Ancestry Imbalance in Case-Control Sampling

Why Case-Control Ancestry Imbalance Matters

In a case-control genetic study, investigators compare genetic variation in people sampled because they have a disease (cases) with variation in people sampled as controls. If cases and controls have different ancestry composition, their pooled allele frequencies can differ even when the allele has the same frequency among cases and controls within every ancestry group. The pooled difference can therefore look like a genetic association when it reflects ancestry composition instead.

An imbalance can arise because disease occurrence varies across ancestry groups, because recruitment or control selection differs across groups, or because both processes operate. The distinction matters for study design, but neither source makes an unadjusted pooled comparison automatically reliable. The immediate diagnostic question is whether cases and controls have comparable ancestry composition and, if not, whether the genetic association is supported within the ancestry groups represented in both.

For a diploid genetic variant, let \(G\) be the number of copies of a specified allele, taking values 0, 1, or 2. Let \(p_{da}\) be that allele's frequency among sampled individuals with disease status \(d\) and ancestry stratum \(a\). The pooled allele frequency in group \(d\) is a weighted average of the stratum-specific frequencies:

Thus, a pooled case-control difference can come from differences in allele frequency within ancestry strata, differences in the ancestry weights, or both. When the within-stratum frequencies are equal between cases and controls, any pooled difference is due to the weights. This is the case-control form of the composition effect underlying ancestry confounding.

How Composition Alone Creates a Pooled Difference

A useful diagnostic calculation separates the composition component from the within-stratum component. For two ancestry strata, label them A and B, and define \(w_{dA}\) as the fraction of sampled group \(d\) in stratum A. If the allele frequency is the same in cases and controls within each stratum, so \(p_{1A}=p_{0A}=p_A\) and \(p_{1B}=p_{0B}=p_B\), then:

The pooled difference is the product of two contrasts: the difference in ancestry composition and the difference in allele frequency across ancestry groups. If either contrast is zero, this particular composition-induced difference is zero. If both are nonzero, the pooled frequencies can differ despite no within-stratum case-control difference. This calculation describes how the pooled contrast arises; it does not, by itself, establish a disease effect or a causal effect.

The same idea applies when ancestry is measured continuously rather than divided into a few categories: the case and control distributions of ancestry can differ, and allele frequency can vary across that ancestry distribution. Coarse strata may fail to capture this pattern. Conversely, a broad label is not a biological explanation by itself; it is a way of describing population structure that may track allele-frequency differences and other relevant factors.

Design Responses: Matching and Ancestry-Aware Sampling

A design response aims to make case and control ancestry distributions comparable before fitting an association model. In ancestry-aware matching, each case is paired with one or more controls similar on ancestry, or is matched within a defined ancestry stratum. A related option is to recruit cases and controls in planned numbers within each stratum. The objective is comparable representation, not identical genotypes: controls should not be selected based on the genetic variant being tested.

When sampling is balanced across strata, the pooled comparison is less exposed to differences in ancestry weights. Matching can also make the comparison more direct by restricting it to sets with similar ancestry. The design does not guarantee that all confounding is removed: ancestry may be measured imprecisely, important structure may remain within a stratum, and other case-control differences may persist. Analysis should respect the matching or stratification used in sampling, rather than treating the observations as if the design had never occurred.

There is a practical trade-off. Strict matching or stratified recruitment can reduce the number of eligible controls, limit the study to ancestry groups with cases and controls available, and make it harder to estimate effects for underrepresented groups. A central requirement is overlap: each ancestry range or stratum used for an adjusted comparison needs both cases and controls. If a stratum contains only cases or only controls, the data do not provide a direct within-stratum case-control comparison there. Statistical adjustment cannot manufacture that missing comparison.

Regression Adjustment for Ancestry

When ancestry is measured for each participant, logistic regression can estimate a genotype association conditional on ancestry covariates. For one binary ancestry indicator, a basic model is:

Here \(I(A=\mathrm{A})\) equals 1 for stratum A and 0 for stratum B. The coefficient \(\beta\) is the change in log odds of being a case for one additional copy of the allele, conditional on the ancestry term in this model. The corresponding conditional odds ratio per additional allele copy is \(\exp(\beta)\). This is an association estimate, not automatically a causal effect. With continuous ancestry measures or multiple components, the model includes suitable terms for those measures; their number and form should reflect the structure of the data and the modeling assumptions.

A common-effect model assumes that the genotype log-odds association is the same across ancestry values represented by the model. If the genetic association may differ across ancestry groups, include an interaction term, such as \(\delta G I(A=\mathrm{A})\). Then the genotype log-odds coefficient is \(\beta\) in the reference group and \(\beta+\delta\) in stratum A; a single common odds ratio is no longer an adequate summary when \(\delta\) is meaningful.

In case-control sampling, the numbers of sampled cases and controls are often set by design, so their ratio does not estimate disease prevalence in the source population. Under outcome-dependent sampling that depends on disease status alone, logistic regression can still estimate odds-ratio slopes under the model, while the intercept generally does not recover the population disease probability without additional information. If recruitment also depends on ancestry or other covariates, the sampling process and analysis need to be considered together.

Assumptions and Practical Checks

An ancestry-adjusted result is informative only to the extent that the data and model support the comparison it represents. Before interpreting the genotype coefficient, check whether ancestry was measured with enough resolution for the study, whether cases and controls overlap across the measured ancestry range, and whether the model represents the relevant relationship. A linear term in a component may not capture a nonlinear association; a common genotype coefficient may be inappropriate if effects differ across ancestry. These are modeling and design questions, not problems that can be solved merely by adding more covariates.

Matching and adjustment address observed ancestry information. They do not necessarily control unmeasured or poorly measured structure, differences in recruitment, or ancestry-correlated environmental factors. They also do not make a genetic association causal. The result should be described as an association conditional on the specified ancestry covariates and sampling framework. For a broad population claim, investigators should also consider which ancestry groups the sample actually represents.

Common Mistake: Treating a Covariate as a Cure

Wrong reasoning: “I added an ancestry variable to the regression, so any association that remains is independent of population structure and must be biological.” This conclusion goes beyond what the model establishes. The covariate may be too coarse, the functional form may be wrong, or important ancestry ranges may lack case-control overlap. A remaining association is conditional on the measured covariates and assumptions, not proof that all confounding has been eliminated.

Correction: Describe the adjustment actually made, check overlap and model fit, and interpret the genotype coefficient as a conditional association. Where effects may vary by ancestry, evaluate an interaction rather than assuming one common effect. Where the data have no case-control support for a comparison, acknowledge that limitation rather than treating model adjustment as a substitute for information.