How Ancestry Creates Genetic Confounding

Why Ancestry Can Confound a Genetic Association

A genetic association is a relationship between a genetic measurement and a phenotype in observed data. It does not, by itself, show that changing the measured variant would change the phenotype. One reason is that genetic variants are not distributed identically in every population. If ancestry is also related to phenotype—for example, through ancestry-correlated environments or other genetic differences—then ancestry can connect genotype and phenotype even when the variant under study has no causal effect.

This mechanism is called population stratification when differences in population structure create confounding in a genetic association analysis. The central issue is not that ancestry is a biological explanation for every trait difference. Rather, ancestry can index population history and correlate with other influences on the phenotype. The resulting association may reflect those shared differences instead of an effect of the specific variant being tested.

The Causal Diagram

A directed acyclic graph (DAG) represents proposed causal relationships with arrows. In the diagram below, ancestry influences the distribution of the tested genotype and is also related to the phenotype. The arrow from \(G\) to \(Y\) is included to represent a possible causal effect; under the null scenario considered here, that arrow is absent.

The route \(G \leftarrow A \rightarrow Y\) is a backdoor path: it links genotype to phenotype through a common cause rather than through a causal effect of the tested variant. If both arrows from \(A\) are present, that path is open unless the analysis accounts for the relevant ancestry variation. Thus, an observed \(G\)-\(Y\) association can arise even when \(G\) does not cause \(Y\).

A simple structural description makes the point precise. Suppose \(G\) is generated partly by ancestry, and \(Y\) is generated by ancestry and other factors, but not by \(G\):

where \(U_G\) and \(U_Y\) represent other influences. There is no \(G\) argument in the equation for \(Y\), so this model contains no causal effect of the tested genotype on the phenotype. Nevertheless, different values of \(A\) can produce different genotype distributions and different phenotype distributions. Comparing genotype groups that contain different mixtures of ancestry groups can therefore compare people with different background risks.

This explanation depends on the causal structure, not on a claim that ancestry itself must directly cause the phenotype. For example, an ancestry grouping may correlate with environmental exposures, access to care, or other genetic variants that influence \(Y\). A causal diagram is a way to state and examine such assumptions; it is not a discovery that the arrows are true simply because the diagram has been drawn.

How Allele-Frequency Differences Enter

For a diploid autosomal variant, define \(G\) as the number of copies of a specified allele carried by an individual, so \(G\) takes values 0, 1, or 2. Let \(p_a\) be the frequency of that allele in ancestry group \(a\). If the two allele copies are sampled independently with probability \(p_a\), then the expected dosage in that group is \(2p_a\). This simple relationship shows why different allele frequencies create different genotype distributions.

The independence assumption in that calculation corresponds to random pairing of allele copies within the group, as in Hardy–Weinberg proportions. It is used here only to illustrate the expected dosage; real genotype distributions can depart from it because of factors such as inbreeding or nonrandom mating. The core confounding mechanism does not require this assumption: it requires that genotype distributions differ with ancestry and that phenotype distributions also differ with ancestry.

An allele-frequency difference alone is not confounding. If \(G\) varies with ancestry but \(Y\) does not, the ancestry-to-phenotype link needed for this path is absent. Likewise, if phenotype differs with ancestry but the tested genotype distribution does not, this particular path does not connect \(G\) and \(Y\). Confounding is plausible when the relevant links coexist, and the size of the resulting association depends on how strongly they operate and on how ancestry groups are represented in the sample.

A Hypothetical Pooled Association With No Within-Group Effect

The following example makes both links visible in observed counts. It is a small hypothetical dataset, not a real study. Let \(G\) be the number of copies of an allele, and let \(Y=1\) indicate a binary phenotype. Within each ancestry group, suppose the phenotype rate is the same for all three genotypes. The rates differ between ancestry groups, and genotype counts also differ between groups.

What the Example Does—and Does Not—Establish

The example demonstrates a mechanism, not a universal outcome. Population structure can generate a spurious association, but the direction and size of the association are not fixed. They depend on the allele-frequency pattern, phenotype differences across ancestry, and the mixture of groups among the people being compared. Other sources of variation can also be present, including a real causal effect of the tested variant.

An ancestry grouping may also be too coarse to capture the relevant structure. People within one broad group can differ in ancestry, and a sample can include continuous or mixed ancestry. Conversely, group labels can correlate with social or environmental conditions for reasons that vary by setting. A label should not be treated as a complete measurement of ancestry or as a biological cause of a phenotype.

The appropriate conclusion from an observed association is therefore limited: genotype and phenotype differ together in the analyzed sample. To interpret that relationship causally, one must consider whether ancestry or other common causes could produce it, whether the study’s ancestry information captures the relevant structure, and whether the modeling assumptions are credible. This tutorial identifies the source of the concern; choosing and evaluating specific adjustment strategies requires additional tools.

Common Mistakes

• Equating association with causation. “The allele is associated with the phenotype, so the allele causes it” ignores the open path \(G \leftarrow A \rightarrow Y\). The observed association may be due to ancestry-related differences instead.
• Assuming every allele-frequency difference creates confounding. A frequency difference establishes a possible ancestry-to-genotype link, not an ancestry-to-phenotype link. Both are needed for this confounding path.
• Treating ancestry labels as complete or interchangeable with race. Ancestry is multidimensional, and broad labels may not represent the genetic or environmental structure relevant to a particular analysis.
• Assuming a null within-group association guarantees a null pooled association. The worked counts show why that reasoning can fail: genotype groups can contain different proportions of ancestry groups whose phenotype rates differ.
• Interpreting a confounded association as evidence about all members of a population. The pattern depends on the sample’s composition and may not transport to a population with different ancestry proportions or allele frequencies.