Pooled Data Views

What a Pooled Data View Contains

In an adaptive trial, a decision may be made using interim data without providing decision-makers with the outcomes for each randomized arm separately. A pooled data view combines observations across arms and summarizes them without using the treatment-arm identities to report arm-specific outcomes. It can describe the trial data overall; by itself, it usually cannot show which arm performed better, or by how much.

Pooling does not mean that every useful feature of the data disappears. Depending on what has been collected and made available, an overall view can include the number of participants enrolled, the overall distribution of an endpoint, its mean and variance, an overall event count or proportion, and summaries of baseline characteristics or follow-up. Each quantity must still be defined carefully: for example, an event proportion needs a specified numerator, denominator, and observation period.

The arithmetic of an overall summary does not require treatment identity. Its interpretation does require context. Outcomes must refer to the same defined endpoint and relevant time point; otherwise an aggregate number can combine quantities that do not represent one coherent summary. The earlier material on estimands emphasizes why population, outcome, and follow-up definitions matter. Pooling cannot repair incompatible definitions or missing outcomes.

Suppose arm 1 contributes \(n_1\) observed outcomes with mean \(\bar{x}_1\), and arm 0 contributes \(n_0\) with mean \(\bar{x}_0\). The pooled mean is the participant-weighted mean:

It is not generally the unweighted average of the two arm means. That shortcut is correct only when the arm sample sizes are equal. Even when the allocation counts are known, the pooled mean provides only one weighted average; it does not, in general, determine both arm means or their difference.

Pooled Rates and Their Denominators

For a binary endpoint, let \(e_1\) and \(e_0\) be the event counts in the two arms, with corresponding denominators \(n_1\) and \(n_0\). If every participant in those denominators has a defined binary outcome for the specified period, the pooled event proportion is

This is the weighted average of the arm proportions, not generally their simple average. If observation periods differ, or outcomes are missing for some participants, the denominator and handling of those observations need to be made explicit. An overall proportion calculated from only participants with observed outcomes describes that observed set; it does not automatically describe all randomized participants.

Why Pooled Variance Can Include an Arm Difference

An overall variance is calculable from pooled numerical outcomes, but it need not estimate the variability within either arm. It can reflect both variation among participants within arms and separation between the arm means. The following identity makes that distinction explicit.

This population identity assumes that \(Y\) has a defined distribution within each arm and that the pooled distribution is the mixture of those two arm distributions with arm 1 weight \(q\). Its first two terms are the weighted within-arm variances. Its final term is the variance of the arm-specific means around the overall mean. That term is nonnegative and is positive when the arm means differ and both arms have positive weight.

To verify the identity, write the overall mean as \(\mu=q\mu_1+(1-q)\mu_0\). Decompose variation around that mean into variation around each arm mean and variation of the arm means around \(\mu\):

The first term is \(q\sigma_1^2+(1-q)\sigma_0^2\). For the second, let \(\Delta=\mu_1-\mu_0\). Then \(\mu_1-\mu=(1-q)\Delta\) and \(\mu_0-\mu=-q\Delta\), so

Combining the terms gives the result. In particular, if both arms have the same variance \(\sigma^2\), then \(\operatorname{Var}(Y)=\sigma^2+q(1-q)(\mu_1-\mu_0)^2\). Thus an overall variance can be larger than the common within-arm variance precisely because the arm means differ. Observing the overall variance does not, by itself, tell us how much came from within-arm variation and how much from separation between arms.

Assumptions, Diagnostics, and Practical Limits

The arithmetic formulas require that the included observations and their definitions be clear. For a pooled mean, each included participant must contribute an outcome on a common scale, and the denominator must match the observations used in the sum. For a pooled event proportion, the event definition, time window, and denominator must be specified. These are descriptive requirements; they do not establish that a pooled summary estimates a particular population quantity or supports a particular adaptation.

Randomization gives a basis for comparing assigned arms when assignment-linked outcomes are available and the comparison is appropriately conducted. It does not make an overall pooled outcome a treatment-effect estimate. Nor does masking treatment identities guarantee that pooled summaries contain no information about treatment effects. As the variance identity shows, an overall variance can depend on the arm difference; a pooled mean or rate can also change with the arm mixture and allocation proportions.

Before interpreting a pooled summary, check the following:

• Weights and allocation. Confirm the numbers contributed by each arm. A pooled mean or rate weights participants, so changes in allocation can change the overall summary even when arm-specific outcomes do not change.
• Outcome availability. Check the number with observed outcomes and whether missingness or differing follow-up changes who contributes to the summary.
• Consistent definitions. Confirm that the endpoint, assessment time, and population are coherent across the included observations.
• Composition over time. In a trial with staged enrollment or an adaptation, the pooled data available at one interim may represent a different mix of participants or follow-up than a later pooled snapshot.
• What the summary can identify. Ask whether the reported quantity is an overall description or a contrast. A total count or overall mean does not become an arm comparison simply because randomized arms generated the observations.

A pooled view is useful for describing combined data without reporting separate arm outcomes. The main alternative for assessing relative arm performance is to retain assignment-linked outcomes and calculate an arm-specific comparison, with the relevant treatment identities available to appropriately authorized decision-makers. That alternative answers a different question and involves different access to interim information. A pooled summary may be adequate for a decision that needs an overall quantity, but it cannot substitute for a contrast when the decision depends on which arm performs better.

Common Mistake

A related mistake is to call an overall standard deviation a within-arm standard deviation. The variance decomposition shows why that inference can fail: overall variability may include a between-arm component. Conversely, a pooled view is not useless merely because it cannot identify the contrast. It can still give valid descriptive totals and summaries when the endpoint and included observations are well defined.