Why Separate Arms Change What a Summary Reveals
A pooled summary can conceal differences between randomized arms: it combines their observations without showing which outcomes came from which arm. Reporting separate arm counts or outcome summaries adds a new kind of information. Even if the arms are labeled only “A” and “B,” a reviewer can see how their observed outcomes differ. Thus, hiding treatment names and hiding comparative outcome information are not the same thing.
This distinction matters when trial information is reviewed before the trial is complete. Separate summaries may be useful for understanding the data, but they can reveal whether one masked arm appears to have better outcomes. That observation does not, by itself, identify which arm received which treatment, establish a treatment effect, or justify an adaptation. It does change what the reviewer knows compared with a pooled view.
The interpretation depends on what information is actually shown. Separate enrollment counts alone reveal the allocation totals, not outcome differences. Separate outcome counts or summaries reveal between-arm outcome patterns. If the labels are masked, the comparison is still visible, but its treatment identity is not supplied by those labels alone. Prior knowledge or other information could nevertheless make the identity inferable.
What Arm-Specific Summaries Show
Consider a two-arm randomized trial with arms labeled A and B. Suppose the endpoint and its assessment time have been specified, and the summaries use the same population definition and outcome rules in both arms. A count \(n_A\) is the number of participants assigned to A in the stated analysis set; \(n_B\) is defined correspondingly. For a quantitative endpoint, \(\bar{x}_A\) and \(\bar{x}_B\) are the arm-specific sample means, in the endpoint’s units. For a binary endpoint, \(\hat{p}_A\) and \(\hat{p}_B\) are the observed proportions meeting the stated response definition in each arm.
These summaries are descriptive. For example, \(\hat{p}_A-\hat{p}_B\) describes the observed difference in response proportions, in proportion units (or percentage points after multiplying by 100). A positive value means the observed response proportion is higher in A under the chosen subtraction order. Without knowing the treatment identities, it does not mean that a named treatment performed better. Nor does a descriptive difference alone establish its precision, statistical significance, or causal interpretation.
A usable comparison requires more than two columns of numbers. Check that the summaries refer to the same endpoint, assessment time, participant definition, and handling of missing outcomes. Always inspect both numerators and denominators for a binary endpoint. A higher response proportion based on few observed outcomes may be less informative than the same proportion based on many participants; a proportion with a different denominator or outcome rule may not be comparable at all.
Worked Example: A Masked Difference in Response Proportions
Hypothetical data. At a prespecified interim assessment, a two-arm trial has response information for 40 participants in A and 40 in B. The response definition and assessment time are the same in both arms. There are 24 responders in A and 18 in B. The labels A and B do not reveal the treatment names.
The observed response proportions are
Using A minus B, the observed difference is
In this hypothetical snapshot, the observed response proportion is 0.150 higher in A, or 15.0 percentage points higher. Because the two arm labels are masked, the reviewer cannot name the treatment associated with the higher proportion from these summaries alone. However, the view does reveal the direction and size of the observed contrast between the labeled arms. It is not a 15-point estimate for a named treatment until treatment identity is known, and the calculation alone does not quantify uncertainty.
Pooled Information Does Not Determine the Arm Contrast
Pooling removes the arm-by-arm allocation of outcomes. For a quantitative endpoint, if all \(n_A+n_B\) observations contribute to the pooled mean, the pooled mean is a weighted average of the arm means:
The weights are the numbers of observations in the arms. Unless the arm means are also available, a pooled mean generally cannot recover their difference. The following derivation makes the information loss explicit.
The sum of all observed values is the sum of the arm totals: \(n_A\bar{x}_A+n_B\bar{x}_B\).
Dividing that total by \(n_A+n_B\) gives the pooled mean shown above.
One pooled mean gives one weighted-average constraint on two arm means. Many pairs of arm means can satisfy it, so the pooled mean does not determine \(\bar{x}_A-\bar{x}_B\).
Worked Example: Same Pooled Mean, Different Arm Patterns
Hypothetical data. Suppose 30 participants contribute an endpoint measured in points: 15 in A and 15 in B. In one possible set of summaries, \(\bar{x}_A=12\) points and \(\bar{x}_B=8\) points. The pooled mean is
A different possible set of arm means, 9 points in A and 11 points in B, produces the same pooled mean:
In the first case, A exceeds B by 4 points; in the second, A is 2 points below B. The pooled mean of 10 points does not reveal which pattern occurred. Showing the arm-specific means would reveal the observed contrast, even if the labels remained masked. These calculations assume the same endpoint and that all 30 observations are included in both the arm and pooled means.
Counts, Denominators, and Incomplete Outcomes
Arm-specific counts can be informative even when outcome summaries are unavailable. For a binary endpoint, it is useful to show the number meeting the response definition alongside the denominator used to calculate the proportion. In a snapshot with missing outcome data, distinguish the number assigned to an arm from the number with an observed outcome. A proportion calculated among participants with observed outcomes describes that observed subset; it is not automatically the proportion among all assigned participants.
Differences in denominators or follow-up can distort a simple comparison. Suppose one arm has had more time for participants to reach the assessment, or more outcomes are missing in one arm. Then comparing raw counts of responders would mix outcome patterns with differences in the number assessed. Even proportions require caution if the observed participants differ systematically from those whose outcomes are missing. State the analysis set and assessment window, and do not present a count as though it were a rate.
Worked Example: Why Both Numerators and Denominators Matter
Hypothetical data. At a common assessment time, response is observed for 36 of 40 participants assigned to A and 30 of 40 assigned to B. There are 21 observed responders in A and 15 in B. The other participants have no observed response at this assessment.
Among participants with an observed outcome, the proportions are
The observed-case difference is \(0.5833-0.5000=0.0833\), or 8.33 percentage points in favor of A by the stated subtraction order. But the denominators show that 4 of 40 assigned participants in A and 10 of 40 in B lack an observed response. The observed-case proportions do not establish what the missing outcomes would have been, and they are not the same as treating all assigned participants as nonresponders. A clear arm-specific view therefore reports the assigned counts, observed-outcome counts, responder counts, and the exact denominator for each proportion.
Comparing Views and Diagnosing Problems
| View | What it shows | What it does not show by itself |
|---|---|---|
| Pooled summary | Overall counts or outcome summaries across arms | The separate arm outcomes or their contrast |
| Masked arm-specific summary | Separate outcomes for A and B and their observed contrast | Which treatment corresponds to A or B, if identity is genuinely unavailable |
| Treatment-identified arm summary | Separate outcomes linked to named treatments | Whether a descriptive difference is precise, statistically significant, or sufficient to guide a decision |
The choice is not simply between “blinded” and “unblinded.” A view may conceal treatment names while still exposing comparative outcomes. Conversely, a pooled view can preserve more separation between the reviewer and the arm contrast, though it can obscure arm-specific problems. Which view is appropriate depends on the purpose and access rules of the review; the summaries themselves do not decide those rules.
When reviewing an arm-specific display, check for the following failure modes:
- Assuming masked means noncomparative. Separate outcome summaries permit an A-versus-B comparison even when the labels do not name treatments.
- Ignoring the denominator. Raw responder counts are not comparable if the numbers assessed differ; show the denominator and analysis set.
- Mixing definitions. Arm summaries using different assessment times, endpoint rules, or participant sets cannot be interpreted as a like-for-like contrast without qualification.
- Overreading a descriptive difference. A visible contrast is not, by itself, a significance test, a precise treatment-effect estimate, or proof of causation.
- Assuming identity can never be inferred. A label may be formally masked but recognizable from outside knowledge, unusual outcome patterns, or information already available to a reviewer. Masking labels alone does not guarantee identity remains unknown.
Arm-specific summaries also have limits as snapshots: they describe the data included at that time and may change as more outcomes accrue or are verified. A reviewer should not treat an interim contrast as a final result. The earlier material on blinding levels and pooled data views provides the context for distinguishing who can see which information; here, the key point is that separating outcomes by arm itself reveals a comparison.
Common Mistake
Correction: The treatment names may be hidden, but the arm-specific outcomes still show whether A or B has the higher observed summary and by how much. What remains concealed is the link between the labels and treatment identities, provided that link cannot be inferred from other information. Identity masking and concealment of the arm contrast are distinct properties.
For any arm-specific display, describe its scope precisely: name the endpoint, time point, analysis set, outcome status, arm labels, and summary being shown. Then state the contrast using those labels and units. This avoids implying more than the display supports while preserving the information it does provide.
Check Your Understanding
Answer each question using the distinctions between pooled summaries, arm-specific outcomes, and treatment identity.
- A report gives \(\bar{x}_A=7\) days and \(\bar{x}_B=5\) days, with treatment names hidden. What can a reader conclude from these summaries, and what can the reader not conclude?
- For a binary outcome, A has 12 responders among 20 assessed and B has 15 among 30 assessed. Calculate each observed response proportion and their difference, A minus B.
- Explain why a pooled mean of 10 points cannot, in general, determine the difference between two arm means.
- A table gives 8 responders in A and 11 in B, but omits the numbers assessed. What important interpretation is blocked?
Solutions. (1) A’s observed mean is 2 days higher than B’s, in the stated endpoint units; the summaries alone do not identify which named treatment is A or establish statistical significance or a causal effect. (2) \(\hat{p}_A=12/20=0.600\), and \(\hat{p}_B=15/30=0.500\). Thus, A minus B is \(0.600-0.500=0.100\), or 10.0 percentage points among those assessed. (3) The pooled mean is one weighted-average constraint on two arm means, so multiple pairs of means, with different contrasts, can produce the same pooled value. (4) The response proportions and their comparison cannot be calculated; the raw counts may reflect different numbers assessed rather than different response rates.