Choosing a Borrowing Structure
Robust borrowing is not one model with a universal conflict switch. A robust mixture and a commensurate prior encode different views of how historical and current treatment effects may relate. The mixture specifies alternative prior regimes, such as informative borrowing or a protective, weakly informative prior. A commensurate prior instead models the difference between historical and current effects, with a precision parameter controlling how tightly they are linked.
That distinction matters before any computation. If the intended assumption is “the current effect is either sufficiently similar to the historical effect for borrowing, or it is not,” a mixture makes that discrete alternative explicit. If the intended assumption is that differences can range smoothly from small to large, a commensurate model is more direct. Neither structure guarantees protection from a poor historical source; protection depends on the component choices, prior probabilities, and model assumptions.
Both formulations need a well-defined, comparable treatment-effect estimand. The cohorts should be modelled with appropriate likelihoods and their sampling uncertainties, and any shared or different populations, outcome definitions, follow-up periods, and treatment versions must be considered. A mixture does not make incompatible estimands comparable; a commensurate prior does not remove bias from a flawed historical analysis. If a current likelihood is conditionally independent of historical data given the parameters, that independence also needs to be plausible.
How Conflict Changes a Robust Mixture
In a two-component mixture, one can introduce a latent indicator \(Z\), with \(Z=I\) for the informative component and \(Z=P\) for the protective component. The prior probabilities are \(P(Z=I)=w\) and \(P(Z=P)=1-w\). Once current data are observed, the posterior component probabilities depend on how well each component predicts those data. The component with the larger predictive density gains posterior weight; closeness of component means alone does not determine the result.
For a normal illustration, suppose the current estimate \(Y\mid\theta_C\sim N(\theta_C,s^2)\), and component \(k\) gives \(\theta_C\mid Z=k\sim N(m_k,v_k)\). Integrating over \(\theta_C\), the component predictive distribution is \(Y\mid Z=k\sim N(m_k,v_k+s^2)\). Bayes’ rule then gives:
where \(\pi_k\) is the prior component probability and \(\phi(\,\cdot\,;m,v)\) is a normal density with mean \(m\) and variance \(v\). These expressions assume the stated normal likelihood and proper normal component priors. The overall posterior for \(\theta_C\) is still a mixture: average the component posterior distributions using the updated component probabilities.
Worked Example: Updating Mixture Weights Under Conflict
Suppose a hypothetical trial has current estimate \(y=0.10\) treatment-effect units with standard error \(s=0.20\), so \(s^2=0.04\). Set prior probability \(0.5\) on each component. Let the historical-informed component be \(N(0.40,0.005)\), and the protective component be \(N(0,0.50)\). These variances are in squared treatment-effect units.
The predictive variances are \(0.005+0.04=0.045\) for the informative component and \(0.50+0.04=0.54\) for the protective component. The respective predictive densities at \(y=0.10\) are:
Thus the posterior probability of the informative component is \(0.5(0.692)/[0.5(0.692)+0.5(0.538)]\approx0.563\); the protective component retains probability \(0.437\). The current estimate is closer to the protective component’s prior mean of 0 (distance \(0.10\)) than to the informative component’s mean of \(0.40\) (distance \(0.30\)). Nevertheless, the informative component has the higher predictive density at this observation and therefore gains posterior probability. The broader protective component has a lower peak density, which is relevant to this comparison.
Within the informative component, the posterior mean is \(\{0.40/0.005+0.10/0.04\}/(1/0.005+1/0.04)\approx0.367\). Within the protective component it is \(\{0/0.50+0.10/0.04\}/(1/0.50+1/0.04)\approx0.093\). Weighting those means gives an overall posterior mean of \(0.563(0.367)+0.437(0.093)\approx0.247\) treatment-effect units. It is a weighted summary of two posterior regimes, not evidence that one regime is certainly true.
The calculation also illustrates why a mixture’s protective behaviour is not automatic. It depends on the protective prior’s scale and shape, its prior probability, and its predictive fit. A very diffuse protective prior can have low density across much of the parameter space, while a protective prior centred in an inappropriate region may not protect against the conflict of concern.
How Commensurability Encodes Borrowing
A basic commensurate model says that the current effect \(\theta_C\) is centred at the historical effect \(\theta_H\), with variance \(1/\tau\). Conditional on a fixed \(\tau\), this produces continuous normal shrinkage rather than a choice between two prior regimes. If \(\tau\) is uncertain, a prior on it allows the degree of commensurability to be learned, subject to the information in the current and historical data and the chosen prior. A spike-and-slab construction is a special way to represent distinct degrees of commensurability; it does not make the general commensurate model identical to a protective mixture.
A short normal calculation makes the role of precision explicit. Suppose historical information gives \(\theta_H\mid D_H\sim N(h,V_H)\), and the commensurate relationship is \(\theta_C\mid\theta_H,\tau\sim N(\theta_H,1/\tau)\). Conditional on fixed \(\tau\), integrating over \(\theta_H\) gives \(\theta_C\mid D_H,\tau\sim N(h,V_H+1/\tau)\). If a current estimate \(Y\mid\theta_C\sim N(\theta_C,V_C)\) is independent of the historical information conditional on the parameters, multiplying the two normal densities yields:
The posterior mean is a precision-weighted average of the historical-informed mean \(h\) and current estimate \(y\). Increasing \(\tau\) reduces \(1/\tau\), making the historical-informed prior variance smaller and ordinarily increasing the historical contribution. At small \(\tau\), the commensurate link is weak, so current data have more influence. This is conditional on the specified model and fixed \(\tau\); with uncertain \(\tau\), the posterior averages across its possible values.
Worked Example: Interpreting a Commensurate Precision
Consider hypothetical historical information summarised by \(h=0.40\) with \(V_H=0.04\), and a current estimate \(y=0.10\) with \(V_C=0.04\), all in treatment-effect units or their squares as appropriate. Set \(\tau=4\), so \(1/\tau=0.25\) squared units. The historical-informed prior variance for the current effect is \(0.04+0.25=0.29\).
The posterior variance and mean are:
The resulting posterior standard deviation is \(\sqrt{0.03515}\approx0.1875\) treatment-effect units. The estimate is pulled toward the historical value, but the link is sufficiently weak that the current estimate remains influential. This is a conditional calculation for the chosen \(\tau\), not a data-driven estimate that \(\tau=4\) is appropriate.
Model Choice, Computation, and Diagnostics
The central choice is the borrowing structure the scientific assumptions support. A robust mixture is attractive when investigators can defend a meaningful set of prior regimes and want explicit prior probability on a protective alternative. Its posterior response is interpretable through component weights. A commensurate prior is attractive when differences are naturally graded and a continuous relationship between historical and current effects is plausible. Its response is governed by the commensurability precision and its prior.
| Consideration | Robust mixture | Commensurate prior |
|---|---|---|
| Basic structure | Weighted alternative prior components | Conditional relationship between current and historical effects |
| Conflict response | Posterior component weights change according to predictive fit | Posterior for commensurability precision controls the degree of shrinkage |
| Prior elicitation | Choose component distributions and their prior weights | Choose a scale or distribution for commensurability precision |
| Computation | Finite mixtures may permit componentwise calculations; latent indicators can also be used | Hierarchical inference must account for the historical effect and precision parameter |
| Potential weakness | A poorly chosen protective component or weight can give inadequate protection | A poorly calibrated precision prior can impose too much or too little linkage |
For normal models with a finite number of components, marginal likelihoods and component-specific posterior distributions can often be evaluated directly and combined. A latent-indicator implementation is another option, but the indicator must be updated in a way that explores the components adequately. Commensurate models add hierarchical dependence and often require inference over \(\tau\); the computational burden depends on the likelihood and precision prior. These are structural differences, not a guarantee that one model is always computationally easier. In either case, check that numerical results are stable and that the implementation matches the stated model.
Before fitting, examine whether the informative prior and the commensurate model predict plausible current data. After fitting, compare the observed estimate with the component-specific or commensurate predictive distributions, inspect posterior component weights or the posterior for \(\tau\), and check whether conclusions are being driven by a narrow modelling choice. Prior predictive simulation can expose implausible scales or predictions. These checks diagnose model-data tension; they do not establish that historical and current populations are exchangeable, nor do they identify the cause of a discrepancy.
Specify the current and historical effects, their likelihoods, and which differences between cohorts could undermine borrowing.
Use a mixture when distinct borrowing regimes are defensible; use a commensurate prior when a graded relationship is the intended assumption.
Inspect prior predictive implications and then assess posterior component probabilities or commensurability, along with numerical stability.
A practical failure mode is to treat either mechanism as a detector of every source of conflict. If the historical treatment effect is biased, the current and historical estimands differ, or covariate distributions create a misleading comparison, a high posterior weight on the informative component or strong commensurability may still be inappropriate. Conversely, a model can reduce borrowing because of random variation in a small current sample. Interpret conflict measures alongside design knowledge and the clinical meaning of the effects.
Common mistake: “The current estimate is closer to the protective component’s mean, so the protective component must gain posterior probability.” Distance between means is not the posterior-weight calculation. The correction is to evaluate each component’s predictive density at the observed data, multiply by its prior weight, and normalise. In the worked mixture example, the estimate is closer to the protective mean, yet the informative component gains weight because its predictive density is higher.
The limitations are consequential. A finite mixture only represents the regimes and distributions that were specified; it cannot automatically adapt to an omitted kind of conflict. A commensurate prior imposes a parameterisation and a prior scale on differences; results can be sensitive to those choices. Both approaches can understate uncertainty if historical information is treated as more comparable or more precise than it is. The model should therefore follow the scientific borrowing rationale, rather than being selected solely because it produces a more favourable estimate.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- In a robust mixture with prior component probabilities \(0.6\) and \(0.4\), the predictive densities at the observed data are \(0.20\) and \(0.50\). What is the posterior probability of the first component?
- What does increasing \(\tau\) do to the variance \(1/\tau\) in a commensurate prior, and what is the usual implication for borrowing?
- Why can the component whose mean is farther from the current estimate still gain posterior probability?
- Give one modelling situation in which a robust mixture is more natural than a single continuous commensurate relationship.
Solutions. 1. The first component’s posterior probability is \(0.6(0.20)/[0.6(0.20)+0.4(0.50)]=0.12/0.32=0.375\). 2. Increasing \(\tau\) decreases \(1/\tau\), tightening the prior link between current and historical effects and usually increasing shrinkage toward the historical effect, conditional on the rest of the model. 3. Posterior component probabilities depend on predictive densities, which incorporate both distance from the mean and predictive variance, as well as prior component probabilities. 4. A mixture may be natural when investigators explicitly regard the historical evidence as either sufficiently comparable to borrow from or sufficiently non-comparable to rely on a protective alternative.