Designing Ancestry-Diverse Association Studies

Designing for an Answerable Cross-Population Question

An ancestry-diverse association study is useful only if its design supports the comparison it intends to make. Recruiting participants from multiple populations is a starting point, not a guarantee that effects can be compared: phenotype definitions may differ, genotype information may be uneven, covariates may be measured inconsistently, and some groups may contribute too little information for precise estimates. The study plan should therefore state which associations will be estimated, which differences will be tested, and what evidence will count as replication.

The target should be an association estimand, defined on a stated effect scale and conditional on a specified set of covariates. This is not automatically a causal effect. Earlier material on association effect scales and replication emphasizes that effect estimates can differ because of scale, allele frequency, linkage disequilibrium, measurement, or underlying effects. Study design should collect enough information to assess these possibilities without treating population labels as biological mechanisms.

Before enrollment, specify the primary phenotype, the association model and effect scale, the variant or region strategy, and the intended comparisons. For each analysis population or ancestry stratum, plan to report the number recruited, the number passing quality control, the phenotype distribution, genotype availability, and the ancestry and covariate measurements used. These details make differences in precision and interpretation visible rather than hiding them behind a single pooled result.

Representation, Measurement, and Information

Representation targets should be set in terms of the intended analysis, not only a desired fraction of the overall cohort. A stratum can have many participants but little usable information at a particular variant if genotype variation is scarce, missingness is high, or the genotype is nearly predicted by adjusted covariates. Conversely, a smaller stratum may be informative for some variants. When feasible, use pilot data to estimate genotype availability, phenotype variance, covariate distributions, and the genotype variation remaining after covariate adjustment.

Plan to measure global ancestry consistently across the cohort and to retain the underlying continuous estimates or ancestry coordinates when they are relevant to the model. For admixed participants, local ancestry at the analyzed locus can support a distinct question about ancestry-specific genomic backgrounds. It requires suitable reference information, a stated inference method, and quality checks; uncertain local ancestry calls should not be treated as error-free. Decide in advance how uncertain calls will be handled, for example through a validated dosage or posterior-based procedure, rather than selecting a rule after seeing the association results.

Covariates should be chosen for the scientific question and data-generating context. Common planning considerations include age, sex-related variables when relevant to the phenotype, study site, genotyping batch, and ancestry summaries used to address population structure. Measure them with consistent definitions across recruitment groups. Avoid assuming that the same variable has the same meaning or measurement quality at every site. Also record missingness and the timing or source of measurements, since complete-case analysis can reduce effective sample size and can change the analyzed cohort.

For a quantitative outcome under a linear model, a simple planning approximation illustrates why marginal genotype variance is not enough when covariates are present. Suppose independent strata \(j=1,2\) have additive genotype dosage \(G\), covariates \(C\), and common residual variance \(\sigma^2\). Within each stratum, residualize \(G\) on the covariate design matrix, including an intercept, and call the residuals \(G_{\mathrm{res}}\). If \(V_{\mathrm{res},j}\) is their mean squared value, then the genotype information is approximately \(n_jV_{\mathrm{res},j}/\sigma^2\). Thus the standard error for the difference between independent stratum-specific additive estimates is approximately:

This approximation assumes a correctly specified linear mean model, independent observations within and across strata, homoskedastic residual errors with variance \(\sigma^2\), and a full-rank design. It also treats the residualized genotype variance as adequately estimated for planning. If there is relatedness, heteroskedasticity, complex sampling, or a binary outcome, use a design-appropriate information calculation or simulation instead. The displayed formula is not a universal sample-size rule.

Choose Stratified Estimates and Interaction Tests Together

A design should make clear whether the primary objective is to estimate associations separately, test whether effects differ, or both. Stratified estimates show the direction and magnitude in each analysis group and expose differences in precision. A pooled interaction model can test a specified contrast while using the individual-level data. These are complementary summaries: a nonsignificant interaction does not establish that effects are equal, and significance in one stratum but not another is not itself a test of effect difference.

For two groups, a simple quantitative-trait model uses \(A=1\) for group 2 and \(A=0\) for group 1:

Here \(G\) is the effect-allele dosage, \(C\) is the prespecified covariate vector, and \(\epsilon\) is the model error. Under this coding, \(\beta_G\) is the dosage association in group 1, \(\beta_G+\beta_{GA}\) is the association in group 2, and \(\beta_{GA}\) is their difference on the modeled outcome scale. The test of effect heterogeneity is \(H_0:\beta_{GA}=0\), with its alternative specified in advance. This interpretation requires the model to be appropriate, the group and genotype terms to be estimable, and covariates to be handled consistently. For more than two groups, use prespecified indicators and a joint test of the genotype-by-group terms when the target is equality across all groups.

A group interaction is not the only relevant design. If the scientific target concerns continuous genetic ancestry, use a prespecified continuous ancestry measure and an appropriate genotype-by-ancestry interaction rather than imposing arbitrary bins. If the target concerns local ancestry, specify the local ancestry coding and the interaction or stratification that corresponds to the question. In every case, check that the data contain enough joint variation in genotype, ancestry, and covariates to estimate the target. A coefficient is not identifiable merely because software returns a number.

Local Ancestry, Stratification, and Replication Targets

Local ancestry should be included when it answers a prespecified question, not simply because it is available. For example, a study may ask whether an additive genotype association differs with the number of copies assigned to a particular local ancestry at the locus. The analysis plan must define the ancestry coding, the genotype coding, any centering, and the covariates. Because genotype and local ancestry can be correlated, the design must provide adequate combinations of genotype and local ancestry; otherwise, their separate contributions may be weakly identified. Local ancestry inference quality, reference representation, and uncertainty should be checked at the analyzed region.

Replication targets should be defined before results are reviewed. Specify the phenotype and its scale, effect allele and strand harmonization rules, variant or region being tested, covariate model, ancestry or population target, and the effect measure to compare. State whether replication means a same-direction association, a prespecified minimum effect, a statistically significant result under a stated test, or a combination. A replication sample with a different phenotype scale or a different proxy variant may test a related but not identical target. Allele harmonization and effect-scale comparability are essential for interpreting direction and magnitude.

Do not base replication planning only on the discovery point estimate, which can be imprecise and inflated when selected for significance. Use a scientifically meaningful target effect and the expected information in the replication sample. Plan enough data for the intended inference, including a useful confidence interval or a test of a specified difference if cross-population portability is the aim. For binary outcomes or complex designs, calculate power using the relevant model and expected covariate-adjusted information rather than importing the quantitative-trait approximation above.

Diagnostics, Alternatives, and Common Failure Modes

Preplanned checks should ask whether the measurements and model support the intended comparison. Examine per-group sample and missingness counts, phenotype distributions, genotype and allele frequencies, imputation or dosage quality, covariate overlap, and the distribution of global and local ancestry. Check model rank, influential observations, residual patterns for linear models, and whether the genotype has meaningful residual variation after covariate adjustment. For an interaction, inspect the joint distribution of the interacting variables and report the standard error of the interaction estimate. Serious imbalance or sparse combinations can make an interaction unstable even when the overall sample is large.

The main analytic alternatives answer different questions:

These approaches are not interchangeable. A pooled main-effect-only model imposes a common association on its modeled scale; it is a poor substitute when heterogeneity is a target. Stratified analyses can be transparent but may be underpowered in smaller strata. Meta-analysis can combine evidence but cannot repair incompatible phenotypes, allele coding, or poorly estimated stratum-specific effects. Study plans should select the primary approach and identify supplementary summaries without treating a single combined estimate as proof of equal effects.

Common mistake: “The study has 10,000 participants, so it must have enough power to compare ancestry groups.” Total sample size alone does not establish precision. If one group is small, the phenotype is noisy, or genotype is almost entirely explained by adjusted covariates, the information for its effect can be low. Plan using the expected analyzable sample and residualized genotype information (or an appropriate model-specific calculation), then check that the observed data support the assumptions. A second mistake is calling an association a replication because its p-value is below a threshold, without confirming that the allele, phenotype, effect scale, and target population match the prespecified question.