What Masking Must Conceal in a Multi-Arm Trial
In a two-arm trial, a masked code may leave two possible identities for each treatment. In a multi-arm trial, coded data can reveal more than identity: even if nobody knows which real treatment is “Arm B,” an interim summary may make it obvious that Arm B has the most favorable observed outcome. That ranking can affect expectations and behavior, and outside information may then connect the code to a real treatment.
The key distinction is between treatment identity (which intervention a code represents) and comparative performance (how the coded arms appear to be doing). Hiding the identity key does not necessarily hide the comparative performance. Conversely, a report can conceal arm-specific results while authorized trial personnel retain the identity key for a legitimate operational reason.
A trial with three or more arms therefore needs to ask two separate questions about each interim output: could it identify what intervention a code represents, and could it identify which coded arm appears favored? The second question matters even when the first answer is “no.” A person who learns that Arm C is performing best may change recruitment, co-interventions, assessments, or expectations without knowing whether Arm C is the new treatment, an active control, or placebo.
What a Summary Reveals
Suppose \(K\) randomized arms have endpoint means \(\bar{x}_1,\ldots,\bar{x}_K\), with \(n_j\) observed participants in arm \(j\). The pooled mean is a weighted average of the arm means:
A pooled summary does not state any one arm’s result. But if the arm sample sizes and all but one arm mean are known, the remaining mean is determined by the pooled total. Reporting a pooled estimate is thus not necessarily equivalent to making arm-specific information unrecoverable. The same concern applies to event counts, rates, and other summaries when enough related quantities are available.
The numerical endpoint also matters. A larger mean may be better for one outcome and worse for another. Before calling an arm favored, specify the endpoint, its direction, and the interim time point. If a favorable ranking is based on noisy interim data, it is not proof that the arm is truly superior; it can still be a potent signal to recipients.
Worked Example: A Coded Arm Becomes the Apparent Leader
In a hypothetical randomized trial, the interim mean improvement in a symptom score is reported for three coded arms. Larger improvements are better. Each arm has 40 participants. The results are 5.8 points for Arm A, 8.1 for Arm B, and 6.4 for Arm C. The treatment-code key is held separately.
The pooled mean is:
The pooled result alone gives no arm ranking. The arm-specific results do: \(8.1>6.4>5.8\), so Arm B is the apparent leader for mean improvement at this interim look. The identity key remains hidden, so this report does not establish which real treatment Arm B represents. Nevertheless, anyone receiving it can identify the favored code, which may be enough to alter behavior or prompt guesses about its identity.
The arithmetic checks out because all arms have the same sample size and the pooled mean is their simple average: \((5.8+8.1+6.4)/3=6.767\) points. This comparison is descriptive, not evidence that the true treatment effects differ: the example gives no variability estimates or inferential analysis.
When Pooled Reporting Is Not Enough
Pooling can reduce direct disclosure, but its protection depends on what else is known. With unequal allocation, arm counts are essential to interpret the weighted mean. If recipients know those counts and the means for all but one arm, they can recover the last arm’s mean. Even without exact recovery, repeatedly released pooled results alongside other trial information may narrow the possible arm-specific values. Suppression of a single table cell is therefore not a guarantee that the information is hidden.
Worked Example: Recovering a Hidden Arm Mean
In a hypothetical trial, three coded arms have \(n_A=30\), \(n_B=50\), and \(n_C=20\) participants with observed outcomes. Higher values are better. A report gives the pooled mean as 7.0 units, the mean for Arm A as 6.0, and the mean for Arm B as 8.0, but omits Arm C’s mean. Recover it using the weighted-average equation.
There are \(30+50+20=100\) observations. Let the omitted mean be \(\bar{x}_C\). Then:
Multiplying by 100 gives \(700=180+400+20\bar{x}_C\). Thus \(20\bar{x}_C=120\), and:
The supposedly omitted result is exactly recoverable from the other reported quantities. Since Arm B’s mean is 8.0 units, the highest of the three, it remains the apparent leader. This calculation assumes the pooled and arm-specific means use the same outcome definition, participants, and interim cutoff, and that the sample sizes refer to those same observations. If those conditions do not hold, this particular reconstruction need not be valid.
A short justification makes the risk precise. Write \(N=\sum_j n_j\). Multiplying the pooled-mean equation by \(N\), then isolating an unreported arm \(r\), gives:
Thus, if every quantity on the right is available, the hidden mean is not hidden. Rounding, missing observations, or different analysis populations may prevent exact recovery, but should not be relied on as a masking safeguard: they may still leave a narrow range or reveal which arm is most favorable.
Codes, Rankings, and Indirect Identification
Codes such as A, B, and C are arbitrary only if their mapping is genuinely protected and no other clue links them to interventions. In practice, a code can become recognizable through a distinctive dose schedule, formulation, known adverse-effect pattern, enrollment pattern, or operational handling. Such clues can be combined with an arm ranking. For example, someone who already suspects that one code is the high-dose intervention may infer its identity when that code is repeatedly described as the leading arm.
This does not mean every arm-specific statistic unmasks the treatment. The risk depends on recipients, prior knowledge, the level of detail, and whether interim information is repeated or combined with other data. A masked statistician who can see coded arm summaries but not the key may be able to compare coded arms while remaining unable to identify their interventions. Yet the comparative information itself may be sensitive, especially if people who influence trial conduct receive it.
Worked Example: Separating Code Identity From Performance
Consider a hypothetical three-arm trial of treatments X, Y, and Z. The independent data team holds a sealed mapping to codes A, B, and C. A masked operational group receives only the following interim event rates (lower is better): A, 12%; B, 7%; C, 10%. No treatment names or individual assignments are included.
The coded ranking is \(7\%<10\%<12\%\), so Arm B is the apparent leader for this endpoint and cutoff. The report does not establish whether B is X, Y, or Z: all three mappings remain possible from the information stated. If, however, the operational group knows from distinctive dosing that Arm B is likely treatment X, the ranking can increase the chance that recipients infer X is favored. A code key protects identity only to the extent that the mapping cannot be reconstructed or guessed from other information.
The rates are calculated as event counts divided by the respective numbers observed. For example, if each arm has 100 participants and the counts are 12, 7, and 10, the rates are \(12/100=0.12\), \(7/100=0.07\), and \(10/100=0.10\), or 12%, 7%, and 10%. This example supports a descriptive ranking only; without uncertainty estimates and a prespecified inferential analysis, it does not show that the underlying event probabilities differ.
Design Choices, Diagnostics, and Limits
There is no single reporting rule that preserves every kind of masking while meeting every operational need. The main alternatives differ in what recipients can learn:
| Reporting approach | Potential benefit | Masking concern |
|---|---|---|
| Pooled summaries only | Withholds direct arm-by-arm comparisons | Can be algebraically combined with counts or other summaries to recover arm information |
| Coded arm-specific summaries | Allows comparison of coded arms without necessarily revealing treatment names | Reveals rankings and may support indirect identification |
| Arm-specific summaries with the treatment key | Directly supports interpretation by those who need treatment identities | Removes identity masking for recipients and can expose comparative performance |
A practical review asks what a recipient could infer from the complete set of available outputs, not just from one table. Check whether arm labels are stable across reports, whether denominators and pooled totals are supplied, whether small cells or unusual response patterns are revealing, and whether known intervention features provide clues. Also check that the outcome’s favorable direction is stated: an arm ranking can be misread if lower values are better but the report’s wording suggests otherwise.
The protection offered by coded-arm reporting has limits. It can conceal the treatment mapping while intentionally exposing relative results; it can also fail to conceal either if the code is recognizable or summaries are reconstructable. Conversely, hiding all arm-specific information may frustrate a task that genuinely requires comparison. The relevant choice is not simply “masked” versus “unmasked,” but which information is available to which recipients and what they can infer from it.
A common mistake is to reason, “The arms are called A, B, and C, so the results are masked.” That confuses a label with protection. Correct reasoning asks separately whether the real treatment mapping is hidden and whether any recipient can identify the apparent leader. A second mistake is to assume pooled reporting always prevents arm-level inference. The weighted-mean equation shows why that can fail when related counts and summaries are also known.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A report gives coded-arm means of 4.2, 5.1, and 4.7 points, where higher is better, but the treatment key is withheld. What can recipients learn, and what can they not necessarily learn?
- A two-arm hypothetical trial has 40 participants per arm, pooled mean 10 units, and Arm A mean 8 units. Recover Arm B’s mean, assuming both arm means and the pooled mean use the same observations.
- Why can a pooled summary fail to conceal an omitted arm mean?
- Name two ways a coded arm’s real treatment identity could become guessable even if the key is withheld.
Solutions. 1. They can rank the coded arms: the arm with mean 5.1 points is the apparent leader. They cannot determine its real treatment identity from these numbers alone if the mapping is otherwise protected. 2. The total sample size is 80, so \(10=(40(8)+40\bar{x}_B)/80\). Multiplying by 80 gives \(800=320+40\bar{x}_B\); hence \(\bar{x}_B=480/40=12\) units. 3. If the pooled total, arm sample sizes, and all other arm summaries are known, rearranging the weighted-average equation determines the omitted mean. 4. Examples include a distinctive dose schedule or formulation, and a known adverse-effect or enrollment pattern associated with the intervention. These clues can support an inference but do not guarantee one.