Why Combine Discovery Evidence Across Ancestry Groups?
A polygenic score depends on estimated genetic effects, and those estimates can be imprecise when a discovery GWAS has limited information. Evidence from multiple ancestry groups may improve precision, include variants that are informative in different groups, and make discovery less dependent on one population’s allele frequencies and linkage disequilibrium (LD). But combining studies is not automatically beneficial: the evidence must refer to comparable effects, and differences in LD, allele frequency, phenotype scale, and true effects must be handled rather than averaged away.
Here, an ancestry group is a study group characterized by genetic ancestry and its associated allele-frequency and LD patterns. These groups are not discrete biological types. The goal is to use their GWAS evidence appropriately, not to assume that every genetic effect is identical or that one group’s results automatically transfer to another.
Precision Gains Under a Shared-Effect Model
Start with a single variant and \(K\) ancestry groups. Let \(\hat{\beta}_a\) be the estimated effect for the same aligned allele in group \(a\), and let \(V_a\) be its variance. If the estimates are approximately unbiased for one common effect \(\beta\), are approximately normal, and are independent across groups, then \(\hat{\beta}_a \sim N(\beta,V_a)\). Independence is appropriate only when the contributing samples do not overlap; otherwise the covariance between estimates must be included.
Under those assumptions, the inverse-variance weighted estimate gives more weight to more precise evidence. Its variance is smaller than that of any individual estimate when each group contributes positive information. The weighting is about precision, not about a claim that one ancestry group is intrinsically more important.
Why the inverse-variance weights arise
Consider a linear combination \(\tilde{\beta}=\sum_a c_a\hat{\beta}_a\) that is unbiased for \(\beta\). Unbiasedness requires \(\sum_a c_a=1\). Under independence, its variance is \(\sum_a c_a^2V_a\). Minimizing this variance subject to the unbiasedness constraint gives the most precise linear unbiased combination.
Thus \(c_a=\lambda/(2V_a)\). Applying \(\sum_a c_a=1\) yields \(\lambda/2=1/\sum_a(1/V_a)\), so \(c_a=(1/V_a)/\sum_j(1/V_j)\). Substituting these weights into the variance gives \(1/\sum_a(1/V_a)\), as stated. The result depends on the common-effect and independence assumptions; it is not a general guarantee that pooling reduces error for every target population.
Worked Example: Combining Two Precise, Aligned Estimates
Suppose two hypothetical GWAS estimates concern the same aligned allele, phenotype scale, and underlying effect. The estimates are \(\hat{\beta}_1=0.20\) with \(SE_1=0.10\), and \(\hat{\beta}_2=0.10\) with \(SE_2=0.05\). Assume independent samples and a common effect. The variances are \(V_1=0.10^2=0.0100\) and \(V_2=0.05^2=0.0025\), giving weights 100 and 400.
The combined estimate is \(0.120\) effect units, with standard error \(0.04472\) effect units. It is more precise than either input estimate because both contribute information, and the second estimate contributes more because its variance is lower. This calculation does not establish that the shared-effect assumption is true; if the true effects differ, the result instead summarizes evidence in a precision-weighted way.
Sample Size and Allele Frequency Determine Information
For a single additive variant in a simple linear regression, suppose the phenotype residual variance is \(\sigma^2\), the genotype is coded as allele count \(0,1,2\), the allele frequency is \(f\), and there are \(N\) independent individuals. Under Hardy–Weinberg proportions, genotype variance is approximately \(2f(1-f)\). Ignoring covariates and other complications, the effect estimate has approximate standard error
This approximation shows why a larger sample generally improves precision and why, at the same \(N\), a variant with frequency near one half may provide more information than a rare variant. It is an illustrative linear-model approximation, not a universal formula: phenotype type, covariates, relatedness, imputation quality, and study design can alter the effective information. For a binary phenotype, for example, the relevant information depends on case-control balance as well as genotype variation.
Worked Example: Equal Sample Sizes, Different Allele Frequencies
Consider two hypothetical single-variant analyses, each with \(N=10{,}000\) independent individuals, residual standard deviation \(\sigma=1\), and the same true effect measured per allele. Assume the simple linear-model approximation above applies and ignore covariates. In the first analysis, the allele frequency is \(f=0.30\); in the second, it is \(f=0.05\).
The estimated standard error is about \(0.03245/0.01543=2.10\) times as large in the lower-frequency group under these assumptions. Thus equal sample sizes do not imply equal precision. A multi-ancestry analysis may add meaningful information where a variant is more frequent, but frequency differences alone do not justify treating effects as equal or interpreting an imprecise estimate as evidence of no effect.
LD Changes What a Marginal Association Represents
A GWAS association at a marker can reflect the marker’s own causal effect, LD with other causal variants, or both. For a simplified model with standardized genotypes, let \(\boldsymbol{\gamma}\) be the vector of causal effects and \(R_a\) the genotype correlation (LD) matrix in ancestry group \(a\). The vector of marginal association effects is approximately
This relation assumes compatible phenotype and genotype scaling and a model in which the causal effects are represented by \(\boldsymbol{\gamma}\). Even if causal effects are shared, different LD matrices can produce different marginal effects at the same marker. Conversely, a weaker marginal association in one group need not mean the underlying causal effect is weaker. Allele frequencies also affect the standard errors and the set of variants that can be estimated reliably. The earlier material on diagnosing portability loss from frequency and LD differences provides related context; here the emphasis is how those differences alter the evidence used at discovery.
Worked Example: The Same Causal Effect, Different Tag-SNP Evidence
Consider two hypothetical ancestry groups with two standardized genotypes: a causal variant and a tag variant. Suppose the causal-effect vector is \(\boldsymbol{\gamma}=(0.10,0)\), so the tag itself has no causal effect. In group A, let the genotype correlation matrix be \(R_A=\begin{pmatrix}1&0.80\\0.80&1\end{pmatrix}\); in group B, let \(R_B=\begin{pmatrix}1&0.10\\0.10&1\end{pmatrix}\). Apply the marginal-effect relation in each group.
The causal variant has the same effect in both groups, but the tag’s marginal effect is \(0.08\) in group A and \(0.01\) in group B. Combining tag-SNP estimates as if they were estimates of the same causal parameter would therefore be inappropriate. A method that uses ancestry-relevant LD information can interpret the evidence more meaningfully than a method that treats all marginal effects as directly interchangeable.
Effect Heterogeneity Changes the Target
The shared-effect model is an assumption to assess, not a default truth. Let the true effect in group \(a\) be \(\beta_a\). Under independent, unbiased estimates with variances \(V_a\), the inverse-variance combination has expectation \(\sum_a w_a\beta_a\), where \(w_a=(1/V_a)/\sum_j(1/V_j)\). It therefore estimates a precision-weighted average of group effects if they differ—not necessarily the effect in any particular target population. This distinction matters when a score is intended for a particular population.
Effect heterogeneity can arise from biological or environmental effect modification, differences in phenotype measurement, or differences in study design and bias. Before interpreting a difference as biological, check allele alignment, phenotype and effect scales, ancestry assignment, and analysis quality. A heterogeneity statistic can flag disagreement beyond estimated sampling error, but it cannot by itself identify the cause of that disagreement.
Worked Example: A Precise Pooled Estimate Can Hide Heterogeneity
Suppose two hypothetical groups have aligned estimates \(\hat{\beta}_1=0.20\) and \(\hat{\beta}_2=0.05\), each with \(SE=0.05\). Assume independent samples. Under a common-effect model, both weights are \(1/0.05^2=400\), so the pooled estimate is \(0.125\) with standard error \(\sqrt{1/800}=0.03536\). For a simple heterogeneity check, calculate \(Q=\sum_a w_a(\hat{\beta}_a-\hat{\beta}_{\mathrm{IVW}})^2\); under the common-effect model and its usual approximations, \(Q\) is compared with a chi-squared distribution with \(K-1\) degrees of freedom.
With \(df=1\), the upper-tail probability for \(Q=4.50\) is approximately \(0.034\). This is evidence of more between-group disagreement than expected from sampling error alone under the common-effect model. The pooled estimate remains a calculable weighted average, but it should not be presented as a shared effect without qualification. The result also does not prove a biological cause of heterogeneity.
Dependence, Diagnostics, and Alternatives
The independent-estimate formula fails when studies share participants or controls, or when other dependencies induce correlated estimation errors. If the vector of estimates has covariance matrix \(\Sigma\), a common-effect generalized least squares estimate is
This form requires an appropriate covariance matrix and still assumes the estimates refer to a common effect. Ignoring positive sample overlap can overstate the precision gained. When dependence is unknown, do not simply add sample sizes or inverse variances as though all evidence were independent.
Useful checks before combining evidence include confirming that effect alleles and reference alleles are aligned, that effect units and phenotype definitions are comparable, that each group’s sample size and allele frequency are plausible, and that the LD information matches the relevant group. Inspect group-specific estimates and standard errors rather than only a combined result. Large, systematic disagreements, unexpectedly extreme estimates, or a combined estimate driven almost entirely by one group may indicate heterogeneity, poor precision, or a data-harmonization problem.
The main approaches differ in what they assume and target:
| Approach | When it is useful | Main limitation |
|---|---|---|
| Pool individual-level data and fit a common-effect model | When data access, consistent harmonization, and a shared-effect model are available | Requires careful control of study and population structure; pooling does not remove heterogeneity |
| Inverse-variance meta-analysis | When group-specific summary estimates target the same effect and their dependence is handled | Can be misleading if marginal effects differ because of LD or true effects differ |
| Analyze groups separately | When effects or study scales may differ and group-specific evidence is important | Smaller groups can have low power and noisy estimates; estimates are not automatically combined into a single target |
| Use a model allowing group-specific effects | When heterogeneity is plausible and the data can support its estimation | Additional flexibility can increase uncertainty, especially with few groups or weak evidence |
Multi-ancestry discovery is useful when it contributes relevant information under a model that respects each group’s precision and genetic structure. Its limitations include sparse data for some groups, uncertain LD estimates, phenotype or design differences, and insufficient information to distinguish effect heterogeneity from noise. It is not a substitute for evaluating score performance in the intended target population.
Common Pitfall
Wrong reasoning: “The largest GWAS has the smallest standard error, so its estimate should determine the score for every ancestry group.” The correction is that precision concerns uncertainty about the effect estimated in that study; it does not establish that the effect is shared, that the study’s marginal association represents the same causal quantity elsewhere, or that its LD pattern is suitable for another group. A larger study can dominate an inverse-variance average while the pooled target remains inappropriate for a particular population.
Another frequent error is to equate a non-significant group-specific result with a different effect. A wide confidence interval may simply indicate low information. Compare estimates and their uncertainty, consider the model’s assumptions, and avoid inferring heterogeneity solely because one group crosses a significance threshold and another does not.
Check Your Understanding
Answer each question, then compare your reasoning with the brief solutions.
- Two independent groups estimate the same effect with standard errors 0.10 and 0.20. Which receives more inverse-variance weight, and how many times more?
- Why can the same causal effect produce different marginal association estimates at a tag variant across ancestry groups?
- In a common-effect analysis, what does the pooled estimate target if the true group effects differ?
- What is one consequence of ignoring sample overlap when combining estimates?
Solutions. (1) Their variances are \(0.0100\) and \(0.0400\), so the first group’s inverse-variance weight is four times the second’s. (2) The groups can have different LD matrices, changing the correlation between the tag and causal variants and therefore changing the tag’s marginal association. (3) It targets a precision-weighted average of group effects under the stated model, not necessarily a shared effect or the effect for a particular target population. (4) If the estimates are positively correlated, treating them as independent generally overstates the information and understates the standard error of the combined estimate.