Hierarchical Models for Cross-Population Borrowing

Why Borrow Across Populations?

A regional treatment-effect estimate can be imprecise when a population has few participants or little local clinical data. An analysis based only on that region may then be unstable. At the other extreme, treating all regions as if they have exactly the same effect can conceal real differences. A hierarchical model offers a middle ground: it allows population effects to vary, while using information across populations to stabilize estimates where the local data are sparse.

The central assumption is not that every population has the same treatment effect. Rather, the effects are modeled as draws from a shared distribution. The resulting borrowing depends on the estimated amount of between-population variation and on the precision of the target region’s own estimate. This is a model-based assumption about how populations relate; it is not guaranteed by the fact that they appear in the same development program.

All effects must refer to a sufficiently comparable estimand, endpoint, time point, and effect scale for this model to be meaningful. For example, an effect measured in symptom-score points cannot be pooled directly with a log odds ratio as though both were on the same scale. The sampling standard errors should represent the uncertainty in each regional estimate. When regional estimates are correlated, such as when they share a control group, their covariance must also be represented; the independent sampling model above is for independent estimates.

What the Variance Components Mean

The sampling variance \(s_j^2\) describes uncertainty in the estimate from population \(j\). The between-population variance \(\tau^2\) describes variation in the underlying effects \(\theta_j\) around the shared mean \(\mu\). These are different sources of uncertainty. A large standard error can arise from sparse data even when the true effects are similar; a large \(\tau\) instead indicates that the model allows substantial differences between population effects.

Conditional on fixed values of \(\mu\) and \(\tau^2\), the marginal distribution of an observed estimate is \(y_j\sim N(\mu,\tau^2+s_j^2)\). Thus, the information about \(\tau^2\) comes from how much the regional estimates vary relative to their sampling variances. If there are few populations, or if their estimates are imprecise, the data may provide little information about \(\tau^2\). An estimate at the boundary \(\tau^2=0\) is possible; it means that the fitted model finds no detectable between-population variation, not that identical effects have been established.

In practice, \(\mu\) and \(\tau^2\) can be estimated by likelihood-based methods, including restricted maximum likelihood. A Bayesian analysis can instead assign distributions to these hyperparameters and estimate their posterior distributions. The calculations below condition on specified fitted values to make the shrinkage mechanism transparent; that plug-in calculation does not include uncertainty in \(\mu\) or \(\tau^2\).

How Partial Pooling Arises

When the hyperparameters are treated as known, the target effect has a normal distribution after observing its regional estimate. Its mean is a precision-weighted average of the local estimate and the shared mean. Completing the square in the product of the normal sampling density and the normal population distribution gives the result.

The coefficient of \(\theta_j^2\) inside the brackets is \(1/s_j^2+1/\tau^2\), so the posterior variance is the reciprocal of that precision. The coefficient determining the posterior mean is \(y_j/s_j^2+\mu/\tau^2\). Therefore,

where \(w_j=\tau^2/(\tau^2+s_j^2)\). Thus, the local estimate receives more weight when its sampling variance is small relative to the between-population variance. If local data are sparse and \(s_j^2\) is large, the estimate moves toward \(\mu\). If \(\tau^2\) is large relative to \(s_j^2\), the model permits more regional variation and the local estimate retains more influence.

Assumptions, Diagnostics, and Alternatives

Partial pooling is useful only if the modeled relationship among populations is defensible. The normal distribution for population effects assumes that deviations around \(\mu\) are reasonably represented by a symmetric distribution on the chosen effect scale. The common \(\tau^2\) assumes that the same degree of between-population variation applies across the modeled regions. Exchangeability means that, after any modeled distinctions, the regions can reasonably be treated as draws from this shared distribution. It does not mean the regions are identical, nor does it establish that a new target is exchangeable with them.

Check that the regional estimates and their uncertainties are comparable and that the model captures relevant differences in design, population, endpoint, and treatment use. Inspect the estimates against the fitted shared distribution and examine whether a particular region strongly influences \(\mu\) or \(\tau^2\). Large discrepancies relative to the reported sampling errors, a fitted \(\tau^2\) near zero despite visibly divergent effects, or substantial changes when one region is omitted are reasons to investigate model fit and population comparability. A target far outside the characteristics represented by the contributing regions is especially difficult to justify by borrowing alone.

The main alternatives differ in how much sharing they impose. Separate regional analyses do not borrow and can leave a sparse target estimate very imprecise. Complete pooling sets all regional effects equal, effectively imposing \(\tau^2=0\); it may be efficient when effects are genuinely alike but can understate regional differences if that assumption is wrong. A hierarchical model estimates a compromise, with the degree of shrinkage governed by the local standard error and \(\tau\). These methods answer different questions and should not be chosen solely by which gives the narrowest interval.

A major limitation of the simple conditional calculations is that \(\mu\) and \(\tau^2\) are estimated from finite data. Treating them as fixed can understate uncertainty, particularly when only a few regions contribute. A full analysis should use a fitting approach that propagates hyperparameter uncertainty and should report sensitivity to reasonable modeling choices. Borrowing cannot correct a biased regional estimate, harmonize incompatible estimands, or supply information about a target population that is not represented by the model’s assumptions.