Meta-Analysis Versus Joint Trans-Ancestry Fine-Mapping

Association Evidence Is Not the Same as Variant Resolution

Combining genetic association results across ancestry groups can increase precision and make a locus easier to detect. But finding a strong association is not the same as identifying which variant at that locus is causal. A correlated, noncausal variant can show a strong association because it is in linkage disequilibrium (LD) with a causal variant. The pattern of that correlation can differ across ancestry groups.

The key distinction is what gets combined. A conventional meta-analysis often combines an estimate and standard error separately for each variant. This can be useful for testing association, but a set of pooled statistics does not, by itself, describe how each ancestry’s LD generated its pattern of associations. Joint trans-ancestry fine-mapping retains ancestry-specific association patterns and their LD information in the likelihood.

These approaches are not opposites in every implementation. Meta-analysis can be one step in a broader analysis, and a meta-analysis result may be followed by fine-mapping that uses suitable LD information. The important distinction is whether ancestry-specific patterns and LD are retained for the fine-mapping calculation, rather than reducing the evidence to pooled association statistics alone.

What a Fixed-Effect Meta-Analysis Combines

Suppose ancestry group \(a\) provides an estimate \(\hat{\beta}_a\) for the effect of the same harmonized effect allele on the same trait, with standard error \(SE_a\). Under a fixed-effect model, the true effect is assumed common across groups, and the estimates are treated as independent with approximately normal sampling distributions. The inverse-variance estimate and its standard error are:

The corresponding Wald statistic is \(Z_{\text{meta}}=\hat{\beta}_{\text{meta}}/SE_{\text{meta}}\). It summarizes evidence for association at that variant under the model. It does not say whether that variant is causal, or whether a correlated variant generated the association.

This calculation requires estimates to refer to the same allele and trait scale, and the groups to be independent for the stated variance formula. The fixed-effect interpretation additionally assumes a common underlying effect. If effects genuinely differ, a pooled estimate can still be computed, but it describes a weighted average rather than a common effect. Harmonization and heterogeneity diagnostics are therefore relevant to interpretation, not optional details.

Why LD Changes the Fine-Mapping Question

For a region containing \(p\) variants, let \(z_a\) denote the vector of ancestry-specific association statistics and \(R_a\) the corresponding LD correlation matrix. In a common large-sample approximation with standardized genotypes and a quantitative trait, the summary-statistic distribution can be written as:

Here \(n_a\) is the sample size for group \(a\), and \(\gamma_a\) is the vector of standardized genetic effects. This approximation assumes the summary statistics and LD reference describe comparable samples and allele coding; actual implementations may use other scales or likelihoods. Its essential message is that LD affects both the expected association pattern and the covariance among statistics. If variant \(j\) is the only causal variant, the mean pattern is proportional to column \(j\) of \(R_a\), not simply a spike at \(j\).

With independent ancestry groups, the joint likelihood combines their contributions while preserving their different \(R_a\) matrices. In a simple shared-effect illustration, \(\gamma_a=\gamma\) for each group. The likelihood is proportional to the product of the ancestry-specific distributions, so the observed pattern in each group can be evaluated against the LD pattern expected under each causal candidate. Other models can allow group-specific effects; the central point here is that the LD structure can remain ancestry-specific.

What Pooling Retains and What It Can Lose

A pooled statistic can retain substantial information about whether a variant is associated. It can also preserve some ability to distinguish variants, especially when the pooled association patterns differ clearly and the fine-mapping procedure uses an appropriate LD model. But pooling the statistics can discard which ancestry-specific pattern contributed to the result.

For illustration, with two equally precise independent groups, a common-effect meta-analysis of \(Z\)-statistics at one variant is \(Z_{\text{meta}}=(Z_A+Z_B)/\sqrt{2}\). The scaling gives a statistic with unit variance under the null. The operation is a sum: it does not record the two input values separately. For instance, ancestry-specific pairs \((4,0)\) and \((2,2)\) produce the same pooled value, \(4/\sqrt{2}\), even though the distribution of evidence across groups differs. If their LD structures differ, those patterns may have different implications for candidate causal variants.

This is a direct information-loss argument. If a pooled statistic is \(T=Z_A+Z_B\), then for any value \(d\), the pair \((Z_A+d,Z_B-d)\) has the same \(T\). The map from two ancestry-specific values to one pooled value is therefore not one-to-one. A joint likelihood can use the separate values and their ancestry-specific covariance structures; the pooled value alone cannot reconstruct them.

A Short Derivation of the Joint Versus Pooled Distinction

Under the normal approximation, the ancestry-specific log likelihood for a proposed effect vector \(\gamma_a\), apart from constants, is:

For independent groups, the joint log likelihood is \(\ell_{\text{joint}}=\sum_a\ell_a\). Thus each group contributes a residual measured against its own expected association pattern and covariance. If the candidate causal variant changes, the relevant column of each \(R_a\) changes, potentially by a different amount in each group.

By contrast, a pooled statistic combines observations before this ancestry-specific comparison. Because summing is many-to-one, the pooled value does not generally determine all the \(z_a\) vectors. When LD differs, those missing vectors can be informative about which candidate better explains the evidence. This justification does not imply that every joint analysis is more precise or that every meta-analysis loses all useful information. It explains why the two procedures answer distinct questions and why a pooled association statistic alone is not a substitute for a joint fine-mapping likelihood.

When the Distinction Matters

The distinction is most important when variants are correlated, LD differs appreciably across ancestry groups, and the analysis aims to prioritize causal variants rather than simply test association. In that setting, a large group may dominate an inverse-variance pooled result, even if a smaller group’s distinct LD pattern is valuable for separating candidates. Conversely, if LD is similar across groups or the smaller group has little information, retaining ancestry-specific data may not materially change resolution.

Before interpreting results, check that variant alleles and trait scales are aligned, that the LD matrices match the relevant ancestry groups and variant set, and that the assumed effect model is suitable. Incomplete variant coverage can also change which candidates can be compared; a variant absent or poorly imputed in one group cannot contribute the same evidence there. A technically correct pooled association analysis may therefore still be inadequate for the causal-resolution question.