Bayesian Borrowing for Model-Based Label Updates

Why Borrow Evidence for a Label Update?

A postapproval data set may be too small to estimate a model parameter precisely, even when relevant preapproval evidence is available. Bayesian borrowing provides a formal way to let that earlier evidence inform an updated estimate. The key question is not simply whether borrowing increases precision; it is whether the earlier and later evidence are sufficiently compatible for that precision gain to be credible.

For a model-based label update, the quantity being borrowed should correspond to the parameter that supports the proposed statement. For example, it might be a treatment effect in a specific population or a model parameter that determines a predicted response. In the examples below, \(\theta\) is a continuous treatment benefit measured in symptom-score points, with larger values indicating greater benefit. All data and estimates are hypothetical.

The calculations below use a normal approximation for the estimate from each evidence source. This is useful when a model-based analysis supplies an estimate and standard error. It is an approximation with assumptions, not a claim that every model parameter or every data set is normally distributed.

Commensurate Borrowing

A commensurate prior expresses the idea that the postapproval parameter may be close to, but need not equal, the preapproval parameter. Let \(m_{\mathrm{pre}}\) and \(V_{\mathrm{pre}}\) denote the mean and variance of the preapproval posterior for the parameter. A simple conditional prior for the postapproval parameter is

Here \(\tau\) is the commensurability precision: a larger \(\tau\) allows less extra variation between the evidence sources, while a smaller \(\tau\) allows more. The term \(V_{\mathrm{pre}}\) carries forward uncertainty in the preapproval estimate. The additional variance \(1/\tau\) represents possible lack of commensurability. This form is a practical normal approximation to borrowing from the earlier posterior; it does not assert that the two true effects are identical.

Suppose the postapproval estimate \(y_{\mathrm{post}}\) has sampling model \(y_{\mathrm{post}}\mid\theta_{\mathrm{post}}\sim N(\theta_{\mathrm{post}},s_{\mathrm{post}}^2)\). If its sampling error is independent of the preapproval evidence conditional on the parameter, conjugate normal updating gives

The posterior mean is a precision-weighted average of the prior mean and the postapproval estimate. The prior precision is \(1/(V_{\mathrm{pre}}+1/\tau)\), not \(1/V_{\mathrm{pre}}\): allowing less commensurability reduces how strongly the earlier evidence pulls the update.

Robust Mixture Priors and Conflict

A fixed commensurate prior can still borrow substantially when the new estimate is far from the historical estimate. A robust mixture prior provides an alternative: it assigns some prior probability to a borrowing component and the rest to a weak, less informative component. For example,

The weight \(w\) is the prior probability of the borrowing component. After observing postapproval data, Bayes' rule updates the probability assigned to each component. If \(y\) has standard error \(s\), the posterior probability of the borrowing component is

Here \(\phi(y;m,v)\) is the normal density at \(y\) with mean \(m\) and variance \(v\). Each density in the expression is a predictive density: it measures how well the observed estimate accords with that prior component, accounting for sampling variance. The posterior distribution is itself a mixture of the two component-specific posteriors; its mean and uncertainty must reflect both components and the updated weights.

From a Posterior to a Label-Relevant Estimate

A posterior distribution can describe uncertainty about a model parameter relevant to a proposed label statement. Suppose the label question is whether the estimated benefit exceeds a prespecified clinically relevant threshold. The posterior probability for that event is calculated under the model and prior; it is not the probability that the model or label statement is true in an unconditional sense.

Assumptions, Diagnostics, and Alternatives

These normal calculations rely on several conditions. The preapproval summary must represent evidence relevant to the same parameter or estimand being updated; the postapproval estimate and its standard error must be suitable summaries of its likelihood; and conditional independence must be credible, or their dependence must be modelled. For the worked examples, the normal sampling distributions and prior distributions are assumed adequate, the estimates are on the same scale and in the same units, and their standard errors are treated as known. In an actual analysis, uncertainty in estimated standard errors or more complex model structure may require a fuller likelihood.

Conflict should be examined rather than inferred from a single visual comparison. Compare estimates and uncertainty on a common scale; inspect the postapproval data's likelihood or predictive fit under the borrowing prior; and assess how posterior estimates, uncertainty, and threshold probabilities change under reasonable prior specifications. In a robust mixture, a low posterior borrowing weight indicates that the new estimate is poorly predicted by the borrowing component relative to the weak component. It is a diagnostic conditional on the chosen components, not a universal test or proof that the sources are scientifically incompatible.

A commensurate prior makes the degree of similarity explicit through its additional variance, or through a model that estimates commensurability from data. A robust mixture instead lets observed data reweight a borrowing component against a weak component. Neither approach removes the need to choose and justify prior structure. With no borrowing, a postapproval likelihood is analysed on its own, avoiding historical-prior conflict but potentially leaving substantial uncertainty. The appropriate comparison depends on the scientific relationship between evidence sources and the consequences of an incorrect update.

Sensitivity analysis is essential. Vary the commensurability precision, mixture weight, and weak-component specification over defensible ranges, then report the resulting label-relevant estimates and uncertainty. A narrow interval under one strongly borrowing prior is not persuasive if modest, reasonable changes in assumptions move the estimate or threshold probability substantially.

Common Pitfall

Wrong reasoning: “The historical estimate is more precise, so the postapproval estimate should be pulled toward it; the narrower posterior proves the label estimate is better.” This treats precision as a substitute for compatibility and scientific relevance. Correction: borrowing is conditional on assumptions linking the evidence sources. Check that they address the same parameter, inspect conflict, and test sensitivity to prior choices. Reduced uncertainty is a mathematical consequence of borrowing, not independent evidence that the assumptions are correct.