Why Information Requirements Depend on the Decision
An adaptive decision should not receive more interim information than it needs. A sample size update based on a nuisance parameter such as a pooled standard deviation may be possible without revealing which outcomes came from which treatment arm. Selecting an arm because its outcomes look better, by contrast, requires arm-specific information. A decision based on an interim treatment effect requires a treatment contrast and its uncertainty. These distinctions matter both statistically and operationally: access to more detailed results can expose emerging differences to people whose behavior could affect trial conduct.
The introductory material on information tiers and treatment-selection information distinguishes the kinds of interim information and the roles that may use them. Here the practical task is to match the information to the decision. The goal is not to make every adaptation blinded, nor to assume that an unblinded analysis is automatically invalid. It is to identify what is necessary, who needs it, and what safeguards are needed for the chosen rule.
Match the Information to the Adaptation
A blinded pooled summary combines observations across treatment arms without showing which arm produced each observation. Depending on the design and the updating rule, that may be sufficient to estimate a nuisance quantity used in a sample size calculation. It does not ordinarily reveal which treatment is doing better. Arm-specific summaries separate the data by treatment, while an estimated treatment effect compares treatments directly, with uncertainty measured by a standard error or another prespecified measure.
The distinction is about the decision rule, not simply the name of the adaptation. If a rule changes sample size because the estimated treatment effect is too small, it needs treatment-effect information even though its operational output is a sample size. Conversely, a rule that updates sample size using a pooled variance can sometimes be run with a blinded summary even though it changes the design.
| Adaptive decision | Minimum interim information typically needed | Role that may need it |
|---|---|---|
| Update sample size using a pooled nuisance estimate | A blinded pooled estimate such as a variance, with relevant design inputs | A statistician responsible for the update; an oversight role may review the procedure |
| Select an arm using comparative outcomes | Arm-specific outcomes and denominators, plus treatment identity for the selection rule | An independent statistician or designated oversight committee |
| Make an interim treatment-effect decision | A treatment contrast, its uncertainty, and the treatment mapping needed to interpret direction | A designated unblinded statistician and/or oversight committee |
| Implement a decision without evaluating efficacy | The operational instruction only, where feasible | Operational sponsor or site staff, as relevant |
This table describes typical minimum needs, not an automatic assignment of access. The prespecified protocol and responsibilities determine who must act. An independent statistician can prepare an analysis for an oversight committee; the sponsor may need to implement an approved change without receiving the comparative results that prompted it. Site staff generally need operational instructions, not arm rankings or interim treatment effects.
Why Pooled Data Do Not Usually Identify Treatment Effects
For a continuous outcome, the overall variability in pooled data can reflect both variability within arms and differences between arm means. Thus, a pooled standard deviation is not generally a substitute for the arm-specific means or their contrast. With arm \(j\) having proportion \(w_j\) of the observations, mean \(\mu_j\), and variance \(\sigma_j^2\), the variance of the pooled mixture is
The first term is the weighted within-arm variance. The second is variation attributable to differences among arm means. A blinded pooled summary does not identify the direction of those differences or which treatment produced a particular mean. It can therefore be useful for some nuisance-parameter updates while being insufficient for selection or efficacy decisions.
Worked Example: A Blinded Variance-Based Sample Size Update
Suppose a hypothetical two-arm trial uses equal allocation and a continuous outcome. The design targets a difference of \(\delta=5\) outcome units, two-sided \(\alpha=0.05\), and 90% power. At an interim review, a blinded pooled estimate of the standard deviation is \(s=12\) units. For illustration, use the normal-approximation planning formula for the required sample size per arm:
Here \(z_{1-\alpha/2}=1.96\), \(z_{1-\beta}=1.282\), and \(\beta=0.10\). Substitution gives
Rounding up gives 122 participants per arm, or 244 in total, before accounting for any other design requirements such as attrition. The update uses the pooled standard deviation and specified planning inputs; it does not require the statistician performing this calculation to know which arm has the higher mean. That does not, by itself, establish that the pooled estimate or the updating rule preserves the intended type I error. The rule must be suitable for the trial’s design and evaluated under its assumptions.
Arm Selection Requires Arm-Specific Outcomes
When a rule selects the treatment with the most favorable interim outcome, the analysis must distinguish outcomes by arm. It also needs denominators or other exposure information relevant to the outcome. If the rule depends on which treatment is which, treatment identity must be available to the decision-maker or to the statistician applying the rule. A ranking alone may be enough to communicate a prespecified choice, but it is not enough to independently verify the calculation.
Selection is not the same as evidence that the selected treatment is effective. An arm can rank first among several candidates because of random variation. The selection rule, the timing of the analysis, and the consequences for later inference must be specified in advance and evaluated as part of the design; a favorable interim ranking does not by itself justify an unadjusted claim of efficacy.
Worked Example: Selecting Among Arms Using Event Rates
In a hypothetical interim analysis, a trial has one control arm and two active arms. The outcome is an adverse event, so a lower event rate is favorable. The observed counts are 40 of 100 participants in control, 28 of 100 in Arm A, and 35 of 100 in Arm B. Suppose the prespecified operational rule is to carry forward the active arm with the lowest observed event rate.
The observed rates are \(40/100=0.40\) for control, \(28/100=0.28\) for Arm A, and \(35/100=0.35\) for Arm B. The absolute observed reductions relative to control are \(0.40-0.28=0.12\) for Arm A and \(0.40-0.35=0.05\) for Arm B. Under the stated ranking rule, Arm A is selected.
The rule needs the event count and denominator for each arm, and it needs the arm labels to identify which active treatment is Arm A. It does not follow that Arm A has been shown to reduce the event rate: this calculation is a ranking of interim estimates, not a confirmatory test. An independent statistician can apply the rule and provide the selected treatment or an authorized oversight committee can review the arm-specific results. Other roles need not receive the rates merely to implement the selection.
Treatment-Effect Decisions Need a Contrast and Uncertainty
A decision based directly on treatment effect needs more than pooled outcomes. At minimum, the analysis requires a defined contrast, the information needed to estimate its uncertainty, and treatment mapping sufficient to interpret its direction. For example, an efficacy rule might compare a standardized statistic with a prespecified boundary. The boundary and the timing of the analysis must be part of the design; an ordinary fixed-sample threshold should not be substituted automatically for a sequential decision rule.
Worked Example: Applying a Prespecified Interim Boundary
Consider a hypothetical trial where a larger outcome is better. At an interim look, the estimated treatment-minus-control difference is 4.2 points, with \(SE=2.0\) points. Suppose the prespecified interim rule requires a standardized statistic of at least 3.0 for an efficacy stop. The statistic is
Because \(2.10<3.0\), the efficacy boundary is not crossed, so this rule does not call for stopping for efficacy at this look. The estimated difference is in the favorable direction, but it does not meet the stated stopping criterion. The analysis needs the contrast and its standard error; the person interpreting benefit also needs to know which treatment is represented by the positive direction. A designated unblinded statistician can calculate the statistic, and an authorized oversight committee may make the decision. The sponsor or sites can receive an implementation instruction without necessarily receiving the estimate.
Assigning Access and Detecting Mismatches
A practical assignment follows the decision’s information needs. A statistician preparing a blinded variance-based update may need pooled summaries but not arm labels. An independent statistician applying an arm-selection rule needs arm-specific results and the treatment key. A monitoring committee may need unblinded comparative results to evaluate safety or efficacy, while operational staff may need only an approved action. The proper arrangement depends on the trial’s responsibilities; no role should be given interim information simply because it is convenient.
Check a proposed arrangement for mismatches. It may be under-informative if the assigned analyst cannot calculate the specified statistic or distinguish treatments needed for the decision. It may be over-informative if a role receives arm-specific results despite needing only an operational instruction. Also check that the information is delivered at the planned time, that treatment codes are interpreted consistently, and that the person applying the rule has the required denominators and analysis definitions.
If a sample size rule uses blinded pooled data, examine whether treatment differences could materially affect the pooled nuisance estimate. If a selection rule uses arm rankings, verify that the ranking is exactly the prespecified one and that the consequences for later inference have been addressed. If a treatment-effect rule uses a boundary, confirm that the estimate, standard error, direction, and boundary all correspond to the planned analysis. Unplanned exploration of interim results can lead to decisions that are difficult to justify statistically and may disclose more than the trial needs.
Limits of the Matching Principle
Minimum access is a design principle, not a guarantee of successful masking or error control. A pooled summary can indirectly reveal information in some settings, and even a nominally blinded update can affect operating characteristics. A rule based on arm-specific estimates can create selection effects, while a treatment-effect decision can require careful control of repeated looks. The appropriate statistical method and access arrangements must therefore be specified together.
There are trade-offs. Restricting access can protect trial conduct, but an oversight role must receive enough information to carry out its safety or efficacy responsibility. Giving a statistician a treatment key may make the calculation feasible, but requires appropriate independence and controls on onward disclosure. The goal is not the smallest possible data set regardless of the task; it is the smallest sufficient information set for the assigned task, delivered to the roles responsible for it.
Check Your Understanding
Answer each question using the distinction between pooled, arm-specific, and treatment-effect information.
- A sample size update uses only a blinded pooled standard deviation and fixed target effect. Does the calculation inherently require treatment labels?
- A team wants to carry forward the arm with the lowest interim event rate. What information is needed beyond a pooled event rate?
- An interim rule compares an estimated contrast with a standardized boundary. Name two essential pieces of information for applying and interpreting it.
- Why does selecting the arm with the best interim rate not establish that it is effective?
- A sponsor must implement an approved change but is not responsible for evaluating interim efficacy. What is the minimum information the sponsor may need?
Solutions. (1) No. The stated calculation can use a blinded pooled standard deviation and the fixed planning inputs, although the estimator and rule still require suitability checks. (2) Arm-specific event counts and denominators are needed, together with treatment labels sufficient to identify the selected arm. (3) The estimated treatment contrast and its standard error are needed to calculate the statistic; treatment mapping is needed to interpret its direction, and the prespecified boundary is needed to make the decision. (4) The observed ranking can arise from sampling variability, especially when several arms are compared; ranking is not a confirmatory efficacy conclusion. (5) The sponsor may need only the authorized operational instruction, provided the sponsor can implement it correctly without seeing comparative interim results.