Fixed-Effect Trans-Ancestry Meta-Analysis

Why Combine Ancestry-Stratified Estimates?

Association studies may estimate an effect separately in each ancestry group, then combine those estimates to obtain a more precise summary. This can be useful when the groups share the same underlying association effect on the chosen scale but have different sample sizes or genotype information. The fixed-effect method gives more weight to estimates with smaller standard errors.

The key qualification is substantive, not merely computational: a fixed-effect analysis assumes one common effect across the ancestry groups being combined. Differing allele frequencies can affect how precisely each group estimates that effect. They do not, by themselves, establish that the effects are equal or unequal. The earlier material on testing effect heterogeneity across ancestries addresses how to assess evidence of differences; here the focus is on combining estimates under a shared-effect assumption.

Model, Assumptions, and Derivation

For each ancestry group, suppose the association analysis estimates the change in a specified outcome associated with a one-unit increase in the same genetic predictor. The predictor coding, outcome definition, effect scale, and relevant model specification must be comparable across groups. For example, if the effect is a per-allele log-odds ratio, all estimates must refer to that same log-odds scale and the same effect allele. The estimates should also come from analyses that adequately control the relevant sources of confounding within each group.

A standard working model for the summary estimates is \(\hat{\beta}_g \sim N(\beta,SE_g^2)\), with independent estimates across groups. The normal approximation is generally justified by large-sample association analyses. The common \(\beta\) is the shared-effect assumption. Independence is reasonable for non-overlapping ancestry samples; if samples overlap, the estimates may be correlated and the ordinary inverse-variance formula is not appropriate without accounting for that covariance. In practice, \(SE_g\) is estimated, and the formula uses those estimated standard errors as plug-in measures of uncertainty.

The pooled estimator can be derived as the value of \(\beta\) that minimizes the weighted squared discrepancies between the group estimates and a common effect. With \(w_g=1/SE_g^2\), define \[ L(\beta)=\sum_{g=1}^{G}w_g(\hat{\beta}_g-\beta)^2. \] Differentiating and setting the derivative to zero gives \[ \frac{dL}{d\beta}=-2\sum_{g=1}^{G}w_g(\hat{\beta}_g-\beta)=0, \qquad \hat{\beta}_{FE}=\frac{\sum_g w_g\hat{\beta}_g}{\sum_g w_g}. \] Under the model, its variance follows from independence: \[ \operatorname{Var}(\hat{\beta}_{FE}) =\frac{\sum_g w_g^2 SE_g^2}{(\sum_g w_g)^2} =\frac{\sum_g w_g}{(\sum_g w_g)^2} =\frac{1}{\sum_g w_g}. \] This gives the stated standard error. A large-sample Wald interval is \(\hat{\beta}_{FE}\pm 1.96SE(\hat{\beta}_{FE})\), and a two-sided test of \(H_0:\beta=0\) uses \(z=\hat{\beta}_{FE}/SE(\hat{\beta}_{FE})\), compared with the standard normal distribution.

Allele Frequencies and Information

Ancestry groups can have different allele frequencies, and that can alter the precision of an association estimate. To see the mechanism in a simple setting, consider a hypothetical linear regression of a continuous outcome on genotype dosage \(G\), with an intercept, independent observations, and errors with variance \(\sigma^2\). If genotype dosage has variance \(\operatorname{Var}(G)\), the standard error of the estimated genotype coefficient is approximately \(\sigma/\sqrt{n\operatorname{Var}(G)}\). Under Hardy–Weinberg equilibrium for a biallelic genotype with effect-allele frequency \(f\), \(\operatorname{Var}(G)=2f(1-f)\). This illustration assumes no additional covariates; in an adjusted analysis, the relevant information depends on genotype variation remaining after adjustment.

Interpreting the Pooled Result and Checking Its Fit

The pooled estimate is a precision-weighted common-effect estimate, not an unweighted average of ancestry-specific effects and not an estimate of the effect in every individual. If the shared-effect assumption is credible, the result summarizes the common association on the specified scale. If true effects differ, the formula still produces a number, but that number is a precision-weighted summary of the supplied estimates; it need not correspond to a single effect that applies to each ancestry group.

Before interpreting the result, check that the group estimates target comparable quantities, that their directions use a consistent allele definition, and that their standard errors are valid. Review the estimates and intervals by group rather than relying only on the pooled result. Large discrepancies in direction or magnitude are a warning that the common-effect model may be inadequate. The earlier material on effect heterogeneity provides a formal assessment; a pooled \(p\)-value alone does not establish homogeneity.

Choosing a Summary and Recognizing Limitations

Fixed-effect pooling is most useful when a common association effect is a defensible target and the group estimates are independent, comparable, and adequately calibrated. Its principal advantage is efficiency: it combines information while giving more influence to estimates with lower sampling variance. Its principal limitation follows from the same weighting rule. Groups with more information can dominate the pooled estimate, so it need not describe the effect in a smaller group or represent an equal-weight average across ancestries.

Reporting ancestry-stratified estimates alongside a pooled estimate makes the weighting and any disagreement visible. If the scientific goal is to describe group-specific effects, the stratified estimates may be more informative than a single summary. If there is credible between-group variation, a model that allows such variation is an alternative, but it answers a different question and requires additional assumptions. A fixed-effect analysis should not be used as a substitute for that model simply because it gives a convenient single estimate.

The method also depends on comparable effect scales. A shared numerical coefficient has a coherent interpretation only when the outcome, predictor, and coefficient scale are aligned. Differences in allele frequency primarily influence information in the idealized model above; other population differences can also alter the estimated association target. Thus, a similar-looking coefficient does not by itself prove that the same underlying biological or causal quantity has been estimated. Keep the interpretation at the level supported by the association model.

Common Pitfalls

• Assuming that a pooled estimate proves a common effect. The formula returns a weighted average even when group estimates disagree. Check the ancestry-stratified results and assess whether the shared-effect assumption is plausible.
• Using sample size alone as the weight. The standard fixed-effect weight is \(1/SE_g^2\). Sample size is only one determinant of precision; genotype variation and other features of the analysis matter too.
• Interpreting different precision as different effect. A group with a larger standard error may simply contain less information. Precision and effect magnitude are distinct quantities.
• Ignoring dependence between estimates. The standard error formula assumes independence. Overlapping samples can induce covariance, which must be accounted for rather than treated as independent information.
• Reporting only the pooled result. A single precise number can conceal meaningful ancestry-specific disagreement. Include the stratum-level estimates when interpreting trans-ancestry evidence.