Population Stratification

The root cause of stratification is the non-random mating and genetic drift that occurs when populations are separated over many generations. This leads to systematic differences in the baseline frequency of single nucleotide polymorphisms (SNPs) between subpopulations. A variant common in one ancestral group may be rare in another, entirely independent of any disease mechanism. When cases and controls in a study are drawn disproportionately from these different ancestral backgrounds, any allele that happens to be more frequent in the over-represented group will appear falsely associated with the disease.

Stratification creates false-positive associations that are statistically significant but biologically meaningless. The classic example is the 'chopstick gene' fallacy: a hypothetical study might find a strong genetic association with chopstick proficiency, but the true underlying variable is ancestral background. The allele is simply more common in East Asian populations, and the cultural trait is the confounder. In real-world studies, this manifests as SNPs reaching genome-wide significance (p < 5×10⁻⁸) not because they are causal, but because they are markers of ancestry that correlate with the differing disease prevalence between the sampled sub-groups.

A subtle and often undetected form of stratification is cryptic relatedness, where unknown distant familial relationships exist between individuals in a study. Even if a study avoids obvious population structure, the inclusion of many distantly related individuals creates subtle allele frequency differences within the sample. This violates the assumption of independent observations in standard statistical tests like the chi-squared test or logistic regression, leading to an inflated Type I error rate. This is particularly problematic in large biobank-scale datasets where exhaustive pedigree checking is impossible.

Stratification becomes a confounding variable specifically when the disease prevalence and the ancestral makeup both vary between the sampled cohorts. If a disease is more common in one subpopulation, and that subpopulation is over-represented in the case group, any allele with a higher frequency in that subpopulation will be statistically linked to the disease. This is the classic definition of a confounder: a variable (ancestry) that is associated with both the exposure (allele frequency) and the outcome (disease status). The effect size of the confounded SNP can be substantial, often dwarfing the subtle signals of true polygenic risk variants.

The primary diagnostic tool for detecting stratification is Principal Component Analysis (PCA) applied to a matrix of genome-wide genotypes. PCA computes the axes of maximal genetic variation, and the top principal components (PCs) almost always correspond to geographic ancestry. By plotting cases and controls against the top PCs, researchers can visually identify clustering by ancestry. A failure of cases and controls to overlap completely on the PC plot is a definitive sign of stratification. The genomic inflation factor (λ_GC) provides a quantitative summary of the overall deviation of the test statistic distribution from the null.

Modern correction methods move beyond simple PCA adjustment to mixed linear models (MLMs). These models explicitly account for the random effect of genetic relatedness by incorporating a genetic relationship matrix (GRM). The GRM estimates the pairwise relatedness between all individuals in the study directly from the SNP data. By modeling phenotype as a function of both the fixed effect of the candidate SNP and the random polygenic background effect captured by the GRM, MLMs effectively control for both population structure and cryptic relatedness, producing well-calibrated p-values.