Why the Review Arrangement Matters
An interim review needs enough information to support its planned decision, but it need not expose every available detail. The central choice is how the review data represent treatment assignment: are outcomes pooled across arms, separated into groups with neutral codes, or shown with the actual treatment identities? These arrangements do not change the underlying trial data. They change what a reviewer can distinguish and therefore which decisions the reviewer can make.
The appropriate arrangement is the least revealing one that still supports the intended adaptation. Pooling may be adequate for a decision based on an overall event rate. It cannot show which arm has the higher rate. Treatment-coded data preserve arm comparisons while concealing which code represents which treatment. Treatment-identified data retain both the comparisons and their clinical meaning. More information can enable a more specific decision, but also exposes more interim treatment information to the reviewer.
These terms describe the information arrangement, not who receives the information, whether the data are patient-level or summarized, or how frequently the review occurs. For example, a statistician could receive a coded summary table or coded participant-level records. In either case, the key masking question is whether treatment-specific results can be distinguished and whether their actual identities are known. The planned role and access route must be considered separately.
What Information Each Arrangement Preserves
Suppose a two-arm randomized trial compares treatments X and Y on a binary outcome. A pooled review might show 24 events among 80 participants, but not the number of events in either arm. A coded review might show 8 of 40 in A and 16 of 40 in B, while keeping the A-to-X or A-to-Y mapping concealed. An identified review would show which actual treatment had 8 events and which had 16.
Pooling removes the ability to estimate an arm contrast from the reviewed results. Coding preserves the contrast but not its clinical label. Identification preserves both. A person with external knowledge, such as a distinctive treatment schedule or an obvious adverse-effect pattern, might infer an identity despite neutral labels; coding is not a guarantee against inference.
This is an information ordering, not a claim that every identified review is better. A decision rule that uses only pooled outcomes can still be applied when more detailed results exist, but the reviewer has then been exposed to information the decision did not require. Conversely, a pooled summary cannot be transformed back into the arm-specific results that were removed.
A Short Justification for the Information Ordering
Let \(D\) denote all interim data, and let \(g(D)\) be the pooled summary. Let \(c(D)\) be the arm-specific results labeled A and B, with the code-to-treatment key withheld. The treatment-identified presentation includes \(c(D)\) and the key. A pooled summary can be computed from arm-specific data, but the reverse is generally impossible: many distinct arm-level datasets can have the same pooled total.
For instance, 24 events among 80 participants could arise from 12 events in each arm, 8 and 16, or 4 and 20. Each configuration gives the same pooled count, so the pooled result alone cannot determine the arm contrast. Coded results retain the contrast; revealing the key then attaches treatment names to the codes. Thus, the presentations form an increasing sequence of distinguishable information: pooled, coded, identified. This justification assumes that the review materials contain only the stated information; ancillary details can weaken masking.
Worked Examples
Worked Example: A Binary-Outcome Review
In a hypothetical trial, 40 participants have been observed in each of two randomized arms. The interim data contain 8 events in one arm and 16 in the other. Which arrangement can support each of the following: an overall event-rate review, an arm comparison, and a decision about which named treatment has the lower observed event rate?
Pooled review. The total is \(8+16=24\) events among \(40+40=80\) participants. The pooled event rate is
This is enough to describe the overall observed event rate, but it does not say which arm had 8 events and which had 16. It cannot support an arm comparison from the pooled result alone.
Treatment-coded review. Suppose the table reports A: \(8/40=0.200\) and B: \(16/40=0.400\). The coded difference in observed event rates, B minus A, is \(0.400-0.200=0.200\), or 20 percentage points. Coding supports an arm-level comparison, but without the key it does not identify which named treatment had the lower observed rate.
Treatment-identified review. If the key shows that A is treatment X and B is treatment Y, the reviewer can state that X had the lower observed event rate in these interim data. If the key instead maps A to Y, the clinical interpretation reverses. The arithmetic is unchanged; the identity mapping determines which treatment the result describes.
The example checks the information requirement rather than establishing a treatment effect. The interim rates are descriptive, and a decision about efficacy would require the trial’s prespecified analysis and decision criteria.
Worked Example: Re-Estimating a Variability Parameter
Consider a hypothetical continuous outcome. At an interim review, each of two arms has \(n=30\) observations. The arm means are 10 and 12 units, and each arm’s sample variance is 4 square units. Which information is useful if the adaptation requires an estimate of within-arm variability?
A coded review can show the two arm-specific variances, each equal to \(4\) square units, without revealing which treatment is A or B. A pooled review instead combines the observations. Its sample variance includes both within-arm variation and the difference between the arm means.
The pooled mean is \((30(10)+30(12))/60=11\) units. The total sum of squares is the sum of the within-arm and between-arm contributions:
There are \(60-1=59\) degrees of freedom for the pooled sample variance, so
The pooled standard deviation is larger than the within-arm standard deviation of 2 units because the two observed means differ. If the adaptation specifically needs within-arm variability, the coded arm-specific summaries directly provide it. The pooled value is not automatically wrong: it estimates variability across the combined observations, a different quantity. The reviewer must match the summary to the parameter required by the planned procedure. This calculation assumes all 60 observations use the same outcome scale and that the stated sample variances were calculated within arms.
Worked Example: Selecting a Coded Arm Versus a Named Treatment
A hypothetical three-arm study compares treatments P, Q, and R on a binary outcome. A coded interim table gives A: \(12/40\), B: \(9/40\), and C: \(6/40\). The protocol calls for identifying the code with the highest observed response rate, but the reviewer does not receive the code key.
The observed rates are \(12/40=0.300\), \(9/40=0.225\), and \(6/40=0.150\). Therefore, code A has the highest observed rate. This coded information is sufficient to rank the arms and select code A under a rule stated in terms of codes. It is not sufficient to tell the reviewer whether A is P, Q, or R.
If an authorized custodian later maps A to Q, then the selected named treatment is Q. If the adaptation instead requires the reviewer to assess the clinical suitability of a specific named treatment—for example, because a treatment-specific safety profile determines whether it may continue—the reviewer needs the identity information or a separate authorized process that applies the identity-dependent rule. The architecture must meet the actual decision need; a coded ranking cannot answer a question that depends on treatment names.
Choosing an Arrangement
A practical choice starts with the adaptation, not with a preferred level of concealment. Specify the decision, the quantity needed to make it, and whether the decision depends on treatment identity. Then choose the least revealing arrangement that supplies those inputs. The earlier material on matching adaptations to information access emphasizes this decision-first logic; here it is applied specifically to the three masking architectures.
Write what will be changed, selected, or continued, and what interim result triggers that action. Avoid vague requirements such as “review the data.”
Decide whether the rule needs an overall rate, arm-specific summaries, a contrast between codes, or a comparison tied to named treatments.
If the same action can be taken using A/B labels, coding may suffice. If the action depends on which treatment is which, identity is decision-relevant.
Do not provide treatment-identifying results simply because they are available. Conversely, do not use pooled results for a rule that requires arm-specific evidence.
Confirm that labels, summaries, and any code key are handled consistently, and that the intended reviewer can apply the planned rule without needing undocumented information.
| Arrangement | Information preserved | Can support | Main limitation |
|---|---|---|---|
| Pooled data | Aggregate outcome information across arms | Decisions based only on a pooled quantity | Cannot recover arm-specific results or contrasts |
| Treatment-coded | Arm-specific results and contrasts, with neutral labels | Code-level comparison or selection when identity is unnecessary | Cannot by itself attach clinical names to codes; differences may still be inferred |
| Treatment-identified | Arm-specific results plus treatment identities | Decisions requiring named-treatment interpretation | Exposes the most treatment-specific interim information |
Diagnostics, Alternatives, and Limitations
A sound choice can be checked by testing whether the planned decision is actually computable from the proposed review materials. If the adaptation rule needs an arm contrast but the materials contain only a pooled total, the arrangement is insufficient. If a coded comparison is enough but the reviewer is also given the code key, the arrangement reveals more than the stated requirement. If treatment identities are essential but are unavailable to anyone authorized to apply the rule, the process is operationally incomplete.
Common warning signs include a rule that uses arm-specific language while specifying only pooled summaries; neutral labels that are easy to infer from treatment-specific patterns; and ambiguous or changing code mappings. Another warning sign is treating the masking label as a substitute for checking what the reviewer can derive from the data. A pooled table may include enough subgroup detail to reconstruct arm differences, while a coded table may effectively disclose identity through distinctive treatment features.
The main alternative to a pooled review is not always full treatment identification. A coded review often provides the needed comparison while withholding the treatment key. The alternative to coding is therefore chosen according to the decision: use pooling when arm distinctions are unnecessary, coding when arm distinctions matter but names do not, and identification when the named treatment is itself relevant. A separate authorized process can sometimes apply an identity-dependent rule without distributing the key broadly, but the responsibilities and information flow must be explicit.
Masking architecture alone does not guarantee trial integrity or preserve statistical error rates. It cannot prevent all inference, behavioral responses, or leakage, and it does not validate an adaptation rule. Prespecification, appropriate statistical operating characteristics, and controlled access remain necessary. The architecture is a way to align information exposure with the task, not a replacement for those safeguards.
Common Mistake
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A pooled interim report shows 30 events among 100 participants. Can it establish which of two equally sized arms had the higher event rate? Explain.
- A coded review reports response rates of 0.18 for A and 0.24 for B. What can the reviewer conclude, and what remains unknown?
- An adaptation rule depends only on the overall event rate. Which arrangement is the least revealing sufficient choice, assuming no other arm-level information is needed?
- Why might a treatment-coded arrangement fail to conceal treatment identity in practice?
Solutions. (1) No. The same pooled 30 events could be split in many ways between the arms, so the pooled report does not identify their separate rates. (2) The reviewer can conclude that B has the higher observed coded response rate, by \(0.24-0.18=0.06\), or 6 percentage points; without the key, the reviewer does not know which named treatment is B. (3) A pooled review is sufficient and least revealing among the three arrangements, provided the rule truly depends only on the overall rate. (4) The reviewer may infer the code mapping from distinctive treatment features, external information, or other disclosed results; neutral labels do not guarantee that identity is unknowable.