Worked Example of Trans-Ancestry Meta-Analysis

From Ancestry-Specific Results to a Pooled Estimate

A trans-ancestry meta-analysis combines association estimates from studies in different ancestry groups. Its appeal is practical: pooling can increase precision, while comparing the group-specific results can reveal whether a single common effect is a reasonable summary. But the pooled number depends on the model and on each result’s precision. It is not automatically an effect that applies equally to every population.

The worked calculations below use hypothetical results for one genetic variant and a binary outcome. The effect allele is harmonized across groups, and each estimate is a log odds ratio per additional copy of that allele, adjusted for the study’s stated covariates. The allele frequency is 0.30 in ancestry group A and 0.08 in group B. The lower frequency in group B may contribute to less information, along with sample size and other study features; the standard errors, rather than allele frequencies alone, determine the inverse-variance weights used here.

Before pooling, check that the effect alleles and effect scales match, that the estimates are suitable for approximately normal inference, and that the groups’ estimates are independent or that their covariance is accounted for. Here we assume independent ancestry-specific samples, valid within-group association analyses, and reported standard errors that adequately represent sampling uncertainty. If samples overlap, the ordinary inverse-variance formulas below do not account for that dependence.

Worked Example: Fixed- and Random-Effects Estimates

Why the Formulas Work

Under the fixed-effect model, suppose the estimates are independent and approximately normally distributed with common true effect \(\beta\) and known sampling variances \(SE_i^2\). The weighted squared distance from a candidate pooled value \(b\) is \(S(b)=\sum_i w_i(\hat{\beta}_i-b)^2\). Differentiating and setting the derivative to zero gives

Because the estimates are independent, the variance of this weighted mean is \(\sum_i w_i^2SE_i^2/(\sum_i w_i)^2=1/\sum_i w_i\), since \(w_i=1/SE_i^2\). This gives the fixed-effect standard error used above. The statistic \(Q=\sum_i w_i(\hat{\beta}_i-\hat{\beta}_{FE})^2\) measures dispersion around that weighted mean; under the common-effect model and the stated normal assumptions, it has approximately a chi-squared distribution with \(k-1\) degrees of freedom.

The DerSimonian–Laird estimate is a method-of-moments estimate: it compares observed dispersion \(Q\) with its expected value under a common effect. Under the random-effects model, that expectation is approximately \((k-1)+\tau^2 C\), where \(C=\sum_i w_i-\sum_i w_i^2/\sum_i w_i\). Solving for \(\tau^2\) gives \([Q-(k-1)]/C\); truncating at zero prevents a negative variance estimate. Random-effects weights then include both sampling variance and estimated between-group variance. When \(\hat{\tau}^2\) is substantial, this added variance reduces the relative dominance of the most precise group.

Precision, Allele Frequency, and Model Choice

Ancestry groups need not contribute equally. In the first example, group A has four times the fixed-effect weight of group B because its standard error is half as large. Allele-frequency differences can help explain such precision differences: a less frequent allele may yield less genotype variation and less information, all else equal. But sample size, phenotype variability, imputation quality, study design, and adjustment also matter. A frequency difference alone does not determine the weight, and a larger weight means greater statistical precision—not greater biological importance.

Diagnostics, Alternatives, and Limitations

A useful review begins before calculating a pooled value. Confirm allele alignment, effect direction, units, and effect scale; check that standard errors and estimates refer to comparable analyses; and establish whether study samples overlap. Then inspect the ancestry-specific estimates and confidence intervals alongside \(Q\), rather than treating a heterogeneity \(p\)-value as a pass/fail decision. A surprisingly large \(Q\), a high \(I^2\), or estimates with opposing directions calls for investigation of harmonization, phenotype definitions, model adjustment, and genuine effect differences. A non-significant \(Q\) with only two groups is weak evidence against heterogeneity, not reassurance that none exists.

The fixed-effect model is appropriate when the scientific target is one common effect and that assumption is credible. The random-effects model targets a mean across a distribution of group-specific effects, but a mean does not imply that every group has an effect near it. With few groups, \(\hat{\tau}^2\) is unstable, and the usual random-effects confidence interval can understate uncertainty in the estimated mean. Neither model corrects confounding, allele misalignment, or incompatible effect scales.

A common mistake is to interpret a non-significant heterogeneity test as evidence that the effect is identical across ancestries, then report the fixed-effect estimate as universally applicable. The correction is to distinguish “no statistically clear heterogeneity in these data” from “effects are equal,” show the group-specific estimates, and state which model defines the pooled target. In cross-population work, frequency, linkage disequilibrium, effect scale, and study design can all shape observed associations; a pooled association does not by itself identify why group estimates differ.