Sequential Beta-Binomial Updating At Interim Looks

Why Update at Interim Looks?

In a clinical trial, outcomes may become available in batches rather than all at once. Investigators may want a current summary of the response probability after each batch has been observed. For a binary endpoint, Bayesian updating makes this possible: the posterior after one look becomes the prior information carried forward to the next look. The procedure is sequential in its bookkeeping, even though—under the model below—the final posterior is the same as if all observed outcomes had been analyzed together.

This equality is useful for checking computations and communicating how evidence accumulates. It does not mean that interim data can be reused as if they were new observations, nor that the model is automatically appropriate for every trial with repeated looks.

Model and Assumptions

Let \(p\) be the probability that a participant has the specified binary response by the specified assessment time. For each observed participant, let the response be 1 for success and 0 for failure. Assume outcomes are conditionally independent given \(p\), and that every participant in this analysis has the same response probability \(p\). Begin with a proper beta prior with positive parameters \(a\) and \(b\).

For sequential updating, divide the observed participants into \(K\) non-overlapping batches. At look \(k\), let \(m_k\) be the number of newly observed outcomes and \(y_k\) the number of successes in that batch. Each participant must appear in exactly one batch, and the recorded outcome must follow the same definition and assessment window throughout. The simple model also assumes that response probabilities do not change across batches after conditioning on \(p\). For example, a major change in eligibility, treatment, or response ascertainment could make one common \(p\) inappropriate.

The Recursive Update

Write the beta parameters available just before look \(k\) as \(a_{k-1}\) and \(b_{k-1}\). The initial values are \(a_0=a\) and \(b_0=b\). The batch contributes \(y_k\) successes and \(m_k-y_k\) failures, so the recursive update is

Thus after each interim look, the success parameter increases by the number of new successes, and the failure parameter increases by the number of new failures. These updated parameters summarize the prior information together with all outcomes observed so far; they are not a replacement prior to be combined again with those same outcomes.

Why the Order of Batches Does Not Matter

The path-independence result follows directly from the likelihood. For a batch with \(m_k\) participants and \(y_k\) successes, the binomial likelihood as a function of \(p\) is proportional to \(p^{y_k}(1-p)^{m_k-y_k}\). The binomial coefficient does not depend on \(p\), so it does not affect the posterior density. Multiplying the likelihood contributions from all batches gives the justification below.

To verify the result, multiply the prior density by the likelihood for each batch. Let \(Y=\sum_{k=1}^{K}y_k\) and \(F=\sum_{k=1}^{K}(m_k-y_k)\). Then the product is proportional to

This is the kernel of a \(\operatorname{Beta}(a+Y,b+F)\) density. The same result follows from the recursion because adding the batch success counts in any order gives \(a+\sum_k y_k\), and adding the failure counts gives \(b+\sum_k(m_k-y_k)\). This derivation requires that each batch contribute only its new outcomes. Counting earlier participants again would multiply in their likelihoods more than once and would produce incorrect parameters.

Predicting a Later Batch

An interim posterior can also describe what outcomes are plausible in a future batch before those outcomes are observed. If the current posterior is \(\operatorname{Beta}(A,B)\), the probability of exactly \(z\) successes in a new batch of \(r\) participants is obtained by averaging the binomial probability over uncertainty about \(p\):

where \(B(\cdot,\cdot)\) is the beta function. This is a beta-binomial predictive probability. It is not generally the same as substituting the posterior mean of \(p\) into a binomial model, because it averages over the full posterior distribution.

Interim Checks, Alternatives, and Limitations

At each look, check that the batch contains only new participants, that successes plus failures equal the batch size, and that the posterior parameters equal the original prior parameters plus the cumulative counts. The posterior parameters must not decrease as valid new observations are added. A mismatch often signals a duplicated batch, an incorrect denominator, or a success/failure coding error.

Also assess whether the assumptions make sense across enrollment time. Different response rates by site, calendar period, dose, or participant subgroup can signal that a single common response probability is inadequate. Dependence between participants or outcome-dependent missingness can likewise undermine the simple likelihood. Incomplete follow-up requires particular care: a participant whose response is not yet known should not automatically be counted as a failure. The analysis may need a model that accounts for the way observation or follow-up occurs.

The computational alternative is to pool all observed outcomes and update once. Under the stated model, pooling and recursive updating are exactly equivalent; sequential bookkeeping is convenient when interim summaries are needed, while pooled updating is a useful independent check. A different model may be needed if the response probability varies over time or across groups. Frequentist sequential procedures address a different issue: repeatedly examining data can affect frequentist error rates, whereas Bayesian posterior updating is a probability calculation conditional on the model and observed data. The equality of the two Bayesian calculations does not by itself establish frequentist operating characteristics or justify a trial decision rule.

The central limitation is therefore not the order of arithmetic but the assumptions behind the likelihood. If batches reflect different response mechanisms, or if participants enter the analysis selectively based on unobserved outcomes, totals alone may not contain all relevant information. A correct sequential update is only as credible as the model and data collection process it represents.