Meta-Analyzing GWAS Effects for Portable Scores

Why Meta-Analyze Ancestry-Specific GWAS Effects?

A multi-ancestry GWAS can produce separate effect estimates for each ancestry group, then combine them to gain precision or summarize effects across groups. The combination is not neutral: its weights determine which groups contribute most, and its assumptions determine what the pooled effect means. A highly precise meta-analysis can therefore be useful for discovery while still failing to represent effects in an underrepresented target population.

This tutorial focuses on inverse-variance fixed- and random-effects meta-analysis of ancestry-specific estimates. The estimates must refer to the same variant, effect allele, phenotype, and effect scale. For example, per-allele log-odds effects can be combined when they use the same outcome definition and covariate conventions. Alleles must be aligned so that each estimate describes the same allele contrast. In the standard formulas below, sampling errors from different ancestry groups are assumed independent; overlapping samples require methods that account for covariance.

Definition: Let \(\hat{\beta}_j\) be the estimated effect for ancestry group \(j\), and let \(SE_j\) be its standard error, for \(j=1,\ldots,K\). The inverse-variance weight is \(w_j=1/SE_j^2\). A fixed-effect meta-analysis assumes that every group shares one true effect \(\beta\), with differences among \(\hat{\beta}_j\) arising from sampling error. A random-effects meta-analysis models group-specific true effects as \(\beta_j=\mu+u_j\), where the \(u_j\) are independent, have mean zero, and have variance \(\tau^2\). Its target \(\mu\) is the mean of the modeled distribution of true effects, not automatically the effect in any particular target ancestry.

Fixed-Effects Weights and Pooled Estimate

Under the fixed-effect model, each ancestry-specific estimate has sampling variance \(SE_j^2\) around the same true effect. More precise estimates have smaller standard errors and therefore receive larger weights. With independent estimates, the inverse-variance pooled estimate and its standard error are

\hat{\beta}_{FE}=\frac{\sum_{j=1}^{K}w_j\hat{\beta}_j}{\sum_{j=1}^{K}w_j}, \qquad SE(\hat{\beta}_{FE})=\sqrt{\frac{1}{\sum_{j=1}^{K}w_j}}.

The pooled estimate is a weighted average, and the weights sum to the total information used in its standard error. This formula is appropriate only when the common-effect assumption is a reasonable model for the intended analysis. It does not say that the groups have identical allele frequencies, linkage disequilibrium patterns, or score performance; it concerns the particular effect parameter being combined.

Worked Example: A Fixed-Effect Estimate on the Log-Odds Scale

Suppose two hypothetical ancestry groups have harmonized per-allele log-odds estimates for the same binary phenotype. Group A has \(\hat{\beta}_A=0.12\) and \(SE_A=0.04\); group B has \(\hat{\beta}_B=0.08\) and \(SE_B=0.08\). Assume independent sampling errors and a common true effect.

The inverse-variance weights are \(w_A=1/0.04^2=625\) and \(w_B=1/0.08^2=156.25\). Thus group A contributes four times the weight of group B. The pooled estimate is

\hat{\beta}_{FE} =\frac{625(0.12)+156.25(0.08)}{625+156.25} =\frac{75+12.5}{781.25} =0.112.

Its standard error is \(\sqrt{1/781.25}=0.03578\) log-odds units per allele. A normal-approximation 95% confidence interval is \(0.112\pm1.96(0.03578)=(0.0419,0.1821)\). This interval describes uncertainty about the common-effect estimate under the model and sampling assumptions; it is not a 95% probability statement about a fixed parameter. Exponentiating the pooled log-odds estimate gives an odds ratio of \(\exp(0.112)=1.119\) per effect allele under that common-effect interpretation.

Measuring Between-Group Heterogeneity

Before interpreting a fixed-effect summary across ancestry groups, examine whether the observed estimates vary more than expected from their standard errors alone. Cochran’s \(Q\) measures their weighted dispersion around the fixed-effect estimate:

Q=\sum_{j=1}^{K}w_j(\hat{\beta}_j-\hat{\beta}_{FE})^2.

If the common-effect model holds and the estimates are approximately normal with known standard errors, \(Q\) has approximately a chi-square distribution with \(df=K-1\). A large \(Q\) relative to that distribution is evidence against the common-effect model. The test can have low power when there are few groups or imprecise estimates, so a non-significant result does not establish that effects are identical.

The commonly reported \(I^2\) summarizes the relative excess dispersion:

I^2=\max\left(0,\frac{Q-(K-1)}{Q}\right)\times 100\%.

It is a descriptive estimate of the proportion of the observed variation in estimates that is beyond what sampling error alone would predict, under the conventional formulation. It is not the percentage of ancestries with different effects, nor a measure of the magnitude of effect differences. Its value also depends on estimate precision and the set of groups analyzed.

Worked Example: Calculating Heterogeneity in Three Groups

Consider hypothetical harmonized log-odds estimates of \(0.20\), \(0.10\), and \(0.80\), with standard errors \(0.05\), \(0.10\), and \(0.20\), respectively. Assume the groups are independent and the estimates are on a comparable scale. The fixed-effect weights are \(400\), \(100\), and \(25\), so the fixed-effect estimate is

\hat{\beta}_{FE} =\frac{400(0.20)+100(0.10)+25(0.80)}{400+100+25} =\frac{110}{525} =0.2095.

The contributions to \(Q\) are \(400(0.20-0.2095)^2=0.0363\), \(100(0.10-0.2095)^2=1.200\), and \(25(0.80-0.2095)^2=8.717\). Therefore \(Q\approx9.952\), with \(df=3-1=2\). The chi-square upper-tail probability is about \(0.0069\), providing evidence against a common effect in this example. The heterogeneity summary is

I^2=\frac{9.952-2}{9.952}\times100\%\approx79.9\%.

This indicates substantial excess dispersion relative to sampling error by this measure. It does not identify why the estimates differ: differences could reflect genuine effect modification, phenotype or analysis differences, or other sources of inconsistency. Those possibilities need scientific and data-quality assessment rather than interpretation from \(Q\) alone.

Random-Effects Weights

A random-effects model allows the true effect to vary across groups. If the between-group variance is \(\tau^2\), the variance of an observed estimate around the mean effect \(\mu\) is \(SE_j^2+\tau^2\). The inverse-variance random-effects weight is therefore

w_j^*=\frac{1}{SE_j^2+\tau^2}, \qquad \hat{\mu}_{RE}=\frac{\sum_{j=1}^{K}w_j^*\hat{\beta}_j}{\sum_{j=1}^{K}w_j^*}, \qquad SE(\hat{\mu}_{RE})=\sqrt{\frac{1}{\sum_{j=1}^{K}w_j^*}}.

Unlike fixed-effect weights, random-effects weights include estimated between-group variance. As \(\tau^2\) grows, differences in sampling precision matter less and the weights become more similar. The value of \(\tau^2\) must be estimated; several methods exist, and the method should be reported. A simple method-of-moments estimate is the DerSimonian–Laird estimator:

\hat{\tau}^2_{DL} =\max\left(0,\frac{Q-(K-1)}{C}\right), \qquad C=\sum_{j=1}^{K}w_j-\frac{\sum_{j=1}^{K}w_j^2}{\sum_{j=1}^{K}w_j}.

This estimate is zero when \(Q\le K-1\). The random-effects weights then equal the fixed-effect weights. With few ancestry groups, \(\hat{\tau}^2\) can be highly uncertain; treating it as known when calculating the standard error can understate uncertainty. The interval from the simple formula above is a conventional plug-in interval, not a guarantee that uncertainty in \(\tau^2\) has been fully represented.

Worked Example: Constructing Random-Effects Weights

Use the three estimates from the heterogeneity example. The fixed-effect weights sum to \(525\), and their squared weights sum to \(400^2+100^2+25^2=170{,}625\). Hence

C=525-\frac{170{,}625}{525}=200, \qquad \hat{\tau}^2_{DL}=\frac{9.952-2}{200}=0.03976.

The random-effects weights are \(1/(0.05^2+0.03976)=23.66\), \(1/(0.10^2+0.03976)=20.10\), and \(1/(0.20^2+0.03976)=12.54\). The pooled estimate and its plug-in standard error are

\hat{\mu}_{RE} =\frac{23.66(0.20)+20.10(0.10)+12.54(0.80)} {23.66+20.10+12.54} \approx0.2979, \qquad SE(\hat{\mu}_{RE})\approx\sqrt{\frac{1}{56.30}}=0.1333.

The random-effects estimate is larger than the fixed-effect estimate of \(0.2095\) because the least precise group, with the largest estimate, receives a larger relative share of the total weight. A normal-approximation 95% interval using this plug-in standard error is approximately \(0.2979\pm1.96(0.1333)=(0.0367,0.5591)\) log-odds units per allele. The estimate describes the mean of the modeled group-effect distribution; it does not estimate the true effect for each group individually.

What the Weights Mean for Portability

A meta-analysis can increase precision without making its pooled effect representative of every population. Fixed-effect weights are determined by standard errors, which in turn depend partly on sample size and allele frequency. When one ancestry group contributes far more information, its estimate can dominate the pooled result. Random-effects weights reduce this dominance when estimated heterogeneity is positive, but they do not give each ancestry equal influence and do not target a particular underrepresented population.

For a polygenic score, a pooled GWAS effect is one ingredient in the score’s variant weights. Whether those weights transfer depends on more than the meta-analytic standard error: allele-frequency differences affect score variance, linkage disequilibrium affects tagging, and true effects may differ. Thus the meta-analysis answers a question about combining effect estimates; it does not by itself demonstrate equivalent prediction performance or calibration across target populations.

Worked Example: A Precise Pooled Effect May Miss the Target

Suppose a hypothetical discovery GWAS has an ancestry group A estimate of \(0.20\) log-odds units per allele with \(SE=0.04\), and an ancestry group B estimate of \(0.50\) with \(SE=0.20\). Assume the estimates are independent and allele-aligned. A fixed-effect calculation gives weights \(625\) and \(25\), respectively, so

\hat{\beta}_{FE} =\frac{625(0.20)+25(0.50)}{650} =\frac{137.5}{650} =0.2115, \qquad SE(\hat{\beta}_{FE})=\sqrt{\frac{1}{650}}=0.03922.

Group B contributes only \(25/650=3.85\%\) of the total fixed-effect weight. The pooled estimate is precise, but it is close to group A’s estimate because group A supplies most of the information. If group B is the intended target and its effect is genuinely near \(0.50\), the pooled value is not a target-specific estimate: it is \(0.2885\) log-odds units per allele below the group B estimate in this hypothetical example. Using it as a score weight could therefore misrepresent the target effect at this variant. The fixed-effect assumption is also questionable given the observed difference, so the narrow pooled interval should not be read as evidence of portability.

Checks Before Reporting a Meta-Analysis

The calculations are only as interpretable as the inputs and model. Before combining estimates, verify that all effects use the same allele orientation and compatible phenotype, covariate adjustment, and effect scale. Confirm whether samples overlap; if they do, the estimates have correlated errors and the independent-estimate formulas above are not valid without adjustment. Also inspect ancestry-specific estimates and standard errors rather than reporting only a pooled value and one heterogeneity statistic.

A useful distinction is between a precision question and a target question. Fixed effects can be efficient when a common effect is scientifically defensible. Random effects represent variation around an average effect and can reduce domination by the largest study. Neither choice automatically answers, “What effect should I use in this particular target ancestry?” That question requires target-relevant evidence and an explicit estimand.

Common pitfall: “The meta-analysis has a narrow confidence interval, so the effect must transfer well to every ancestry.” This reasoning confuses precision for a pooled parameter with representativeness for a target group. A narrow interval can arise because one well-powered group dominates the weights; it does not establish a common effect, eliminate heterogeneity, or validate score performance in an underrepresented population.

Harmonize the estimates.
Check the effect allele, phenotype, effect scale, and analysis conventions before calculating weights.

Choose the model for the estimand.
Use fixed effects for a defensible shared effect; use random effects when the target is a mean across a modeled distribution of effects.

Report heterogeneity and composition.
Give ancestry-specific estimates, standard errors, weights, and heterogeneity summaries, with appropriate caution when only a few groups are available.

Check Your Understanding

Use the definitions and calculations above to answer the following questions.

Why does an ancestry group with a smaller standard error receive more weight in a fixed-effect meta-analysis?
In the three-group example, what does \(I^2\approx79.9\%\) describe, and what does it not establish?
How does positive \(\tau^2\) change the relative weights in a random-effects analysis?
Why can a narrow fixed-effect confidence interval fail to provide evidence that a pooled GWAS effect is suitable for an underrepresented target population?
Name two input or design checks that are needed before applying the independent inverse-variance formulas.