Multi-Ancestry Shrinkage With Heterogeneous Effects

Why Shrinkage Must Allow Heterogeneity

A genetic effect estimate from one ancestry can be imprecise, especially when the available sample is small. Estimates from other ancestries may add information, but treating all effects as identical can conceal genuine differences. Multi-ancestry shrinkage addresses this tension: it lets estimates inform one another while allowing the underlying effects to vary.

The goal is not to declare one ancestry’s result universally correct or to guarantee portability. It is to estimate ancestry-specific effects using a model that states how much similarity is expected, and to carry uncertainty about both the effects and that similarity into the result. In this tutorial, effects are on a common, harmonized scale—for example, log odds per effect allele for the same phenotype and model.

The shared distribution is a modeling assumption, not a biological fact. Ancestry labels describe groups defined for a particular analysis; they do not imply that every person in a group has the same genetic background or effect. Effect differences can also reflect phenotype definition, environment, study design, linkage disequilibrium, or other features of the analysis.

A Hierarchical Model for Ancestry-Specific Effects

Consider one genetic variant and \(A\) ancestry groups. Let \(\hat{\beta}_a\) be the estimated effect in ancestry \(a\), \(s_a\) its standard error, and \(\beta_a\) the underlying effect for that ancestry. A common summary-statistic model is:

The first line represents sampling uncertainty: conditional on the true effect, the estimate is approximately normal with standard error \(s_a\). The second line represents effect heterogeneity: ancestry-specific effects are modeled as draws around a shared mean \(\mu\), with between-ancestry standard deviation \(\tau\). The model commonly assumes conditional independence of the estimates across ancestries. If samples overlap or estimates are otherwise correlated, their sampling covariance must be included rather than ignored.

The model is usually completed by specifying prior distributions for \(\mu\) and \(\tau\). For example, one could give \(\mu\) a normal prior centered at \(m_0\), and \(\tau\) a prior supported on nonnegative values. The prior for \(\mu\) expresses beliefs about the shared effect’s location; the prior for \(\tau\) expresses beliefs about the scale of cross-ancestry differences. These choices should be justified on the effect scale and examined for sensitivity. In a fully Bayesian analysis, uncertainty in \(\mu\) and \(\tau\) is retained. An empirical Bayes analysis estimates one or both from the data, often with less representation of hyperparameter uncertainty.

How the Shrinkage Calculation Works

First suppose \(\mu\) and \(\tau\) are treated as known, and focus on one ancestry. The likelihood for its observed estimate is normal with mean \(\beta_a\) and variance \(s_a^2\); the prior for \(\beta_a\) is normal with mean \(\mu\) and variance \(\tau^2\). Multiplying these normal densities gives a normal posterior. Its variance is the reciprocal of the sum of the two precisions:

The posterior mean is the same precision-weighted combination of the estimate and the shared center:

To verify the final expression, substitute \(V_a=s_a^2\tau^2/(s_a^2+\tau^2)\) into the precision-weighted form. The coefficient of \(\hat{\beta}_a\) becomes \(\tau^2/(s_a^2+\tau^2)\), and the coefficient of \(\mu\) becomes \(s_a^2/(s_a^2+\tau^2)\); they sum to one. Thus a less precise estimate (larger \(s_a\)) is pulled more toward \(\mu\), while a larger heterogeneity scale \(\tau\) leaves more weight on the ancestry-specific estimate. This formula is conditional on fixed hyperparameters; it does not by itself account for uncertainty in them.

What the Prior Assumptions Do

The shrinkage amount depends directly on the heterogeneity assumption. A tight prior on \(\tau\) implies that ancestry-specific effects are likely to cluster closely around \(\mu\). A more dispersed prior permits larger differences, so an individual ancestry’s estimate is less strongly drawn toward that center. The location prior for \(\mu\) matters too, particularly when few ancestries or few informative estimates are available. A prior centered at zero, for instance, can pull effects toward zero through the shared mean as well as through the ancestry-level hierarchy.

For a fixed \(\mu\) and \(\tau\), the marginal distribution of an ancestry’s estimate is:

This follows because the estimate contains both between-ancestry variation, with variance \(\tau^2\), and sampling variation, with variance \(s_a^2\). With several ancestries, this marginal likelihood helps distinguish these sources of variation, although the distinction can be weak when there are few groups or all estimates are imprecise. A prior on \(\tau\), or an empirical Bayes estimate of it, can therefore have substantial influence.

Borrowing When a Target-Ancestry Estimate Is Sparse

If the target ancestry has little or no direct information, a hierarchical model can produce a prediction from the shared distribution. Such a prediction is not a target-specific measurement: it is conditional on the assumed relationship among ancestry effects. The distinction is especially important when the target ancestry is poorly represented in the data used to estimate the shared distribution.

Use in Genetic Prediction and Important Limitations

At a variant, a posterior mean may be used as a less noisy effect estimate, and estimates across variants may contribute to polygenic-score weights. In practice, that use requires care. Effects must refer to compatible phenotypes and scales, and alleles must be aligned so that estimates point in the same direction. If estimates are correlated—for example, because discovery samples overlap—the model should represent their sampling covariance. Similarly, fitting each variant independently does not automatically account for dependence among variants due to linkage disequilibrium.

Hierarchical shrinkage is most valuable when it balances information across groups without erasing real differences. But sparse data create a difficult boundary: the model may report a precise-looking estimate largely because the shared distribution is restrictive, even when the target ancestry contributes little evidence. A prior cannot create target-specific information. A posterior interval may remain broad, and its center may move when reasonable prior assumptions change.

The model also does not establish why effects differ, guarantee prediction performance in a new population, or make an association causal. Estimated differences can reflect biological heterogeneity, measurement and design differences, or model misspecification. These estimates should therefore be interpreted as conditional on the phenotype, analysis, ancestry definitions, and statistical model used.

Common Mistake

Wrong reasoning: “The ancestry-specific estimates were shrunk toward a shared value, so the effects must be the same across ancestries.” Shrinkage is not a test of equality. It is the consequence of a model that gives effects a shared distribution, with the amount of movement determined by its heterogeneity assumptions and the precision of the estimates.

Correction: State whether the model allows heterogeneity, report or describe the assumed or estimated \(\tau\), and retain uncertainty about ancestry-specific effects. If \(\tau\) is near zero, check whether that conclusion is well supported by the number and precision of the available estimates, or is being driven by the prior.