Population Stratification

The most widely used method for correcting population structure. PCA computes continuous axes of genetic ancestry variation by performing eigendecomposition on the genotype covariance matrix. The top principal components capture the major axes of population differentiation.

• Input: A matrix of genotypes across individuals
• Output: Continuous eigenvectors representing ancestry gradients
• Correction: Include top 10–20 PCs as fixed-effect covariates in the GWAS regression model
• Limitation: Cannot fully correct for fine-scale structure or recent admixture

A unified framework that simultaneously accounts for population structure and cryptic relatedness by modeling the phenotype as a mixture of fixed genetic effects and a random polygenic background component.

• Key implementations: GCTA, GEMMA, BOLT-LMM, fastGWA
• Mechanism: Uses a genetic relationship matrix (GRM) to estimate the covariance structure between individuals
• Advantage: Controls for both distant ancestry and close familial relatedness without explicit PC selection
• Trade-off: Higher computational cost than PCA-based correction, though optimized algorithms now scale to biobank-sized datasets

A simple post-hoc correction method that adjusts GWAS test statistics by a uniform inflation factor (λ) to account for population stratification. The genomic inflation factor is estimated from the median observed chi-squared statistic divided by the expected median under the null.

• Correction: Divide all chi-squared statistics by λ
• Assumption: Inflation is uniform across the genome, which is often violated
• Modern use: Primarily reported as a quality control metric rather than a primary correction method
• Limitation: Overly conservative when many true associations exist, as polygenic signals inflate λ

A study design approach that restricts analysis to genetically homogeneous subgroups defined by self-identified race or ethnicity, then meta-analyzes results across strata.

• Application: GWAS conducted separately within European, African, East Asian, and other ancestry groups
• Strength: Reduces confounding when ancestry labels align with genetic clusters
• Critical caveat: Self-reported ancestry is a social construct that imperfectly captures genetic ancestry; admixed individuals may be excluded
• Modern best practice: Combine with PCA within each stratum to capture residual structure

A method that distinguishes true polygenic signal from population stratification by examining the relationship between LD scores and test statistics. Under polygenicity, variants in high-LD regions have inflated statistics; under stratification, inflation is uniform.

• Intercept: An estimate of confounding bias uncorrelated with LD; an intercept near 1.0 indicates no stratification
• Attenuation ratio: (Intercept - 1) / (Mean χ² - 1) quantifies the proportion of inflation attributable to stratification versus polygenicity
• Advantage: Does not require individual-level data; works with GWAS summary statistics alone

For recently admixed populations, local ancestry inference assigns each chromosomal segment to a reference ancestral population, enabling ancestry-specific effect estimation.

• Tools: RFMix, LAMP-LD, ELAI
• Correction: Include local ancestry dosages as covariates or stratify by local ancestry at each locus
• Application: Critical for trans-ancestry GWAS in Hispanic/Latino and African American cohorts where global PCA is insufficient
• Output: Locus-specific ancestry tracts that can be tested for ancestry-enriched disease risk

A dimensionality reduction technique applied to a genotype matrix to infer continuous axes of genetic ancestry. PCA identifies the primary directions of allele frequency variation across individuals, which typically correspond to geographic ancestry. The top principal components are routinely included as fixed-effect covariates in GWAS regression models to adjust for population structure. This method assumes that ancestry differences are captured by linear combinations of genotypes, making it effective for removing broad-scale stratification but less sensitive to fine-scale or cryptic substructure.

A diagnostic metric that quantifies the degree of systemic bias in a GWAS by comparing the median observed chi-squared test statistic to the expected null distribution. Under the null hypothesis of no association, λ should equal 1.0. Values substantially greater than 1.0 indicate inflation due to population stratification, cryptic relatedness, or other technical artifacts. A λ value of 1.05 or lower is generally considered acceptable in large GWAS, though polygenic traits naturally produce modest inflation due to many true small-effect associations.

A technique that distinguishes polygenic signal from population stratification by regressing GWAS test statistics against LD scores. LD scores measure the total amount of genetic variation tagged by each variant. Under polygenicity, variants with higher LD scores show higher test statistics. Under stratification, the relationship is flat. LD Score Regression estimates the intercept of this regression, which captures inflation not attributable to polygenic architecture, providing a robust measure of confounding separate from true genetic signal.

A statistical framework that models both fixed effects (the genetic variant being tested) and random effects (a genetic relationship matrix capturing relatedness and ancestry). By including a kinship matrix as a random effect, LMMs simultaneously account for population structure and cryptic relatedness without requiring explicit ancestry covariates. Implementations like GCTA, GEMMA, and BOLT-LMM are widely used in GWAS to control type I error inflation while maintaining statistical power.

The presence of unrecognized distant familial relationships among study participants that violates the assumption of independent observations in GWAS. Unlike obvious first-degree relatives, cryptic relatedness involves individuals who share small proportions of their genome identical-by-descent (IBD) without known genealogical connections. This can inflate test statistics and produce spurious associations. Methods like Kinship-based INference for GWAS (KING) detect cryptic relatedness from genotype data, enabling removal or modeling of related individuals.

Advanced polygenic risk score approaches that explicitly model or correct for population structure during effect estimation. Methods include:

• LDpred2: Can incorporate PCA covariates to adjust effect sizes
• PRS-CSx: A cross-population extension that leverages GWAS from multiple ancestries
• SDPRX: Models effect size heterogeneity across populations
• PolyPred: Combines within-population and cross-population effect estimates These methods aim to improve PRS portability across diverse ancestral groups by accounting for differences in LD patterns and allele frequencies.