460 participants per arm. The dramatic increase is a consequence of the squared denominator: a difference of 1.2 is much smaller than the planned difference of 5. This result does not establish that the true treatment difference is 1.2, nor does it show that increasing the trial to this size is a sound decision. It illustrates the sensitivity of an unblinded plug-in rule to a noisy interim estimate.
For contrast, if the interim difference were 8 units under the same illustrative rule, the calculation would be
Rounding gives 33 per arm, but the rule's floor of 40 per arm would prevent a reduction below the participants already observed. The example shows why floors, caps, and the exact adaptation rule must be specified, even though choosing those limits is separate from the arithmetic itself.
Derivation: Why the Difference Estimate Has Leverage
The sensitivity in the preceding example follows directly from the planning formula. Holding \(\sigma\), \(\alpha\), and power fixed, define \(C=2\sigma^2(z_{1-\alpha/2}+z_{1-\beta})^2\). Then the calculated target is \(n=C/\delta^2\). For two positive candidate differences \(\delta_a\) and \(\delta_b\),
Thus, changing the input difference from 5 to 1.2 multiplies the unrounded target by \((5/1.2)^2\), approximately 17.36. This is a mathematical property of the calculation, not evidence that the interim estimate is a reliable estimate of the effect that will be observed at the end. A rule that uses the interim difference must account for the uncertainty and consequences of using it; it cannot justify treating the interim value as fixed truth.
Information Requirements, Diagnostics, and Safeguards
The distinction can be summarized by asking what data the rule needs and who can see them.
| Feature | Blinded re-estimation | Unblinded re-estimation |
|---|---|---|
| Interim information | Treatment-combined outcomes or a summary that does not reveal arm-specific results | Arm-specific outcomes and potentially the observed treatment contrast |
| Typical target of the update | A nuisance parameter, with the design effect assumption held fixed | A nuisance parameter, an observed treatment contrast, or a rule using both |
| Key information risk | The chosen combined-data summary may not estimate the intended nuisance parameter well | The noisy observed contrast may drive unstable changes and reveal emerging results |
| Access controls | Limit access to any data or summaries that could reveal assignments | Restrict access to designated personnel or an independent committee |
Diagnostics should be tied to the prespecified rule, rather than improvised after seeing results. Check that the planned interim sample and outcome data are available as specified; that the calculation uses the intended units, allocation, and parameter estimate; and that the resulting target obeys the prespecified bounds. For a blinded procedure, assess whether its summary is plausibly estimating the nuisance parameter required by the formula and whether treatment separation could contaminate it. For an unblinded procedure, assess how much the target changes across plausible interim contrasts, especially when few participants contribute data.
A blinded rule is often preferable when the design only needs a nuisance-parameter update and treatment-effect information is not needed. It does not eliminate all risk: a poor nuisance estimate, an unsuitable calculation, or accidental disclosure of arm information can still undermine the procedure. An unblinded rule may be considered when the design explicitly requires arm-specific information, but it needs tighter safeguards and a clear rationale for using a noisy interim contrast. Neither approach is automatically superior in every trial.
Common Pitfall and Limitations
A second mistake is to interpret an unblinded plug-in calculation as an accurate forecast of the final treatment effect. The interim contrast is uncertain, and a large sample size change can be driven by chance variation. A formula alone does not establish that the resulting rule preserves the intended operating characteristics or that the conduct of the trial remains protected from information leakage. Those questions require evaluation of the complete prespecified design and its implementation.
The normal-approximation calculations here are deliberately limited to a continuous outcome, equal allocation, and a common-variance model. They illustrate the information distinction, not a universal re-estimation recipe. The main practical comparison is therefore not “which formula is shorter?” but “which interim information is necessary, how trustworthy is the resulting estimate, and what protections keep the rule and trial conduct aligned with the design?”
Check Your Understanding
Answer each question, then compare with the solutions below.
- In one sentence, distinguish the information used by a blinded rule from the information used by an unblinded rule.
- In the planning formula, what happens to the calculated sample size if the assumed difference is halved while all other inputs stay fixed?
- Why is a treatment-combined outcome spread not automatically an estimate of within-arm variability?
- A rule uses an observed interim treatment difference to update the target. Name one statistical risk and one operational safeguard.
Solutions. (1) A blinded rule uses treatment-combined information without arm-specific results, whereas an unblinded rule can use arm-specific outcomes or their contrast. (2) Since \(n\) is proportional to \(1/\delta^2\), halving \(\delta\) multiplies the calculated target by \(2^2=4\). (3) If treatment effects separate the arm distributions, the combined spread can reflect that separation as well as within-arm variability, so it may not estimate the variance parameter the formula requires. (4) The noisy interim contrast can cause an unstable target change; restrict access to designated independent personnel and prespecify the calculation and decision process.