Why Ancestry Representation Is a Design Decision
A genome-wide association study (GWAS) can recruit a large sample and still provide limited evidence about people who were sparsely represented. Recruitment composition affects which groups are directly studied, which allele-frequency and linkage-disequilibrium patterns the data capture, and how confidently results can be applied beyond the enrolled sample. Ancestry representation is therefore a design decision, not merely a demographic description added after data collection.
The decision requires care because ancestry is not a set of naturally separate, uniform categories. Genetic ancestry reflects shared patterns of genetic variation and population history; it can be continuous, multidimensional, and shaped by admixture. Self-identified race or ethnicity captures important social and personal information, but it is not interchangeable with genetic ancestry. Recruitment plans may need to consider both, while stating clearly how each is defined and used.
Specify Targets Before Recruitment
A useful plan connects targets to the study’s intended population and scientific goals. It should explain whether the aim is to approximate a target population’s composition, to ensure meaningful representation across groups, or to recruit additional participants from groups historically underrepresented in genomic research. Those aims can call for different allocations. A sample that mirrors the composition of a local health system, for example, need not have the same composition as one designed to support group-specific analyses.
For each target, specify the number of participants or the percentage, the denominator, how group membership will be established, and how recruitment progress will be monitored. If recruitment is across several sites, specify whether targets apply overall, within each site, or both. Site-level planning matters because a group recruited at only one site may also differ in healthcare access, measurement procedures, or other features of the study. A single overall percentage can conceal such concentration.
Use categories that are meaningful and feasible for the intended setting, and avoid treating broad labels as internally uniform. Where appropriate, collect self-identified ancestry or race and ethnicity separately from genetic data used to characterize ancestry. Explain how participants can report multiple identities, how missing or uncertain responses are handled, and whether categories can be revised as recruitment proceeds. Do not infer a participant’s identity from genetic data or assume that one reported category identifies a single genetic background.
Targets are planning goals, not guarantees of adequate sample size for every analysis. A group may be represented in the study but still have too few participants, too few copies of a particular allele, or insufficient data quality for a particular association analysis. Later power calculations can quantify detection ability; the recruitment plan should first make the intended composition and its limitations explicit.
Worked Example: Turning Representation Goals Into Recruitment Numbers
A hypothetical GWAS plans to enroll 6,000 participants across four operational recruitment strata. For planning purposes only, the team sets targets of 40%, 30%, 20%, and 10%. These labels are locally defined recruitment categories, not claims that each category is genetically uniform. Find the target count in each stratum.
For a target proportion \(q_g\), the target count is \(N_g=6{,}000q_g\). The calculations are:
The counts sum to \(2{,}400+1{,}800+1{,}200+600=6{,}000\), and the proportions sum to 1. Suppose an interim projection is 3,000, 1,500, 1,080, and 420 participants. Subtracting projected counts from targets gives gaps of \(-600\), \(+300\), \(+120\), and \(+180\). The first stratum is projected to exceed its target by 600; the other three fall short by the stated amounts. The team can use these gaps to plan outreach, while also checking whether each site and the overall study meet their separate goals.
These calculations describe composition, not statistical power or representativeness of every subgroup within a broad stratum. The team should report how the categories were defined and should not interpret the resulting allocation as an estimate of the composition of all people in the target population unless the sampling design supports that interpretation.
Choose Reference Resources to Match the Design
Reference resources can inform which variants are likely to be observed or imputed well, how genetic ancestry is characterized, and whether a study’s data can be compared with existing datasets. They are not substitutes for recruiting the people whose health and genetic variation the study aims to understand. A reference panel’s value depends on its participants, sequencing or genotyping data, variant content, and the specific task for which it is used.
For imputation, a reference panel should contain relevant genetic variation and provide useful haplotype information for the study sample. The 1000 Genomes Project provides a widely used resource with participants sampled from multiple world regions, though the sample sizes within particular populations are limited. The Haplotype Reference Consortium has been used as a large imputation reference but is predominantly of European ancestry. TOPMed includes diverse whole-genome sequencing data and can provide a useful reference for some study designs. These descriptions do not mean that any one panel is optimal for every participant or variant: evaluate coverage and data compatibility for the planned sample and variant set.
For ancestry characterization, reference individuals can help orient comparisons, but the resulting estimates depend on which populations and variants are represented in the reference. If a participant’s ancestry is not well represented, the estimate can be imprecise or misleading. Record which resources and procedures are used, and interpret ancestry estimates in light of those limitations. A reference label should not be treated as a definitive identity or as a complete description of an individual’s ancestry.
Worked Example: Matching a Reference Resource to a Recruitment Plan
A hypothetical study expects to recruit participants from several regions, including groups with limited representation in the study team’s existing imputation panel. The team is considering continuing with that panel or evaluating a resource such as TOPMed. The choice must be made from evidence about the study data and the candidate resources, not from a resource’s name alone.
Step 1: Define the task. The immediate task is imputation of variants in the planned study sample. The team lists the array content, variant set, expected ancestry composition, and the quality metrics it will use to compare imputation results.
Step 2: Check fit. The team compares how well each candidate resource represents the planned sample and how well it covers variants relevant to the study. It also checks that the study data and reference data can be processed compatibly. A panel with many participants but weak representation of a study group may not perform uniformly across the sample.
Step 3: Make a conditional decision. If evaluation shows that the existing panel provides weaker coverage for some planned groups, the team tests an alternative or combined strategy if technically appropriate. It documents which panel was used and reports group-specific quality summaries. If the evidence does not show a meaningful improvement, the team does not claim that a larger or more diverse-sounding resource automatically solves the coverage problem.
No numerical performance result is assumed here: the decision requires actual validation against the study’s data and reference resources. The example illustrates why matching is an empirical design check rather than a simple rule that one panel fits everyone.
What Uneven Representation Means for Discovery
Allele frequencies and linkage-disequilibrium patterns can differ among populations and ancestry backgrounds. A variant that is common in one group may be uncommon in another, and a typed or imputed marker may tag a causal variant differently across groups. Consequently, the same sample size in two groups does not necessarily provide the same information about a particular variant. Conversely, a larger group does not ensure that every association is well measured or that the results describe smaller groups.
A simple calculation illustrates one component of this issue. Suppose, for illustration, that a biallelic genotype is coded as the number of copies of an allele, and that the two allele copies are independent with allele frequency \(p_g\) in group \(g\). Under Hardy–Weinberg equilibrium, the dosage variance is \(2p_g(1-p_g)\). With \(n_g\) unrelated participants and a quantitative outcome having residual variance \(\sigma^2\), a rough measure of the genotype information for an additive association is proportional to \(n_g2p_g(1-p_g)/\sigma^2\), under a simple model with comparable residual variance and no additional complications. This is an information comparison, not a complete power calculation.
The equality uses independent Bernoulli allele indicators \(A_1,A_2\), each with probability \(p_g\) of carrying the counted allele. Under the stated simplified association model, information about a regression slope increases with the total genotype variation observed. This explains why sample size alone is not a complete description of the information available for a particular variant.
Worked Example: Comparing Genotype Variation Across Two Groups
Consider a hypothetical biallelic variant in two unrelated study strata. In stratum A, \(n_A=2{,}400\) and the counted-allele frequency is \(p_A=0.30\). In stratum B, \(n_B=600\) and \(p_B=0.05\). Assume Hardy–Weinberg equilibrium within each stratum and use \(n_g2p_g(1-p_g)\) only as a rough genotype-variation measure.
For stratum A:
For stratum B:
The ratio is \(1{,}008/57\approx17.7\). Under the simplified assumptions, stratum A contributes much more genotype variation for this variant. This comparison does not show that a true effect exists, does not calculate a p-value, and does not establish the power of either stratum. It shows why both the number recruited and the allele frequency matter when describing the information a study may provide. The frequency, imputation quality, effect, phenotype variance, and analysis model all matter to a full power assessment.
Portability, Generalizability, and Pooled Analyses
A discovery in a heavily represented group may not transfer unchanged to another group. Differences in allele frequency and linkage disequilibrium can alter how well a marker tags an underlying causal variant. Effect estimates can also vary because of genuine effect heterogeneity, differences in phenotype measurement, environmental or social context, or study procedures. These possibilities should be investigated rather than assumed to be absent or present. Evidence from one ancestry background is not automatically evidence of the same predictive performance in all populations.
Pooling participants can be appropriate, but a pooled estimate is tied to the study’s composition and analysis model. If the study combines groups with different allele frequencies or effect estimates, the pooled coefficient need not equal the coefficient in any one group. Genetic-ancestry covariates and other appropriate adjustments can help address some structure-related confounding, but they cannot create observations in groups that were not recruited, guarantee equal measurement quality, or ensure portability. Group-specific estimates and uncertainty can clarify what the data support, provided the sample contains enough information for those estimates.
Purposeful oversampling can improve the amount of data available from groups that would otherwise be sparsely represented. It also changes the sample composition relative to a population-based sample. If the goal is to estimate a population-level quantity, the design and analysis may need to account for the sampling scheme; if the goal is association discovery within strata, the relevant group-specific estimates and their limits should be reported. Recruitment targets and inferential goals should be planned together.
Worked Example: Describing the Reach of a Pooled Result
A hypothetical study enrolls 4,000 participants: 3,200 from one recruitment stratum and 800 combined across three other strata. It reports a pooled association estimate from a model that adjusts for measured genetic ancestry. The team wants to describe what the result establishes.
The pooled estimate summarizes the association in the analyzed sample under the specified model. It does not, by itself, establish that the effect has the same size in all four strata. The adjustment does not change the fact that 800 participants are divided among three strata, nor does it guarantee precise stratum-specific estimates.
A responsible report states the recruitment composition, gives the pooled estimate and its uncertainty, and reports stratum-specific results where they can be estimated reliably. It describes transfer to other populations as a question requiring evidence, not as a consequence of including an ancestry covariate. If stratum-specific estimates are imprecise, that limitation should be stated rather than interpreting a non-significant result as proof of no effect.
Common Mistakes in Representation Planning
- Treating ancestry labels as biological boxes. A category used to manage recruitment is an operational definition, not a claim that everyone in it shares the same ancestry or genetic variation.
- Reporting percentages without a denominator. A target of 20% is unclear unless the plan says 20% of which participants, at which sites, and according to what group definition.
- Equating representation with analytic adequacy. Enrollment in a group does not guarantee sufficient information for every variant or subgroup analysis.
- Assuming covariate adjustment repairs recruitment gaps. Adjustment may address some confounding from population structure, but cannot replace participants or establish portability.
- Choosing a reference by size alone. A large panel can still be a poor match for a study group or variant set; fit should be evaluated for the intended task.
Check Your Understanding
Try these questions before reviewing the tutorial.
- A study plans to enroll 5,000 participants and sets a target of 18% for a defined recruitment stratum. What is the target count, and what additional information must the plan report to make the target interpretable?
- Why are self-identified race or ethnicity and genetic ancestry not interchangeable measures?
- Under the simplified Hardy–Weinberg assumptions, compare the dosage variance for allele frequencies \(p=0.50\) and \(p=0.10\). What does the comparison say, and what does it not say?
- Why might a reference panel with many participants still be a poor match for part of a planned GWAS sample?
- What can adjustment for genetic ancestry contribute to a pooled association analysis, and what recruitment limitation can it not remove?