The core promise of genomic medicine is its ability to predict, diagnose, and treat diseases based on an individual's DNA. However, the mathematical foundation of modern genomics relies heavily on statistical power—the probability that a study will detect a true genetic association if it exists. Statistical power is fundamentally driven by two variables: sample size ($N$) and the frequency of the genetic variant within the study population.
When global genomic databases are overwhelmingly dominated by individuals of European descent, a severe statistical power imbalance occurs. Underrepresented populations (such as African, Hispanic/Latino, and Asian cohorts) are relegated to much smaller sample sizes within biobanks and clinical trials. This disparity does not simply mean we know less about these groups; it fundamentally distorts our understanding of disease biology, creates severe discovery biases, and causes predictive models to fail when applied to non-European patients.
The Mathematics of the Representation Gap
To understand why sample size disparities break genetic discovery, consider how a Genome-Wide Association Study (GWAS) identifies a disease-linked variant. Because a GWAS tests millions of genetic markers simultaneously, it requires an exceptionally high threshold for statistical significance (typically $p < 5 \times 10^{-8}$) to eliminate false positives.
If a cohort's sample size is small, the statistical "noise" remains high. As a result, only genetic variants with massive effect sizes can break through the threshold. As noted by Agarwal et al. and Raptis et al., insufficient sample sizes for non-European populations severely limit the power to identify ancestry-specific risk factors. When a dataset is heavily skewed, independent genome-wide analyses within minority subgroups become statistically unviable, as highlighted by Ding et al..
This sample size deficit creates a vicious cycle. When researchers stratify a diverse cohort into smaller ancestral subpopulations to avoid confounding effects, the statistical power drops even further, as shown by Rappoport et al.. Consequently, rare variants or subtle polygenic signals unique to underrepresented groups remain buried in the noise, a limitation documented across diverse clinical traits ranging from cognitive decline (Aung et al.) to adverse pregnancy outcomes (Khan et al.; Yan et al.).
Discovery Bias and Hidden Architecture
The statistical power imbalance introduces a profound discovery bias. If a genetic variant is rare in European populations but common in African or Hispanic populations, it will likely be missed by large, European-centric meta-analyses. Conversely, variants that are common in Europeans receive disproportionate attention, leading to a biased architecture of global disease risk.
Yang et al. demonstrate that these sample size disparities result in a discovery bias that systematically favors the heavily sampled population, complicating cross-ancestry comparisons. This means that regulatory mechanisms unique to underrepresented ancestries are frequently obscured, as argued by Liu et al..
When researchers actively build cohorts for underrepresented groups, the biological rewards are immediate. For instance, by analyzing a dedicated cohort of individuals of African ancestry, Ng et al. successfully identified novel genetic loci associated with adiposity that had been missed in larger European studies. Similarly, Petty et al. used a large Hispanic/Latino multi-omics cohort to uncover entirely novel genetic signals for type 2 diabetes and blood lipids. Without these population-specific sample sizes, low-frequency and rare variants remain statistically invisible (Chen et al.).
The Portability Crisis in Predictive Medicine
The practical consequence of this power imbalance is the failure of Polygenic Risk Scores (PRS) to "port" across ancestries. A PRS aggregates the effects of thousands of variants to predict an individual's genetic liability for a disease. However, because these scores are trained on massive European datasets, they reflect the specific genetic background—including Linkage Disequilibrium (LD) blocks and allele frequencies—of European populations.
When a PRS optimized for Europeans is applied to a minority population, its predictive accuracy plummets. Moreno-Grau et al. show that the limited representation of non-European individuals in existing biobanks directly reduces the statistical power available to develop and refine ancestry-specific risk models. This data scarcity prevents predictive tools from reaching clinical utility for minority groups, a phenomenon also observed by Venkateswaran et al. in behavioral and systemic health risks.
Furthermore, attempting to replicate novel variants across highly disparate sample sizes introduces statistical anomalies. As explained by Hu et al., genetic associations identified in diverse but smaller cohorts often fail to reach formal significance in replication studies due to the "winner's curse" and the massive sample size disparity between minority and majority ancestry groups.
Resolution Failures in Fine-Mapping and Heterogeneity
Beyond discovering whether a gene is linked to a disease, researchers use cross-ancestry data for fine-mapping—pinpointing the exact causal mutation out of a cluster of highly correlated variants. African populations are uniquely valuable for fine-mapping because they possess shorter LD blocks due to a longer evolutionary history of genetic recombination.
However, the power imbalance undercuts this biological advantage. Gharahkhani et al. demonstrated that when African GWAS datasets have low statistical power relative to European datasets, their mathematical influence on cross-ancestry fine-mapping frameworks is severely diminished.
[European Cohort: Massive Sample Size] ----> High Statistical Weight \
----> Fine-Mapping Result Biased
[African Cohort: Small Sample Size] ----> Low Statistical Weight / Toward European LD Structure
This imbalance also obscures heterogeneous effects, where a single genetic variant behaves differently depending on ancestral background. Detecting an interaction effect (ancestry $\times$ variant) requires significantly more statistical power than detecting a main effect. As a result, even large biobanks that include minority subgroups often lack the power to resolve fine-grained variations in genetic associations, as detailed by Wang et al.. This leaves clinicians blind to instances where a drug target or risk variant might be highly effective or uniquely dangerous in a specific subpopulation.
Statistical and Algorithmic Remedies
Because balancing global biobank sample sizes through enrollment alone will take decades, researchers are developing statistical frameworks to maximize the power of existing diverse data.
Multi-Ancestry Pooling and Meta-Analysis
Rather than analyzing underrepresented groups in isolation, frameworks are shifting toward multi-ancestry pooling. By combining diverse cohorts, researchers can maintain statistical power while adjusting for genetically inferred ancestry, helping to uncover shared genetic architectures while preserving population-specific signals (Budu-Aggrey et al.; Johnson et al.).
Leveraging Local Genetic Structure
Instead of treating ancestry as a global covariate, advanced frameworks like LOG-TRAM, developed by Xiao et al., leverage massive, well-studied European datasets as auxiliary data. By focusing on local genomic structures rather than global averages, these methods transfer statistical power from large cohorts to smaller, underrepresented target populations without introducing ancestry-specific biases.
Heterogeneity-Aware Discovery Frameworks
To detect subtle population-specific effects without losing power to strict multiple-testing corrections, the authors developed frameworks like hetFDR (Wang et al.). This approach models the sample size imbalances directly, allowing researchers to identify ancestry-specific variations across thousands of clinical traits simultaneously.
Ultimately, solving the statistical power imbalance requires a dual approach: aggressively expanding the ancestral diversity of physical biobanks while deploying mathematical frameworks designed to extract robust, equitable insights from uneven data.
Key Foundational Work
Association between Regulated upon Activation, Normal T Cells Expressed and Secreted (RANTES) -28C/G Polymorphism and Susceptibility to HIV-1 Infection: A Meta-Analysis (2013) — contributed 1 concept elements to this direction.
Bayesian identification of admixture events using multilocus molecular markers (2006) — contributed 1 concept elements to this direction.