Beta-Binomial Robust Mixture Example

Why agreement and conflict matter

When a trial has few participants, historical response data may be useful, but they may not describe the current trial well. A robust mixture represents that uncertainty directly: one prior component permits borrowing from historical information, while another protective component limits borrowing when current outcomes are poorly aligned with that information. The current data update both components, and also update the probability assigned to each component.

This tutorial develops that calculation for a binary endpoint. The examples use hypothetical counts. The aim is to see precisely how agreement and conflict affect the posterior distribution of the current response probability, not to treat a historical response rate as automatically exchangeable with a current one.

Model and assumptions

The borrowing component may be specified by translating historical evidence into beta parameters; for example, a historical count can be combined with a prespecified beta baseline prior. The protective component is chosen to represent a defensible, less-borrowing distribution for \(p\). Neither component is guaranteed to be appropriate merely because it has been given a label. The endpoint definition, follow-up period, patient population, response ascertainment, and trial conduct must support the model and the historical information used to construct the borrowing component.

The mixture weight \(w\) is the prior probability assigned to the borrowing component, not the fraction of current participants who are historical participants. For the conjugate calculations below, the component parameters and \(w\) are specified before observing the current count. The binomial likelihood also assumes a fixed number \(n\) of evaluable participants and a common \(p\). Missing outcomes, varying follow-up, clustering, or important patient-level differences may require another likelihood.

Component updates and posterior mixture weights

For observed \(X=x\), the likelihood is proportional to \(p^x(1-p)^{n-x}\). Multiplication by a beta density gives a beta posterior within each component. The posterior probability of the borrowing component depends on how well that component predicted the observed count relative to the protective component.

Here \(Z\) denotes the component from which the prior for \(p\) was drawn. Define the beta function by \(B(a,b)=\int_0^1p^{a-1}(1-p)^{b-1}\,dp\). The probability of the observed count under a component is its beta-binomial marginal likelihood. The binomial coefficient is common to both components and cancels in the posterior weight:

Thus the posterior distribution is itself a mixture, with borrowing weight \(w_{\mathrm{post}}\). Its posterior mean is the weighted average of the two component posterior means. This average is also the model-based posterior predictive probability of response for one additional participant with the same response probability \(p\).

The calculation has a useful interpretation. The data affect the posterior in two ways: they change the response probability within each component, and they change the relative support for the components. Treating the mixture weight as fixed after observing the data misses the second update.

Justifying the reweighting

The weight formula follows directly from conditioning on the component indicator. Before observing the count, \(P(Z=B)=w\) and \(P(Z=P)=1-w\). The probability of observing \(x\) responses, conditional on either component, is the likelihood averaged over that component’s beta prior. Applying Bayes’ rule gives:

For either component \(j\), the averaging step is \(\int_0^1\binom{n}{x}p^x(1-p)^{n-x} p^{a_j-1}(1-p)^{b_j-1}/B(a_j,b_j)\,dp\). Combining powers of \(p\) and \(1-p\) gives \(\binom{n}{x}B(a_j+x,b_j+n-x)/B(a_j,b_j)\), the stated beta-binomial marginal likelihood. This shows why the component weight responds to predictive agreement, rather than simply to whether the observed proportion lies above or below a prior mean.

Diagnostics, alternatives, and limitations

Before interpreting a posterior, check that the response count corresponds to the intended endpoint and analysis population, and that the assumed \(n\) and \(x\) are correct. Confirm that each component has positive shape parameters and that \(0<w<1\). Recalculate both component marginal likelihoods and verify that the posterior weights are nonnegative and sum to one. The component means and the mixture mean should lie between 0 and 1, and the mixture mean should lie between the two component means.

A very small posterior borrowing weight can indicate that the observed count is poorly predicted by the borrowing component, but it is not by itself proof that the historical and current populations differ. Conversely, a large weight does not establish exchangeability: two components can both fit poorly, or the protective component may not cover plausible current response rates. Review the endpoint, follow-up, eligibility criteria, and clinical context alongside the calculation. With very small samples, one or two outcomes can shift the weights substantially.

A single beta prior is simpler, but it cannot represent both the possibility of borrowing and a distinct protective alternative in this way. A pooled analysis that combines historical and current counts directly can amount to stronger borrowing and does not adapt its borrowing weight through this two-component update. The robust mixture makes the alternative component explicit, but it is not a substitute for checking the likelihood or justifying the prior components. In particular, a common-\(p\) binomial model can be inadequate with overdispersion, substantial patient heterogeneity, or time-varying response probabilities. More elaborate regression models address different structures and should not be inferred from this simple count model.

A common mistake is to calculate the two beta posterior means and then average them using the original weight \(w\). That leaves the component probabilities unchanged after seeing the data. The correction is to calculate each component’s marginal likelihood and update its weight with Bayes’ rule before averaging. Another mistake is to call \(w_{\mathrm{post}}\) the probability that the historical data are “true.” It is the posterior probability of a specified prior component under the model, not a general measure of historical validity.