Posterior Mixture Weights Under Conflict

Why Mixture Weights Change

A robust borrowing prior represents uncertainty about how informative historical evidence should be for the current trial. One component may borrow more strongly from historical information; a protective component is intended to limit borrowing when current observations conflict with that information. The component weights assigned before observing the current data need not remain the weights after observing them. Bayes’ rule updates those probabilities according to how well each component predicts the data.

This tutorial concerns that update in isolation: the posterior probability of each component. It does not derive the parameter distribution conditional on each component or combine those distributions into a robust mixture posterior. Those are separate calculations. Here, the key question is: after seeing the current trial data, how much posterior probability is assigned to each prior component?

The marginal likelihood is also called the prior predictive probability or density under a component. It is not the likelihood evaluated at one selected parameter value: it averages the likelihood over that component’s prior. Each component must use the same observed data and sampling model. If the data are continuous, marginal likelihoods are densities evaluated at the same data value, so their ratio is dimensionless even though each density has units.

Bayes’ Rule for Component Probabilities

For two components, the events \(C=B\) and \(C=P\) are mutually exclusive and exhaustive. The total predictive probability or density of the observed data is therefore the weighted sum of the component marginal likelihoods. Applying Bayes’ rule gives the updated weight for each component.

The expressions can also be written in odds form. Provided both prior weights and both marginal likelihoods are positive, posterior odds equal prior odds multiplied by the marginal likelihood ratio:

The ratio \(m_P(y)/m_B(y)\) measures the relative predictive support that the observed data receive under the two components. If it exceeds 1, the data shift the odds toward the protective component; if it is below 1, they shift the odds toward the borrowing component. The shift is relative: a component can be favored by the data and still have a posterior weight below one half if its prior weight was sufficiently small.

Assumptions and a Short Derivation

The calculation assumes that the component set and prior weights were specified before observing the data; that the components are mutually exclusive and exhaustive; and that each component supplies a well-defined predictive distribution for the same data under the same observation process. The marginal likelihood must integrate over all unknown parameters within that component using its stated prior. The data should not be counted twice—for example, by using the same current observations both to choose a component and again to update its weight as if that choice had been independent of the observations.

To derive the formula, consider the joint probability or density of the observed data and a component. By the definition of conditional probability, \(p(y,C=j)=P(C=j)p(y\mid C=j)=\pi_jm_j(y)\). Summing over the possible components gives \(p(y)=\sum_k\pi_km_k(y)\). Dividing the joint quantity by this total yields the posterior component probability:

For more than two components, the same expression applies, with the denominator summed over every component. Since all prior weights and marginal likelihoods are nonnegative, the resulting posterior weights are nonnegative. Their sum is 1 because the sum of their numerators is exactly the denominator. If a component has prior weight zero, its posterior weight remains zero under this update; data alone cannot restore a component excluded by the prior.

Updating Weights in Hypothetical Trial Examples

The following examples use hypothetical predictive probabilities for specified categories of trial outcomes. They illustrate the weight calculation without requiring a particular parametric prior family. In each case, the component probabilities are assumed to have been obtained by integrating the component’s likelihood over its prior. The categories and probabilities are chosen to make those predictive distributions explicit and comparable.

Reading the Update and Checking the Calculation

The central interpretation is comparative predictive fit. A shift toward the protective component means the observed data are more supportive of that component relative to the borrowing component than they were before the data were observed. It does not mean that the posterior probability that a particular historical study is exchangeable has been directly measured, nor that the probability of a parameter value has changed by the same amount. The weights concern the latent component label in the specified prior mixture.

A practical check is to calculate both the normalized weights and their odds form. They should agree, and the posterior weights should sum to 1. It is also useful to inspect the individual marginal likelihoods and the ratio between them. An extreme ratio can be an appropriate response to conflict, but it can also signal a component that assigns implausibly little predictive probability to ordinary outcomes. Check that:

• Each component predicts the same data in the same representation and on a compatible probability or density scale.
• Each marginal likelihood reflects integration over that component’s prior, not just a selected parameter value.
• The prior weights are the prespecified weights before seeing the current data and sum to 1.
• The likelihood includes the observed data once, with any dependence among outcomes handled correctly.
• The protective component has meaningful predictive support for plausible conflict patterns; a component that is too narrow may fail to protect, while an excessively diffuse component may receive a small marginal likelihood even when it allows the observed result.

A zero marginal likelihood for a component makes its posterior weight zero whenever another component gives the data positive predictive probability. That can happen because the model rules out the observation, but in practice it may reveal a coding error, incompatible outcome definitions, or an overly restrictive component. If every component assigns zero probability or density to the observation, the model does not define a posterior update for that data point; the model or data representation must be examined rather than forcing a normalization.

Alternatives, Uses, and Limitations

One alternative is to keep the mixture weights fixed after observing the data. That may be useful when weights are intended as externally imposed design weights rather than probabilities to be updated, but it is not the Bayesian posterior component-weight calculation described here. Another alternative is hard selection: choose whichever component has the larger marginal likelihood and discard the other. That loses uncertainty about component membership. Posterior weighting instead retains both components when both have positive posterior weight. It is a form of Bayesian model averaging at the component-indicator level, but the weights alone do not provide the full parameter posterior.

These weights are sensitive to the prior specifications because the marginal likelihood averages over each component’s entire prior distribution. A component can fit a particular observation poorly on average even if it includes parameter values that fit it well. Accordingly, a posterior weight should not be interpreted without considering how each component was constructed and what data it predicts. In a clinical trial, the update can communicate how strongly current observations shift the prespecified balance between borrowing and protection, but it cannot compensate for an inappropriate likelihood, an unrepresentative historical source, or an incomplete set of components.

A common mistake is to reason, “The observed data are unlikely under the borrowing component, so the borrowing component’s posterior probability must be zero.” The correction is that posterior probability depends on both marginal likelihoods and prior weights. Unless the borrowing marginal likelihood is exactly zero, its posterior weight may remain positive; even a relatively small borrowing marginal likelihood can produce substantial posterior weight if its prior weight is large enough. Conversely, a protective component is not automatically favored just because it was designed to guard against conflict: the observed data must provide relative predictive support for it.