Evaluating multi-ancestry genome-wide association methods: Statistical power, population structure, and practical implications

Summary

The increasing availability of diverse biobanks has enabled multi-ancestry genome-wide association studies (GWASs) to enhance the discovery of genetic variants across traits and diseases. However, the choice of an optimal method remains debated, due to challenges in statistical power differences across ancestral groups and approaches to account for population structure. Two primary strategies exist: (1) pooled analysis, which combines individuals from all genetic backgrounds into a single dataset while adjusting for population stratification using principal components, increasing the sample size and statistical power but requiring careful control of population stratification; and (2) meta-analysis, which performs ancestry-group-specific GWASs and subsequently combines summary statistics, potentially capturing fine-scale population structure but facing limitations in handling admixed individuals. Using large-scale simulations with varying sample sizes and ancestry compositions, we compare these methods alongside real data analyses of eight continuous and five binary traits from the UK Biobank (N ≈ 324,000) and the All of Us Research Program (N ≈ 207,000). Our results demonstrate that pooled analysis generally exhibits better statistical power while effectively adjusting for population stratification. We further present a theoretical framework linking power differences to allele-frequency variations across populations. These findings, validated across both biobanks, highlight pooled analysis as a powerful and scalable strategy for multi-ancestry GWASs, improving genetic discovery while maintaining rigorous population structure control.

Graphical abstract

Multi-ancestry GWASs enhance discovery in diverse populations, but optimal methods remain debated. Using theory, simulations, and analyses from the UK Biobank and All of Us, we show that pooled analysis outperforms meta-analysis in power while controlling for population structure, driven by allele-frequency differences across ancestries.

Introduction

Genome-wide association studies (GWASs) have played a crucial role in identifying genetic variants associated with various traits and diseases.^1,2 However, these studies have historically been dominated by individuals of European ancestry, who comprised approximately 94.5% of study participants as of 2025,³ limiting the broader applicability of genetic discoveries. This imbalance poses challenges for the generalizability of genetic findings across populations, since allele frequencies, linkage disequilibrium (LD) patterns, and genetic architectures vary across ancestries.^4,5,6 To address this gap, researchers have increasingly incorporated participants from diverse genetic backgrounds into multi-ancestry GWASs.^7,8,9,10,11 These studies leverage genetic diversity to identify novel variants and refine polygenic risk scores (PRSs), with recent multi-ancestry PRS methods demonstrating improved cross-population risk prediction.^{12,13,14,15,16,17}

Despite these advances, questions remain regarding the optimal methodology for multi-ancestry GWASs. Two primary strategies are commonly employed: pooled analysis and meta-analysis. In pooled analysis, individuals from all ancestries are analyzed together in a single model, with principal components (PCs) included to account for population stratification. This approach maximizes sample size, accommodates admixed individuals, and can improve statistical power. However, it also raises concerns about residual confounding due to imperfect correction for population structure and may be less sensitive to ancestry-specific or local ancestry effects compared to stratified analyses.

Meta-analysis, in contrast, conducts ancestry-group-specific GWASs and then combines the summary statistics.¹⁸ This method better accounts for fine-scale population structure, facilitates data sharing when individual-level data are restricted, and may better account for heterogeneous effect sizes across populations. An extension, MR-MEGA,⁵ leverages allele-frequency differences among contributing studies to boost power and handle admixed individuals. However, this method introduces additional parameters that can reduce power, especially when dealing with complex admixture. One important limitation of meta-analysis, particularly when aggregating smaller cohorts, is that population structure correction using PCs may be less effective: PCs computed from small samples often fail to capture subtle stratification adequately, potentially leading to residual confounding.¹⁹

Both strategies can be implemented using fixed-effect or mixed-effect models. Fixed-effect modeling (e.g., PLINK2²⁰) assumes genetic effects are constant across individuals, providing computational efficiency but limited ability to handle cryptic relatedness. In contrast, mixed-effect modeling^{21,22,23,24,25} includes both fixed and random effects to account for population structure and relatedness, enhancing robustness at the cost of increased computational demands. This approach is particularly useful in large biobank studies, where cryptic relatedness is common, and case-control imbalances may introduce biases if not properly accounted for.

In this study, we systematically compare three methods, pooled analysis, fixed-effect meta-analysis, and MR-MEGA,⁵ using both fixed-effect and mixed-effect frameworks. Our primary goal is to identify which approach optimally balances statistical power and population stratification control in multi-ancestry contexts. We conduct large-scale simulations with individuals from five ancestry groups, varying sample sizes, ancestry-group proportions, and outcomes (continuous and binary). To further assess the impact of varying levels of admixture, we simulate admixed individuals using the Admix-kit pipeline.²⁶ Finally, we validate our findings in real-world data by analyzing eight continuous and five binary traits from two large biobanks, the All of Us Research Program²⁷ (AoU, N ≈ 207,000) and the UK Biobank²⁸ (UKB, N ≈ 324,000).

Our results demonstrate that pooled analysis achieves higher statistical power than meta-analysis and MR-MEGA across a range of study designs while maintaining well-controlled type I error in realistic scenarios. We further propose a theoretical framework linking these power gains to allele-frequency differences across ancestry groups. These findings support pooled analysis as a robust and scalable approach for multi-ancestry GWASs, improving genetic discovery and enhancing the generalizability of GWAS findings across populations.

Material and methods

Theoretical motivation for method discrepancies

Consider a multi-ancestry cohort comprising $J$ distinct subcohorts (ancestry groups). Let $n_j$ denote the number of subjects in subcohort $j$, and let $f_j$ be the allele frequency of a causal variant in subcohort $j$. We assume that the allelic effect of this variant ($\beta$) is constant across ancestry groups. For individual $i$ in population $j$, the genotype $g_{ij}$ follows a binomial distribution $g_{ij}\sim Bin\left(2,f_j\right)$. The phenotype is defined as

$$y_{ij}=\beta g_{ij}+{\epsilon }_{ij},{\epsilon }_{ij}\sim N\left(0,{\sigma }^2\right)\text{,}$$

where $i=1,\dots ,n_j$ and $j=1,\dots ,J$, and ${\sigma }^2$ represents the residual variance.

Within-population GWAS

If we conduct a separate linear regression in each subcohort (with intercept), we obtain

$${\hat{\beta }}_j=\frac{{\sum }_{i=1}^{n_j}\left(y_{ij}-{\overline{y}}_j\right)\left(g_{ij}-{\overline{g}}_j\right)}{{\sum }_{i=1}^{n_j}{\left(g_{ij}-{\overline{g}}_j\right)}^2}$$

and

$$Var\left({\hat{\beta }}_j\right)=\frac{{\sigma }^2}{{\sum }_{i=1}^{n_j}{\left(g_{ij}-{\overline{g}}_j\right)}^2}\text{,}$$

where ${\overline{y}}_j$ and ${\overline{g}}_j$ are the phenotypic and genotype means, respectively, in subcohort $j$.

Pooled-analysis GWAS

When pooling all individuals across the $J$ subcohorts into a single GWAS, we have

$${\hat{\beta }}_P=\frac{{\sum }_{j=1}^J{\sum }_{i=1}^{n_j}\left(y_{ij}-\overline{y}\right)\left(g_{ij}-\overline{g}\right)}{{\sum }_{j=1}^J{\sum }_{i=1}^{n_j}{\left(g_{ij}-\overline{g}\right)}^2}$$

and

$$Var\left({\hat{\beta }}_P\right)=\frac{{\sigma }^2}{{\sum }_{j=1}^J{\sum }_{i=1}^{n_j}{\left(g_{ij}-\overline{g}\right)}^2}\text{,}$$

where $\overline{y}$ and $\overline{g}$ are the global phenotype and genotype means, respectively.

Fixed-effect meta-analysis

Using inverse-variance weighting, the fixed-effect meta-analysis estimate is

$${\hat{\beta }}_M=\frac{{\sum }_{j=1}^J{\sum }_{i=1}^{n_j}\left(y_{ij}-{\overline{y}}_j\right)\left(g_{ij}-{\overline{g}}_j\right)}{{\sum }_{j=1}^J{\sum }_{i=1}^{n_j}{\left(g_{ij}-{\overline{g}}_j\right)}^2}$$

and

$$Var\left({\hat{\beta }}_M\right)=\frac{{\sigma }^2}{{\sum }_{j=1}^J{\sum }_{i=1}^{n_j}{\left(g_{ij}-{\overline{g}}_j\right)}^2}\text{.}$$

As $n_j\to \infty$, for all $j$, both ${\hat{\beta }}_M$ and ${\hat{\beta }}_P$ converge to $\beta$.

Asymptotic relative efficiency

We define the asymptomatic relative efficiency (ARE) of meta-analysis vs. pooled-analysis as

$$\text{ARE}\left({\hat{\beta }}_P,{\hat{\beta }}_M\right)=\frac{Var\left({\hat{\beta }}_M\right)}{Var\left({\hat{\beta }}_P\right)}=\frac{{\sum }_{j=1}^J{\sum }_{i=1}^{n_j}{\left(g_{ij}-\overline{g}\right)}^2}{{\sum }_{j=1}^J{\sum }_{i=1}^{n_j}{\left(g_{ij}-{\overline{g}}_j\right)}^2}\text{.}$$

It follows that

$$\text{ARE}\left({\hat{\beta }}_P,{\hat{\beta }}_M\right)=\frac{{\sum }_{j=1}^J{\sum }_{i=1}^{n_j}{\left(g_{ij}-{\overline{g}}_j\right)}^2+{\sum }_{j=1}^J{n}_j{\left({\overline{g}}_j-\overline{g}\right)}^2}{{\sum }_{j=1}^J{\sum }_{i=1}^{n_j}{\left(g_{ij}-{\overline{g}}_j\right)}^2}>1\text{,}$$

because of the standard ANOVA decomposition:

$${\sum }_{j=1}^J{\sum }_{i=1}^{n_j}{\left(g_{ij}-\overline{g}\right)}^2={\sum }_{j=1}^J{\sum }_{i=1}^{n_j}{\left(g_{ij}-{\overline{g}}_j\right)}^2+{\sum }_{j=1}^J{n}_j{\left({\overline{g}}_j-\overline{g}\right)}^2\text{.}$$

The term ${\sum }_{j=1}^J{n}_j{\left({\overline{g}}_j-\overline{g}\right)}^2$ represents the sample-size-weighted variance in allele frequencies across populations. When allele frequencies differ substantially across subcohorts, the pooled analysis benefits from leveraging higher variance in carrier counts across populations, improving power. If one population is disproportionately large (e.g., $n_2\ll n_1$ in a two-population settings), the advantage of pooled analysis diminishes because the dominant population drives the association signal, and the sample-size-weighted variance in allele frequencies will be small.

We have further extended this theoretical framework to allow for ancestry-specific heterogeneity in effect sizes (see Note S1). Specifically, we consider a model where allelic effects vary across populations as independent and identically distributed (i.i.d.) deviations from a shared mean effect. This formulation allows the ancestry-specific deviations to follow an arbitrary distribution under the i.i.d. assumption. In this setting, we show that pooled analysis still retains a power advantage over fixed-effect meta-analysis, even when effect sizes are not identical across ancestry groups.

Implementation of methods

We applied the following approaches in both simulations and real data analyses using AoU and UKB datasets.

Pooled and meta-analysis

For pooled analysis, a GWAS was performed on the entire dataset by combining all ancestry groups into a single analysis. Population structure was adjusted using the top ten cross-ancestry-group PCs derived from the full pooled dataset, along with covariate adjustments for age and sex. Variants were included in the pooled analysis if they had minor allele frequency (MAF) >1% in at least one ancestry group. For meta-analysis, a GWAS was conducted separately within each ancestry group. Top ten ancestry-group-specific PCs were used for population structure adjustment, along with age and sex as covariates. Each ancestry-specific GWAS included only variants with MAF >1% within that ancestry group. Summary statistics from individual ancestry analyses were then combined using a fixed-effect meta-analysis approach with inverse-variance weighting. As a result, variants were included in the final meta-analysis if they met the MAF >1% threshold in at least one ancestry group, ensuring comparability with pooled analysis.

Both approaches used REGENIE²¹ (--qt for continuous traits and --bt for binary traits) for the mixed-effect modeling and PLINK2²⁰ (--linear for continuous traits) for the fixed-effect modeling. REGENIE was selected as a scalable approach for mixed-model association testing, as it uses a local ridge regression framework to adjust for population structure and relatedness. While REGENIE does not explicitly introduce a random effect term, its regularization-based approach functionally approximates mixed-effect modeling by accounting for cryptic relatedness and polygenic effects, distinguishing it from standard fixed-effect models implemented in PLINK2.

MR-MEGA

MR-MEGA extends meta-analysis by incorporating ancestry-group-specific summary statistics while explicitly accounting for allele-frequency differences across ancestry groups. MR-MEGA models ancestry-group-specific genetic effects as a function of population structure, using PCs derived from allele-frequency variation across groups. The input for MR-MEGA is similar to that for standard meta-analysis, using summary statistics from either mixed-effect or fixed-effect GWAS results. The number of axes of genetic variation used in MR-MEGA (--pc) was set to the maximum allowed under the package recommendation, which is the number of distinct genetic ancestry groups minus 3 (i.e., strictly less than the number of groups minus 2). For example, we used --pc 2 for UKB (five ancestry groups) and --pc 3 for AoU (six ancestry groups). MR-MEGA was applied only when combining data from at least four ancestry groups. For example, MR-MEGA was not used for prostate cancer in UKB, as only three ancestry groups had a sufficient number of cases. It is important to note that MR-MEGA restricts its output to variants with MAF >1% in all genetic ancestry groups, while meta-analyses include variants with MAF >1% in at least one ancestry group.

Type I error simulations

To evaluate type I error, we used real genomic data from six genetic ancestry groups as defined in AoU,²⁷ namely African (AFR), admixed American (AMR), East Asian (EAS), European (EUR), Middle Eastern (MID), and South Asian (SAS), to simulate null phenotypes under varying degrees of population stratification. Employing actual genotype data allows us to capture realistic LD structures within each ancestry. For both ancestry-group-specific and cross-ancestry-group scales, stratification was introduced by varying the percentage of the null phenotype’s variance explained by the first ten PCs, ranging from 0% to 5%.

Percentage of variation captured by the top ten ancestry-specific or cross-ancestry PCs in AoU

Before simulating phenotypes, we estimated the variance explained by the first ten PCs in real polygenic traits. Five representative traits were selected: height, high-density lipoprotein cholesterol (HDL), low-density lipoprotein cholesterol (LDL), total cholesterol (TC), and waist circumference. To quantify the variance explained by ancestry-group-specific PCs and cross-ancestry-group PCs, we applied a two-step regression approach.

To quantify the variance explained by ancestry-group-specific and cross-ancestry-group PCs, we employed a two-step regression procedure. First, to evaluate ancestry-group-specific PCs, we regressed the phenotype on age, sex, and the first ten cross-ancestry-group PCs (PCGs), combining all ancestry groups in a single model:

$$Y\sim Age+Sex+PC{G}_1+...+PC{G}_{10}\text{,}$$

where $Y=\left[Y_{AFR},Y_{AMR},Y_{EAS},Y_{EUR},Y_{MID},Y_{SAS}\right]$ is the vector of phenotypes for the different ancestries. From this model, We obtained the residuals for each ancestry group by regressing out the effects of age, sex, and the first ten cross-ancestry-group PCs. The residuals are denoted as $\left[{\epsilon }_{AFR},{\epsilon }_{AMR},{\epsilon }_{EAS},{\epsilon }_{EUR},{\epsilon }_{MID},{\epsilon }_{SAS}\right]$. Next, for each ancestry group, we fit two models to estimate the variance explained by the ancestry-specific PCs: a full model with the top ten ancestry-group-specific PCs and a reduced model with only intercept. The ancestry-group-specific variance explained was calculated as

$$R^2\left({\epsilon }_{ANC}\sim P{C}_{ANC1}+...+P{C}_{ANC10}\right)-R^2\left({\epsilon }_{ANC}\sim 1\right)\text{,}$$

where ANC denotes one of the ancestry groups.

To assess the variance explained by cross-ancestry-group PCs, we reversed the conditioning order. For each ancestry group, we first regressed a model within each ancestry regressing our phenotype of interest again, age, sex, and the top ten ancestry-group-specific PCs:

$$Y_{ANC}\sim Age+Sex+P{C}_{ANC1}+...+P{C}_{ANC10}\text{.}$$

We then combined the residuals across all ancestries into a pooled vector $\epsilon$ and computed the variance explained by cross-ancestry PCs as

$$R^2\left(ϵ\sim P{C}_{ANC1}+...+P{C}_{ANC10}\right)-R^2\left({\epsilon }_{ANC}\sim 1\right)\text{.}$$

Confidence intervals are obtained via bootstrapping using 10,000 replicates (Table S1). These empirical estimates guided our selection of $\alpha$ values $\left(0,0.2\%,0.5\%,1\%,2\%,\text{and}5\%\right)$ in the subsequent simulations.

Null phenotype generation

We simulated population stratification separately on ancestry-group-specific and cross-ancestry-group scales. Empirical results (Table S1) indicated that after accounting for ancestry-group-specific PCs, cross-ancestry PCs explained 0% variance, justifying this independent modeling. To simulate ancestry-group-specific population stratification, we set

$$y_{ij}=\sum _{k=1}^{10}{\delta }_{jk}P{C}_{ijk}+{\epsilon }_{ij},{\epsilon }_{ij}\sim N\left(0,\sqrt{1-\alpha }\right)\text{,}$$

where $P{C}_{ijk}$ is the $k^{\text{th}}$ PC for individual $i$ in ancestry $j$. ${\delta }_{jk}=\sqrt{\frac{w_{jk}\alpha }{Var\left(P{C}_{jk}\right)}}$ is the effect for k^th ancestry-group-specific PC in ancestry $j,\alpha$ is the total percentage of variance explained by the first ten ancestry-group-specific PCs, and $w_{jk}={\lambda }_{jk}/{\lambda }_j$, with ${\lambda }_{jk}$ being the eigenvalue of the $k^{\text{th}}$ PC in ancestry $j$ and ${\lambda }_j$ the sum of the ten eigenvalues for ancestry $j$. This formulation of ${\delta }_{jk}$ ensures that population stratification effects are proportionally assigned across PCs while preserving phenotype variance. The resulting null phenotype has mean 0 and variance 1 within each ancestry group.

Similarly, to simulate cross-ancestry-group population stratification, we used

$$y_{ij}=\sum _{k=1}^{10}{\delta }_{k}P{C}_{ik}+{\epsilon }_{ij},{\epsilon }_{ij}\sim N\left(0,\sqrt{1-\alpha }\right)\text{,}$$

where $P{C}_{ik}$ is the $k^{\text{th}}$ cross-ancestry-group PC for individual i, ${\delta }_k=\sqrt{\frac{w_k\alpha }{Var\left(P{C}_k\right)}}$ is the effect for the k^th cross-ancestry-group PC, $\alpha$ is the total percentage of variance captured by the top ten cross-ancestry-group PCs, and $w_k={\lambda }_k/\lambda$, with ${\lambda }_k$ as the eigenvalue of the $k^{\text{th}}$ cross-ancestry-group PC and $\lambda$ as the sum of the eigenvalues of top ten cross-ancestry-group PCs. As in the ancestry-group-specific case, this design ensures that the null phenotype remains standardized with mean 0 and variance 1 while incorporating a realistic population structure.

To assess the robustness of our eigenvalue-based weighting, we additionally simulated ancestry-specific stratification using empirically estimated PC weights from real phenotypes in AoU. For five traits (height, waist circumference, HDL, LDL, and TC), we estimated the relative contribution of each PC to trait variance within each ancestry using nested $R^2$ comparisons. Specifically, we computed

$$w_{jk}=\frac{R^2\left({\epsilon }_{ANC}\sim P{C}_{ANC,1\text{to}10\text{including}k}\right)-R^2\left({\epsilon }_{ANC}\sim P{C}_{ANC,1\text{to}10\text{excluding}k}\right)}{R^2\left({\epsilon }_{ANC}\sim P{C}_{ANC,1\text{to}10\text{including}k}\right)-R^2\left({\epsilon }_{ANC}\sim 1\right)}\text{,}$$

where ${\epsilon }_{ANC}$ is the phenotype residual after regressing out age, sex, and the first ten cross-ancestry PCs. We replaced the eigenvalue-derived $w_{jk}$ with empirically estimated weights derived from height in the AoU. These empirical weights were then used to define ${\delta }_{jk}$ as before. The resulting null phenotype again had mean 0 and variance 1 within each ancestry group. Full empirical weights are reported in Table S2.

We simulated $\alpha \in \left(0,0.2\%,0.5\%,1\%,2\%,5\%\right)$ to assess different degrees of population stratification. These scenarios reflect both observed values (e.g., 0%–5%) and hypothetical extremes (e.g., 1%–5%). This setup allowed us to assess whether each GWAS method (pooled, meta-analysis, or MR-MEGA) controlled type I error appropriately in the presence of population structure on different scales.

Statistical power simulations

Continuous phenotypes

We used a previously published simulated dataset of 600,000 independent subjects¹³ (120,000 from each of five ancestry groups: AFR, AMR, EAS, EUR, and SAS), generated by HAPGEN2²⁹ and the 1000 Genomes Project (1000G)³⁰ as reference for LD and allele-frequency structures (Figure S1). To generate continuous phenotypes, we randomly selected 5% of variants from the HapMap3 variant list³¹ to be causal. Phenotypes for individual $i$ in population $j$ were simulated as

$$Y_{ij}={\gamma }_j+\sum _{l=1}^L{X}_{ijl}{u}_l+\sum _{k=1}^{2}P{C}_{ijk}{\delta }_{jk}+{\epsilon }_{ij}\text{,}$$

where ${\gamma }_j$ ∼ $N\left(0,1\right)$ is an ancestry-group-specific intercept term. $X_{ijl}=G_{ijl}/\sqrt{2f_{jl}\left(1-f_{jl}\right)}$ is the standardized genetic value for variant $l$ with mean 0 and variance 1. Under the assumption that variant effects on the standardized scale are constant across ancestry groups, we sampled causal effect sizes $u_l\sim N\left(0,\frac{h^2}{L}\right)$, with $L$ being the number of causal variants, and the heritability $h^2$ being set to 0.4. This implies that the per-allele effect size varies with respect to the MAF in each group. Two ancestry-group-specific PCs were included, with ${\delta }_{jk}$ assigned based on the empirical estimates from UKB. The residual error variance was modeled as

$${\epsilon }_{ij}\sim N\left(0,\sqrt{1-\left(h^2+{\delta }_{1k}^{2}var\left(P{C}_{1k}\right)+{\delta }_{2k}^{2}var\left(P{C}_{2k}\right)\right)}\right)\text{.}$$

Ten replicates of continuous phenotypes were generated.

Binary phenotypes

To generate binary traits, we simulated a larger dataset of 5 million individuals (1 million per ancestry group), using HAPGEN2 with 1000G as reference, to ensure adequate sample sizes for case-control sampling and accurate representation of allele-frequency and LD structures. We randomly selected 1% causal variants across the genome. The binary outcomes were generated using a logistic model:

$$g\left(\mathrm{Pr}\left(Y_{ij}=1\right)\right)={\alpha }_j+\sum _{l=1}^L{X}_{ijl}{u}_l+\sum _{k=1}^{2}P{C}_{ijk}{\delta }_{jk}\text{,}$$

where $g\left(x\right)=\log \left(\frac{x}{1-x}\right)$ is the logit function. Similarly as continuous outcomes, we assume $u_l\sim N\left(0,\frac{h^2}{L}\right)$, where logit-scale genetic variance $h^2$ was set to 0.57 to mimic breast cancer.³² The intercept term ${\alpha }_j$ was used to control disease prevalence. We assigned different ${\alpha }_j$ to mimic the estimated prevalence of breast cancer from SEER^∗Stat based on the past 28 years (Table S3). After generating case-control status for all individuals, we randomly sampled cases and controls in a one-to-one ratio to create datasets with an average total sample size of 237,241 cases and 237,241 controls (detailed sample sizes by ancestry group are described in Table S4).

Admixed individuals

To evaluate admixture effects, we simulated an additional dataset of 240,000 individuals using Admix-kit²⁶ and 1000G as the reference. The simulated dataset included four groups: (1) 60,000 individuals with 100% EUR ancestry; (2) 60,000 individuals with 100% AFR ancestry; (3) 60,000 subjects with 50% EUR and 50% AFR admixture; and (4) 60,000 subjects with 20% EUR and 80% AFR admixture (Figure S2).

Admix-kit is a simulation tool that generates realistic admixed genomes by combining haplotypes from multiple ancestral populations. It models admixture as a generational process, where individuals inherit ancestry segments based on recombination patterns over multiple generations. The simulation was conducted in two steps. First, we expanded the reference haplotype data using HAPGEN2, generating a larger pool of phased haplotypes for EUR and AFR populations. To improve computational efficiency, we restricted the variants to those included in the HapMap3 variant list.³¹ In the second step, we simulated admixture by drawing haplotypes from the expanded EUR and AFR datasets using the admix-simu function in Admix-kit. We modeled ten generations of admixture (--n-gen 10), allowing for recombination to shape local ancestry patterns. Ancestry proportions were assigned using the --admix-prop flag to reflect the desired admixture levels in each group. Continuous phenotypes for admixed individuals were simulated following the same methodology described in the previous section.

Statistical power evaluation

We used two metrics to evaluate the performance of GWAS methods: (1) exact recovery of causal variants, which measures the percentage of causal variants that are directly identified at genome-wide significance with $p<5\times {10}^{-8}$, being defined as the number of causal variants reaching genome-wide significance; and (2) LD-based recovery of causal variants, which assesses the percentage of causal variants located within LD regions of genome-wide significant variants. Specifically, this is calculated as the number of causal variants with at least one genome-wide significant variant within a 500-kb region divided by the total number of causal variants.

Real data analysis in AoU and the UKB

For both AoU and UKB datasets, we analyzed eight continuous traits and five binary traits. The continuous traits included height, waist circumference, LDL, TC, HDL, calcium, creatinine, and estimated glomerular filtration rate (eGFR). The binary traits examined were asthma, coronary artery disease (CAD), type 2 diabetes (T2D), breast cancer (BC), and prostate cancer (PC). The corresponding concept IDs for AoU and data field IDs for UKB are provided in Table S5. Subjects with measurements outside predefined ranges for each phenotype were excluded (Table S5).

Data processing and sample selection in AoU

For AoU analyses, we utilized phenotype and genotype array data from AoU v.7.1. Detailed procedures related to genotyping, ancestry classification, quality control measures, and the exclusion of related participants are comprehensively documented in the AoU Research Genomic Research Data Quality Report. The AoU data include both genotype array data and short-read whole-genome sequencing (srWGS) data. Although genotype array data encompass a larger sample size, ancestry and family relatedness information was only available for individuals with srWGS data. To ensure consistent ancestry classification, we restricted analyses to individuals with srWGS data, but we used genotype array variants for GWASs. Variants were filtered based on a MAF threshold of >0.01 within any ancestry group, resulting in approximately 1.2 million variants.

We excluded flagged related subjects, subjects with “sex_at_birth” missing (e.g., “skip,” “I prefer not to answer,” “no matching concept,” or “none”), and individuals with a predicted ancestry probability was ≤0.75. This threshold was applied to ensure a consistent comparison between pooled and meta-analysis approaches, as meta-analysis requires predefined ancestry groups. Including these individuals, who are likely admixed, would have been inconsistent with the ancestry-group-based framework used in meta-analysis. After these exclusions, the final dataset included 207,305 individuals across six genetic ancestry groups: 47,207 AFR, 36,500 AMR, 5,153 EAS, 115,701 EUR, 497 MID, and 2,247 SAS (Figure S3). Phenotype values were taken from the most recent available measurement, and age was calculated as the difference between the date of birth and the date of the latest phenotype measurement.

Data processing and sample selection in UKB

For UKB, we classified individuals into five major continental ancestry groups (AFR, AMR, EAS, EUR, and SAS) using the 1000G³⁰ as a reference. A random forest classification model was trained on the 1000G dataset, using the top 20 PCs as predictors and known ancestry group labels as outcomes. This trained model was then applied to the top 20 PCs within UKB to infer genetic ancestry. To minimize population structure confounding, we restricted analyses to unrelated individuals, resulting in a final dataset of 323,908 subjects across five genetic ancestry groups (6,864 AFR, 590 AMR, 586 EAS, 311,053 EUR, and 5,734 SAS) (Figure S4). Age in UKB was based on “age at visit” (data field ID 21003). Variants for UKB analyses were restricted to HapMap3³¹ + Multi-Ethnic Genotyping Arrays³³ (MEGA) chip array, and further filtered using a MAF threshold of >0.01 within at least one ancestry group, resulting in approximately 1.53 million variants. To ensure a consistent comparison across methods, we used the same set of genotypes within each dataset (AoU and UKB) for LD clumping of genome-wide significant variants across pooled analysis, meta-analysis, and MR-MEGA.

PC analysis computation

To compute PCs, we pruned autosomal variants with MAF >1% across the pooled dataset using an $r^2$ threshold of 0.1, a window size of 500 kb with PLINK2²⁰ (--indep-pairwise). This resulted in a set of 125,692 variants in AoU and 51,922 variants in UKB. The first ten PCs were then computed separately within each genetic ancestry group and within the pooled dataset using PLINK2²⁰ (--pca-approx).

Results

We evaluated three multi-ancestry GWAS strategies: meta-analysis, MR-MEGA,⁵ and pooled analysis (Figure 1). Meta-analysis was implemented with standard fixed-effect meta-analysis using inverse variance weighting (see material and methods). All primary analyses employed mixed-effect modeling with REGENIE²¹ in the main text, while analyses using fixed-effect modeling was performed with PLINK2²⁰ (see supplemental information. In pooled analysis, individuals from all ancestry groups were combined into a single dataset, with cross-ancestry-group PCs controlling for population stratification. In contrast, meta-analysis stratified individuals into genetically homogeneous ancestry groups before combining their summary statistics. MR-MEGA refined meta-analysis by explicitly modeling allele-frequency differences among ancestry groups.

Figure 1.

Overview of analysis

In the pooled analysis (top), principal components (PCs) are computed for the entire dataset, followed by a single GWAS. In the meta-analysis (bottom), PCs are calculated separately for each genetic ancestry group, and independent GWASs are conducted for each. Summary statistics are then combined through meta-analysis using inverse-variance weighting or MR-MEGA. The two approaches are compared based on type I error and statistical power. In simulations with causal variants, statistical power is assessed by the percentage of causal variants exactly recovered and those within the 500-kb linkage disequilibrium (LD) region of a causal variants. In real data analyses, statistical power is measured by the number of significant variants ($p\le 5\times {10}^{-8}$) and the number of independent signals (LD $r^2$ < 0.1).

Our comparisons of multi-ancestry GWAS methodologies focused on two core metrics: (1) type I error, evaluating whether each method properly controls false positives in the presence of population stratification; and (2) statistical power, measured by exact causal variant detection, which measures the proportion of causal variants reaching genome-wide significance, and LD-based detection, which measures the proportion of significant variants within 500 kb of a causal variant.

Statistical power discrepancies between pooled analysis and meta-analysis

Under the assumption of homogeneous allelic effects across ancestry groups, differences in statistical power between pooled analysis and meta-analysis depend on allele-frequency variance across populations (Figure 2). When MAFs are similar across populations, both methods yield comparable power. However, as MAF divergence increases, the relative advantage of pooled analysis becomes more pronounced (see material and methods). This advantage arises because pooled analysis integrates information across populations, allowing individuals from groups with high MAF to contribute more effectively to association detection. In contrast, meta-analysis treats each population separately before combining results, limiting the ability of high-MAF groups to enhance power in low-MAF populations. The power gain of pooled analysis is maximized when allele frequencies vary across ancestries and sample sizes are balanced. However, if one population is disproportionately large, the benefits of pooling diminish as the dominant population drives association signals, reducing the influence of smaller ancestry groups. Overall, pooled analysis achieves better statistical power at a particular locus when allele frequencies differ across populations, with its advantage further amplified when highly divergent alleles are well represented in larger sample sizes.

Figure 2.

Illustration of pooled vs. meta-analysis approaches in multi-ancestry GWASs based on allele-frequency differences

This figure demonstrates that the theoretical differences in power between pooled and meta-analysis approaches in multi-ancestry GWASs are mainly driven by allele-frequency differences. Manhattan plots show simulated results for two known causal variants under two scenarios: (1) similar allele frequencies across ancestry groups (chromosome 4, rs11936438) result in equal statistical power for both pooled and meta-analysis ($-{\log }_{10}p=11.8$, shown in left panel); (2) divergent allele frequencies (chromosome 5, rs404167) show higher power in pooled analysis ($-{\log }_{10}p=9.6$) compared to meta-analyses (${-\log }_{10}p=5.8$, shown in right panel). The lower panels display ancestry-group-specific results for simulated ancestry groups: African (AFR), admixed American (AMR), East Asian (EAS), European (EUR), and South Asian (SAS). The green line represents the minor allele frequency. The tables at the bottom summarize allele frequencies of the causal SNPs across ancestry groups, highlighting how different allele frequencies impact the statistical power of each approach.

Type I error simulations

We evaluated type I error by simulating null phenotypes across six populations using AoU data, ensuring that LD patterns and population stratification reflected real genomic architectures. We investigated both local population stratification, where the phenotype was associated with ancestry-group-specific PCs, and global population stratification, where the phenotype was associated with cross-ancestry-group PCs. To model varying degrees of stratification, we varied the proportion of the null phenotype’s variance explained by the first ten ancestry-group-specific and cross-ancestry-group PCs (see material and methods).

Under both fixed- and mixed-effect modeling approaches, type I error was well controlled with meta-analysis, regardless of the magnitude of ancestry-group-specific or cross-ancestry-group PC effects. Pooled analysis also controlled type I error when cross-ancestry-group PCs were associated with the outcome but exhibited inflation when more than 1% of the null phenotype’s variance was explained by ancestry-group-specific PCs. (see Figure 3 for mixed-effect modeling and Figure S5 for fixed-effect modeling). However, real-world traits typically exhibit much lower ancestry-specific PC variance. In AoU, the median variance explained by the first ten ancestry-specific PCs is only 0.26% across the three largest ancestry groups (Table S1). Higher ancestry-group-specific PC variance (>1%) was primarily observed in smaller ancestry groups such as EAS, SAS, and MID, where estimates may be less stable due to limited sample sizes. While large multi-ancestry GWASs may include some populations with small samples sizes, it is unlikely that most participating ancestry groups exhibit extreme PC variance exceeding 1%. Notably, our simulations only showed inflated p values when multiple ancestries had PC variance above this threshold. This suggests that the level of ancestry-group-specific population structure required to induce meaningful bias in pooled analysis is substantially greater than what is typically observed in well-powered studies.

Figure 3.

Quantile-quantile plots comparing the observed versus expected −log₁₀(p) values for meta-analysis and pooled analysis under varying levels of population structure

Null phenotypes were generated using the All of Us dataset, which includes individuals from six genetic ancestry groups: 47,207 African (AFR), 36,500 admixed American (AMR), 5,153 East Asian (EAS), 115,701 European (EUR), 497 Middle Eastern (MID), and 2,247 South Asian (SAS) (see material and methods). Orange , meta-analysis; blue, pooled analysis.

(A) Quantile-quantile (QQ) plots of mixed-effect GWAS results for null phenotypes with varying percentages (0%–5%) of phenotypic variance are explained by the first ten cross-ancestry-group principal components (PCs).

(B) The same analysis of ancestry-group-specific PCs, with similar ranges of variance explained. Divergence from the diagonal line indicates p value inflation due to population structure in both local and global population stratification scenarios.

To test the robustness of our PC weighting scheme, we repeated the ancestry-group-specific PC simulations using empirical weights estimated from real traits in AoU. Specifically, we replaced the PC weights originally defined based on eigenvalues (i.e., proportional to each PC’s eigenvalue; see material and methods) with weights empirically derived from height (Table S2). The resulting simulations remained consistent and demonstrated that type I error in pooled analysis remained well controlled when total PC variance was ≤1% (Figure S6), even though the empirical PC weight distributions deviated markedly from the eigenvalue-based assumption.

Statistical power simulations

Continuous phenotypes

To assess statistical power across different ancestry-group proportions, we considered three scenarios: (1) equal sample sizes across ancestries while increasing total sample size; (2) a fixed EUR sample size while gradually increasing non-EUR sample size; and (3) a fixed total sample size with increasing non-EUR sample sizes. In all scenarios, pooled analysis consistently outperformed meta-analysis and MR-MEGA (Figure 4; see Figure S7 for corresponding absolute power values).

Figure 4.

Comparison of meta-analysis and MR-MEGA with pooled analysis for detecting causal variants and variants in linkage disequilibrium across different ancestry-group proportions

This figure compares the relative efficiency of meta-analysis (orange) and MR-MEGA (green) in recovering causal variants and variants in LD, using pooled analysis (dashed line) as the benchmark (relative efficiency = 1.0). The top row evaluates the exact recovery of causal variants, measured as the relative efficiency of identifying genome-wide significant ($p\le 5\times {10}^{-8}$) causal variants. The bottom row assesses LD-based recovery of causal variants, which measures the relative efficiency of detecting causal variants that have at least one genome-wide significant variant within 500 kb. Each column represents a different study design: (A) equally sized ancestry groups, (B) a fixed EUR sample size of 120,000 with varying non-EUR sample sizes, and (C) a fixed total sample size of 124,800 with varying non-EUR proportions. All methods are evaluated using GWAS summary statistics generated via REGENIE from simulations of continuous traits in a multi-ancestry dataset composed of African (AFR), admixed American (AMR), East Asian (EAS), European (EUR), and South Asian (SAS) groups.

In the first scenario, where sample sizes were equal across ancestry groups (Figure 4A), meta-analysis initially outperformed MR-MEGA at smaller sample sizes, likely due to the additional parameter fitting required in MR-MEGA. However, once the total sample size exceeded 100,000 individuals (∼20,000 per ancestry across five ancestry groups), MR-MEGA surpassed meta-analysis in identifying causal variants. For LD-based detection (i.e., identifying variants within 500 kb of a causal SNP), MR-MEGA outperformed meta-analysis when each ancestry group had at least 56,000 individuals (total $n=280,000)$.

In the second scenario, we fixed at 120,000 EUR individuals, while gradually increasing non-EUR sample sizes (Figure 4B). This setup reflects the impact of recruiting non-EUR participants within an existing EUR-dominated biobank. When non-EUR groups reached approximately 20% of the EUR sample size (i.e., 120,000 EUR and 24,000 of each non-EUR ancestry, total $n=216,000$), MR-MEGA outperformed meta-analysis in identifying causal variants. This advantages of MR-MEGA became more pronounced when non-EUR groups reached approximately 50% of the EUR sample size.

For the third scenario, we maintained a fixed total sample size of 124,800 individuals while varying non-EUR proportions (Figure 4C). Here, meta-analysis consistently outperformed MR-MEGA for LD-based detection, regardless of ancestry-group proportions. However, MR-MEGA showed a slight advantage in identifying causal variants when all five ancestry groups were equally represented (24,960 per ancestry group). Pooled analysis exhibited the highest statistical power across all ancestry-group compositions particularly when ancestry-group proportions were more balanced. This advantage was most evident in LD-based region detection, where pooled analysis consistently outperformed both meta-analysis and MR-MEGA.

Across all scenarios, pooled analysis demonstrated the highest statistical power, regardless of ancestry-group proportions or total sample size. When only ancestry-group-specific summary statistics were available, MR-MEGA performed best when each ancestry group had at least 20,000 individuals, while meta-analysis remained preferable when some ancestry groups were smaller. The absolute power values for all methods across scenarios are shown in Figure S7, providing a direct view of method-specific performance.

Binary phenotypes

The results for binary phenotypes were consistent with those observed for continuous phenotypes (Table 1). Simulations were conducted using an average of 237,241 cases and 237,241 controls, with detailed sample sizes provided in Table S3. Across all evaluated metrics, pooled analysis exhibited the highest statistical power. When measuring power by the proportion of causal variants reaching genome-wide significance ($p\le 5\times {10}^{-8})$, pooled analysis identified 0.33% of causal variants, compared to 0.17% for meta-analysis and 0.19% for MR-MEGA. Similarly, for LD-based detection, pooled analysis achieved a power of 27.19%, exceeding both meta-analysis (21.25%) and MR-MEGA (19.18%).

Table 1.

Power comparison of multi-ancestry GWAS approaches on simulated binary phenotypes

	No. of samples on average	No. of variants in GWAS output	No. of significant variants	Power 1	Power 2
Pooled analysis	474,482	2,017,677	4,567	0.33%	27.19%
Meta-analysis	474,482	2,024,717	2,375	0.17%	21.25%
MR-MEGA	474,482	1,304,330	2,551	0.19%	19.18%

This table compares the performance of pooled analysis, meta-analysis, and MR-MEGA, under mixed-effect modeling. The simulated dataset comprises 237,241 individuals and 237,241 control individuals from five ancestry groups: African (AFR), admixed American (AMR), East Asian (EAS), European (EUR), and South Asian (SAS) populations. Detailed sample size by ancestry group is provided in Table S4. The number of significant variants is determined using a genome-wide significance threshold of $p\le 5\times {10}^{-8}$. Power 1 represents the percentage of causal variants recovered as GWAS-significant variants. Power 2 represents the percentage of causal variants recovered within a 500-kb region of GWAS-significant variants. The number of significant variants, power 1, and power 2 are averaged over ten replicates. Further details on the simulation setup for phenotype generation can be found in material and methods.

Admixed individuals

To assess statistical power in samples including admixed individuals, we simulated four groups: two reference groups (100% AFR and 100% EUR) and two groups of admixed individuals (50%:50% AFR/EUR and 80%:20% AFR/EUR) using the Admix-kit²⁶ pipeline, each with 60,000 subjects (see material and methods). Across all groups, pooled analysis demonstrated the highest power for both detecting causal SNPs and identifying variants within LD-based causal regions (Table 2). For exact causal variant detection, pooled analysis achieved a power of 1.37%, outperforming both meta-analysis (0.83%) and MR-MEGA (0.91%). For LD-based detection, pooled analysis reached 29.93%, exceeding meta-analysis (24.31%) and MR-MEGA (23.56%).

Table 2.

Power comparison of multi-ancestry GWAS approaches on simulated continuous phenotypes in equally sized samples of admixed individuals

	No. of samples	No. of variants in GWAS output	No. of significant variants	Power 1	Power 2
Pooled analysis	240,000	1,052,000	4,062	1.37%	29.93%
Meta-analysis	240,000	1,052,000	2,633	0.83%	24.31%
MR-MEGA	240,000	1,011,016	3,044	0.91%	23.56%

This table compares the performance of three multi-ancestry GWAS approaches, namely pooled analysis, meta-analysis, and MR-MEGA, under mixed-effect modeling. The dataset comprises one set of 60,000 samples simulated based on 1000G European (EUR) genotypes and one set of 60,000 samples simulated based on 1000G African (AFR) genotypes, one set of 60,000 individuals with 50/50 EUR/AFR admixture, and one set of 60,000 individuals with 20/80 EUR/AFR admixture. Simulations were performed using Admix-kit.²⁶ The number of significant variants was determined using a genome-wide significance threshold of $p\le 5\times {10}^{-8}$. Power 1 represents the percentage of causal variants recovered as GWAS-significant variants. Power 2 represents the percentage of causal variants recovered within 500 kb of GWAS-significant variants. The number of significant variants, power 1, and power 2 are averaged over ten replicates. Further details on the simulation setup for phenotype generation can be found in material and methods.

Real data analysis in AoU and UKB

We conducted a GWAS using mixed-effect modeling with REGENIE and fixed-effect modeling with PLINK2 on real data from AoU and UKB, analyzing eight continuous and five binary traits (see material and methods). Sample sizes for each trait within ancestry groups are provided in Tables S6 and S7, while Manhattan and quantile-quantile (QQ) plots for all GWAS results are shown in Figures S8–S11. To assess inflation in test statistics, we examined the scaled genomic inflation factor (${\lambda }_{1000}$), which adjusts for differences in sample size. Across all analyses, ${\lambda }_{1000}$ ranged from 0.999 to 1.003, indicating that no or little systematic inflation was present. Since the true causal variants in real data are unknown, we measured statistical power by the number of genome-wide significant variants detected ($p\le 5\times {10}^{-8})\text{.}$ Specifically, we reported both the total number of genome-wide significant variants and the number of independent genome-wide significant variants after LD clumping.

Under both mixed- and fixed-effect modeling, pooled analysis consistently identified the highest number of significant variants, particularly for highly polygenic traits such as height. Across both AoU (Tables 3 and S8) and UKB (Tables 4 and S9), pooled analysis outperformed both meta-analysis and MR-MEGA for all traits in both continuous and binary outcomes. On average, across all analyzed traits in UKB, pooled analysis detected 4% and 52% more independent genome-wide significant variants than meta-analysis and MR-MEGA, respectively. The advantage of pooled analysis was even more pronounced in AoU, where it identified, on average across all traits, 34% and 117% more independent significant variants than meta-analysis and MR-MEGA, respectively. This greater advantage of pooled analysis in AoU is driven by its larger proportion of non-European participants (∼41% non-EUR) compared to UKB (∼4% non-EUR). This pattern aligns with our theoretical derivations (see material and methods), which predict that when MAFs vary across ancestries, pooled analysis has a greater advantage over meta-analysis when ancestry proportions are more balanced.

Table 3.

Comparison of significant and independent variants across multi-ancestry GWAS methods in All of Us with mixed-effect modeling

	Height	Waist	LDL	TC	HDL	Calcium	Creatinine	eGFR	Asthma	CAD	T2D	Breast cancer	Prostate cancer
No. of significant variants
pooled analysis	5,435	2,120	354	921	1,258	73	513	37	101	125	268	20	35
meta-analysis	4,668	823	339	550	760	62	73	11	99	124	287	20	35
MR-MEGA	3,695	627	303	475	628	31	25	3	40	120	280	20	22
No. of independent significant variants
pooled analysis	580	316	42	50	101	18	213	1	7	3	28	2	11
meta-analysis	559	80	36	39	87	15	43	1	7	3	24	2	8
MR-MEGA	394	44	22	27	52	6	3	1	4	2	16	2	5

This table compares the number of significant and independent variants identified using three different multi-ancestry GWAS approaches: pooled analysis, meta-analysis, and MR-MEGA. Variants were filtered based on a minor allele frequency (MAF) >0.01 in at least one ancestry group, resulting in approximately 1.2 million variants analyzed. Significant variants were determined using a genome-wide significance threshold of $p\le 5\times {10}^{-8}$. Independent variants are generated with linkage disequilibrium (LD) clumping with an r² threshold of 0.1. The eight continuous phenotypes include height, waist circumference, low-density lipoprotein cholesterol (LDL), total cholesterol (TC), high-density lipoprotein cholesterol (HDL), calcium, creatinine, and estimated glomerular filtration rate (eGFR). The five binary phenotypes include asthma, coronary artery disease (CAD), type 2 diabetes (T2D), breast cancer, and prostate cancer. The number of subjects varies by phenotype, and the sample size for each trait is provided in Table S6.

Table 4.

Comparison of significant and independent variants across multi-ancestry GWAS methods in the UK Biobank with mixed-effect modeling

	Height	Waist	LDL	TC	HDL	Calcium	Creatinine	eGFR	Asthma	CAD	T2D	Breast cancer	Prostate cancer
No. of significant variants
pooled analysis	44,509	5,393	6,580	7,903	10,132	3,758	10,588	14,174	1,010	752	742	207	205
meta-analysis	45,001	5,196	6,532	7,792	9,624	3,750	9,267	12,738	1,039	675	745	211	186
MR-MEGA	33,844	3,629	4,743	5,884	6,871	2,690	7,111	9,054	791	453	509	135	125
No. of independent significant variants
pooled analysis	2,576	351	345	401	604	271	568	743	37	43	52	15	28
meta-analysis	2,574	338	344	399	595	270	542	731	34	41	53	14	25
MR-MEGA	1,851	224	232	266	404	184	448	505	26	24	32	11	15

This table compares the number of significant and independent variants identified using three multi-ancestry GWAS approaches: pooled analysis, meta-analysis, and MR-MEGA. Variants were filtered using a MAF threshold of >0.01 within at least one ancestry group, resulting in approximately 1.53 million variants analyzed. Significant variants were determined using a genome-wide significance threshold of $p\le 5\times {10}^{-8}\text{.}$ Independent variants are generated with linkage disequilibrium (LD) clumping with an $r^2$ threshold of 0.1. The eight continuous phenotypes include height, waist circumference, low-density lipoprotein cholesterol (LDL), total cholesterol (TC), high-density lipoprotein cholesterol (HDL), calcium, creatinine, and estimated glomerular filtration rate (eGFR). The five binary phenotypes include asthma, coronary artery disease (CAD), type 2 diabetes (T2D), breast cancer, and prostate cancer. The number of subjects varies by phenotype, and sample sizes for each trait are provided in Table S7.

Additionally, mixed-effect models consistently outperformed fixed-effect models in detecting genome-wide significant variants. Across all three GWAS approaches, the mixed-effect model identified 8% more independent significant variants than the fixed-effect model in AoU (Tables 3 and S8) and 19% more in UKB (Tables 4 and S9). This increased statistical power of mixed-effect models is likely due to their ability to adjust for both known covariates (top PCs, age, and sex) through fixed effects and additional genomic structure through random effects. By modeling these sources of variation, mixed-effect models reduce residual error variance, thereby increasing statistical power.

Discussion

We evaluated three multi-ancestry GWAS approaches—pooled analysis, meta-analysis, and MR-MEGA—under both fixed- and mixed-effect modeling. Our findings show that pooled analysis consistently provides the highest statistical power across various ancestry-group compositions and trait architectures while maintaining well-controlled type I error in most realistic scenarios. This advantage is particularly pronounced when allele frequencies vary across ancestry groups, supporting our theoretical derivation that pooled analysis benefits from increased minor allele carrier counts across populations.

Through large-scale simulations, we showed that meta-analysis and MR-MEGA are effective but less powerful than pooled analysis, particularly when non-EUR individuals made up a larger proportion of the sample (Figure 4C). Our theoretical framework explains this disparity: as allele-frequency divergence increases, meta-analysis becomes less efficient compared to pooled analysis. Real data analyses from UKB ($n\approx 324,000$) and AoU ($n\approx 207,000$) further validated our findings, with pooled analysis detecting the most independent genome-wide significant variants across both continuous and binary traits (Tables 3, 4, S8, and S9).

Beyond allele-frequency differences, sample-size distribution across ancestry groups plays a key role in determining statistical power. When ancestry groups are more balanced, the power advantage of pooled analysis over meta-analysis and MR-MEGA is maximized. This is particularly evident in AoU, where a higher proportion of non-EUR participants amplifies the advantage of pooled analysis, further reinforcing our theoretical predictions.

Our findings also replicate previous findings showing the higher power of mixed-effect models over fixed-effect models³⁴ (Tables 3, 4, S8, and S9). By incorporating random effects, mixed-effect models account for genome-wide polygenic effects, which helps reduce residual error variance and improve the detection of true associations. Additionally, they adjust for cryptic relatedness and population structure, enhancing robustness while maintaining proper type I error control. These results support the continued adoption of efficient mixed-effect frameworks such as REGENIE for large-scale GWAS analyses.

Historically, fixed-effect meta-analysis has been the standard approach for multi-ancestry GWASs due to its ability to combine summary statistics from independent studies without requiring individual-level data.^{8,10,11,35,36} However, our results suggest that pooled analysis is preferable when individual-level data are available, as it avoids binning individuals into discrete ancestry groups and leverages shared genetic architecture across populations. MR-MEGA, which models allele-frequency differences, performed slightly better than standard meta-analysis when non-EUR sample sizes were sufficiently large, but it still lagged behind pooled analysis in most scenarios.

Both pooled and fixed-effect meta-analysis aim to estimate a shared genetic effect across populations. Our theoretical framework (see Note S1) shows that pooled analysis retains a statistical power advantage even when ancestry-specific effect heterogeneity is present, provided that the ancestry-specific effects are independently drawn from a common, arbitrary distribution. In our simulations, effect sizes were assigned on the standardized genotype scale, leading to MAF-dependent variation in per-allele effects across ancestries. This introduces realistic marginal effect heterogeneity, which our theoretical model accommodates. Additionally, our simulations incorporate realistic LD patterns via 1000G haplotypes, further contributing to marginal effect variability across populations. In all these scenarios, pooled analysis consistently outperformed meta-analysis.

While recent studies suggest that most genetic effects are broadly shared across populations,^37,38 limited sample sizes in non-European groups hinder our ability to comprehensively assess heterogeneity in marginal effects. In situations where variant effects are strongly population specific, both pooled and fixed-effect meta-analysis may be suboptimal compared to ancestry-stratified GWASs. Therefore, we recommend a practical dual strategy: using pooled analysis to enhance discovery across ancestries while also conducting ancestry-specific analyses to identify potential population-specific signals.^39,40,41

The observed power differences have important implications for PRS development and fine-mapping efforts in multi-ancestry cohorts. Pooled analysis may improve PRS portability by increasing statistical power for variant discovery, ultimately enhancing risk prediction models in different populations.^39,40 Additionally, the superior power of pooled analysis suggests that functional fine-mapping and transcriptome-wide association studies may benefit from using pooled GWAS summary statistics, as these analyses rely heavily on accurate effect-size estimates.⁴¹ Another key implication is for admixture mapping studies. Our findings indicate that pooled analysis provides better power than meta-analysis when analyzing admixed individuals, supporting its use in large, diverse cohorts such as AoU and the Million Veteran Program.⁴² While meta-analysis remains necessary when individual-level data are unavailable, recent studies have begun to explore hybrid approaches that combine pooled analysis with ancestry-aware meta-analysis.^43,44 Future work could continue to build on these efforts to enhance cross-ancestry analysis in GWASs and PRSs.

Our study has several limitations. First, pooled analysis requires individual-level genotype data, which are often restricted due to privacy and regulatory constraints. This highlights the need for more powerful meta-analytical methods. Second, our simulations assume homogeneous effects on the standardized scale, which may not fully reflect gene-environment interactions or population-specific selection. Future work should explore adaptive models that balance power with flexibility in modeling heterogeneity. Third, while pooled analysis controls type I error in most settings, caution is warranted when ancestry-specific PCs explain a large share of trait variance (>1%), which may lead to spurious associations. We recommend evaluating this variance prior to pooled GWASs, especially for traits with strong structure. Fourth, our simulations assume Hardy-Weinberg equilibrium (HWE), including for admixed individuals. While this assumption simplifies model construction and does not artificially inflate power, it may not fully capture biological or demographic deviations from HWE in real-world multi-ancestry data. In practice, SNPs that are highly differentiated across populations, those contributing most to pooled analysis power, are also more likely to appear out of HWE when tested in a combined sample. This creates a risk that standard pooled HWE filtering may remove precisely the variants most informative for cross-ancestry analysis. Recent work has shown that such deviations often reflect ancestry heterogeneity rather than genotyping error or true selection.⁴⁵ To mitigate this, we recommend performing HWE filtering within ancestry groups rather than in pooled samples, consistent with current best practices in large diverse cohorts. Alternatively, multi-ancestry-aware HWE tests, such as the method proposed by Shyr et al.,⁴⁵ could offer a principled solution in diverse cohorts.

In summary, our study demonstrates that pooled analysis provides the highest statistical power for multi-ancestry GWASs, particularly when allele frequencies vary across populations. While meta-analysis remains a valuable approach when individual-level data are unavailable, pooled analysis should be prioritized when feasible. Our findings emphasize the importance of balanced ancestry representation and highlight the advantages of mixed-effect modeling for large-scale genetic studies. Future research should focus on developing hybrid approaches that integrate the strengths of pooled analysis and meta-analysis, ensuring that multi-ancestry GWASs continue to uncover novel genetic insights across diverse populations.

Data and code availability

The code used to conduct the simulations and data analyses is available at the GitHub repository multiancestryGWAS (https://github.com/juliealexia/multiancestryGWAS).

Acknowledgments

The analysis utilized the high-performance computation Faculty of Arts and Sciences Research Computing Cluster at Harvard University, Biowulf cluster at National Institutes of Health, Polaris cluster at Argonne National Laboratories, and All of Us Researcher Workbench. We gratefully acknowledge All of Us participants for their contributions, without whom this research would not have been possible. We also thank the National Institutes of Health’s All of Us Research Program for making available the participant data examined in this study. We acknowledge the UK Biobank, a major biomedical database, along with their participants. This research has been conducted using the UK Biobank resource under application number 52008. This research was supported in part by the Intramural Research Program of the National Institutes of Health (NIH). The contributions of the NIH authors (H.X., X.W., P.K., and H.Z.) were made as part of their official duties as NIH federal employees, are in compliance with agency policy requirements, and are considered Works of the United States Government. However, the findings and conclusions presented in this paper are those of the authors and do not necessarily reflect the views of the NIH or the US Department of Health and Human Services. This research was funded by NIH research grant U01CA261339 (J.-A.D.), NIH training grant T32GM135117, and NSF Graduate Research Fellowship DGE-2140743 (T.C.). This research used resources of the Argonne Leadership Computing Facility, a U.S. Department of Energy (DOE) Office of Science user facility at Argonne National Laboratory and is based on research supported by the U.S. DOE Office of Science-Advanced Scientific Computing Research Program, under Contract No. DE-AC02-06CH11357.

Author contributions

J.-A.D., P.K., and H.Z. conceived the project. J.-A.D. carried out all data analyses and derivations, simulated admixed genotypes, ran GWASs on AoU data using the All of Us research workbench, and ran GWASs on UKB data with supervision from P.K. and H.Z. T.C. advised on best practices when running cluster analyses and preprocessed genotype and phenotype data for UKB analyses. J.-A.D. and H.X. preprocessed the AoU data. X.W. advised on methods for binary analyses. R.K.M. and A.A.R. provided computing resources from Argonne National Laboratory to simulate genotypes and phenotypes. J.-A.D., P.K., and H.Z. drafted the manuscript. T.C., H.X., X.W., A.A.R., and R.K.M. provided comments. All authors reviewed and approved the final version of the manuscript.

Declaration of interests

The authors declare no competing interests.

Published: September 2, 2025

Web resources

• •Admix-kit, https://github.com/KangchengHou/admix-kit
• •MR-MEGA, https://tools.gi.ut.ee/tools/MR-MEGA_v0.2.zip
• •PLINK2, https://www.cog-genomics.org/plink/2.0/
• •REGENIE, https://rgcgithub.github.io/regenie/install