Ancestry & Equity

Who lives in Loudoun

2023 ACS 5-year estimates (2019-2023), Loudoun County, Virginia. Retrieved 2026-10-06 from the Census Bureau API. Labels below are the Bureau’s own, verbatim.

Detailed Asian groups

Table B02015, Asian Alone by Selected Groups. This is a detailed table and publishes no percentage, so none is shown and none is derived here.

Who GWAS has enrolled

Share of genome-wide association study participants by ancestry group, from a single 2024-09 snapshot. Source: GWAS Diversity Monitor (Mills & Rahal^[mills2020]. Denominator: GWAS Diversity Monitor live dashboard, September 2024.

Two cautions on this chart. The snapshot itemises only these groups and does not break out the remainder, which is therefore not plotted rather than being zero. And “Asian” here is not South Asian-specific: it combines East, South-East, South and Central Asian ancestries, so it cannot be read as a figure for South Asian participation.

South Asian participation, itemised

Because the snapshot above does not separate South Asian ancestry, the table below gives the one verified source that does. It is a different source with a different denominator, so it is shown as a table rather than merged into the chart: the two sets of percentages are not interchangeable and must not be added together.

Fatumo et al. 2022, Nature Medicine^[fatumo2022]. One earlier survey of the first decade of polygenic scoring studies found 67% included only European-ancestry participants, and 3.8% were conducted among African, Hispanic or Indigenous cohorts; it noted South Asian samples had been included in very few such studies ^[duncan2019].

Why accuracy drops across ancestry groups

Five quantities account for most of the measured loss, and the literature treats them separately. They are not interchangeable, and a score can be affected by several at once.

1. Discovery sample size. Score performance tracks the size of the genome-wide association study it was built from. In one multi-ancestry type 2 diabetes evaluation, single-ancestry score performance correlated positively with GWAS sample size across all five ancestry groups ^{[huertachagoya2026]}, and in a large type 2 diabetes meta-analysis the groups with the smallest effective sample sizes produced the weakest ancestry-specific scores^{[mahajan2022]}.
2. Linkage-disequilibrium differences. A score carries tag variants rather than the underlying functional ones. How well a tag stands in for its neighbour depends on local haplotype structure, which differs between populations ^[wang2020ra]. Earlier work named the handling of linkage disequilibrium as a methodological requirement for applying scores to non-European cohorts^[duncan2019].
3. Allele-frequency differences. A variant common enough to detect in one population can be too rare to detect in another, and ancestry-specific variants tend to be rarer^{[mahajan2022]}. Averaged across traits, allele-frequency and linkage-disequilibrium differences together account for about 86% of the accuracy loss in African ancestry and about 37% in South Asian ancestry^[wang2020ra], which leaves a substantial remainder in both.
4. Effect heterogeneity. The cross-population correlation of effect sizes at the underlying functional variants, and the ratio of heritability between populations, both enter the accuracy ratio directly ^[wang2020ra]. For type 2 diabetes and asthma, association signals themselves appear mostly shared across groups ^{[mahajan2022][tsuo2022]}, so for those traits this term is smaller than the tagging and frequency terms.
5. Environment and gene-environment interaction. Reviews of transferability name modifiable risk factors alongside genetic ones ^{[kachuri2024]}. Asthma score accuracy differed across biobanks in part through disease prevalence, environmental exposures and phenotype definition ^[tsuo2022]. In blood-pressure score evaluations, stratified performance patterns were clearer in some self-reported background groups than others, which the authors read as a sign of environmental and gene-environment contributions^{[kurniansyah2023]}. Environment is not the whole story, though: when training and testing are confined to a single cohort, reducing environmental and genotyping confounding, accuracy still falls with genetic distance ^[prive2022].

Underneath all five, accuracy varies continuously with genetic distance from the training data rather than stepping between labelled groups. Across 84 traits, genetic distance correlated with score accuracy at −0.95^[ding2023].

Disease modules

One module per disease. The headline figure below is each module’s South Asian-ancestry row, with its metric named. Metrics differ between modules and are not comparable across cards: each number is only meaningful against the comparator given beside it.

Hypertension and breast cancer were evaluated as a fourth module and excluded. Neither has a verified study reporting a transferability metric for a South Asian-ancestry group: the available blood-pressure evaluation groups participants by self-reported race and ethnicity, and its “Asian” category is not a South Asian-ancestry grouping^{[kurniansyah2023]}. Shipping it would have meant treating a self-identification category as an ancestry group, which the policy below forbids.

One score, one metric, five groups

Every bar below is the same metric, the same trait (type 2 diabetes), the same score and the same study, which is what makes them comparable with each other. Metric: incremental AUC of the multi-ancestry PRS, the gain in area under the curve from adding the score to a model of age, sex and principal components. Higher is more accurate. Source: Huerta-Chagoya et al. 2026, The Lancet Diabetes & Endocrinology^{[huertachagoya2026]}.

Figure data

One honest complication, from the same paper: this multi-ancestry score significantly outperformed the best previously published score for the African or African American, Admixed American and European groups, but not for the East Asian or South Asian groups, where the best previously published South Asian score reached a slightly higher incremental AUC of 0.068 ^{[huertachagoya2026]}. A newer score is not automatically a better one for every group.

Mechanism brief

Methods and definitions

Relative accuracy

The relative accuracy of a polygenic score in an ancestry-divergent target population is the ratio of its prediction accuracy in that population to its prediction accuracy in a population of the same ancestry as the discovery sample: RA = R²₂ / R²₁. Loss of accuracy is (1 − RA) × 100%. Relative accuracy decomposes into the squared cross-population correlation of effect sizes at the underlying functional variants, the ratio of heritability between populations, and terms for allele-frequency and linkage-disequilibrium differences^[wang2020ra]. The definition and the decomposition are that paper’s, not GENARCH’s.

Other metrics appear on these pages because the source studies used them: incremental AUC, variance explained on the liability scale, odds ratio per standard deviation, pseudo R², and relative effect size indexed to a reference group. These answer different questions and do not share a scale. Every value on every page is printed with its metric name beside it for that reason, and no chart here places two metrics on one axis.

Population descriptors

• “South Asian ancestry”, “African ancestry”, “European ancestry” refer to genetic-ancestry groupings as defined in the studies cited.
• “Black”, “Hispanic or Latino”, “Asian Indian” are Census self-identification categories and appear only beside Census figures.
• Hispanic and Latino populations are admixed and sit on a continuum of genetic ancestry rather than forming a discrete group^{[ding2023][ge2022]}. Where a study’s own label spans both framings, GENARCH records the study’s label verbatim in a “study labels” field beside its own, so the two can always be told apart.
• A Census category is never equated with a genetic-ancestry group in the same sentence without this caveat. The National Academies published a 2023 consensus study report on population descriptors in this field^[nasem2023].

Verification of every number

Each reference behind these pages was checked against Crossref by DOI, with a normalised title match required, and each number was read from text actually retrieved: the Crossref abstract, the Europe PMC abstract, or the Europe PMC open-access full text. A page, table, or figure locator is recorded alongside every value in pipeline/sources/cross_ancestry/literature.json, and the pipeline validator refuses to build a module whose value does not match its recorded extraction. Where a figure could not be located in retrieved text, it is omitted and the gap is stated rather than filled from another study. Loudoun population figures come from the Census Bureau API, with variables selected by matching the Bureau’s own label text rather than by variable code.

What GENARCH does not do

• No polygenic score is computed, stored, or displayed for any person. There is no input on this site that accepts genetic data.
• No individual-level statement is made or implied. Every figure is a group-level statistic from a published study population.
• No clinical recommendation is made. GENARCH does not state whether a polygenic score should or should not be used in care; it reports what studies measured.
• No new analysis was performed. GENARCH ran no polygenic scoring and no reanalysis of any cohort, and claims no finding of its own: it synthesises published work.
• No causation is asserted. These are measurements of how a statistical model transfers between study populations.