Population Stratification in GWAS: Why an Association Can Be Real, Strong, and Still About Ancestry
Statistical Genomics & GWAS
Population Stratification in GWAS: Why an Association Can Be Real, Strong, and Still About Ancestry
A genome-wide association study can hand you a variant with a crushing p-value that replicates in a second cohort and still has nothing to do with the trait. If ancestry differs between your cases and controls, the association test will faithfully report ancestry as biology — and the signal can look indistinguishable from a real one.
A GWAS is, at bottom, a very large stack of simple comparisons: for each of millions of variants, is the allele more common in people with the trait than in people without it? When it is, you have an association. The machinery is so standardized that the danger is easy to forget — an association is a statement about correlation, and correlation between an allele and a trait can arise from something other than the allele affecting the trait. The most important something, and the one that has haunted the field since its beginning, is ancestry.
The formal name is population stratification, and its definition is precise: a systematic difference in allele frequencies between subpopulations, which can lead to spurious association findings in case–control GWAS.1 The word “spurious” is doing heavy lifting. These are not weak or noisy false positives. They can be among the strongest, most reproducible signals in the study — which is exactly what makes stratification dangerous rather than merely annoying.
How ancestry manufactures an association
The confounding needs two ingredients, and human population history supplies both abundantly. First, allele frequencies differ across ancestral groups — the overwhelming majority of common variants are more frequent in some populations than others, for reasons of drift and history that have nothing to do with any trait. Second, many traits also differ across those same groups, often for entirely non-genetic reasons: diet, environment, healthcare access, socioeconomic factors. When both hold, ancestry is a common cause of both the genotype and the phenotype, and any variant whose frequency tracks ancestry will correlate with the trait.
The mechanism is worth stating exactly: spurious associations arise when the tested variant differs in frequency across ancestral populations and ancestry is also correlated with the phenotype — via environmental factors, or via causal loci elsewhere in the genome that differ in frequency across groups.2 Note what is not required: the variant need not be near any causal locus, need not be functional, need not have anything to do with the trait’s biology. It only needs to differ in frequency between groups that also differ in the trait. The classic cautionary tale is a hypothetical “chopstick gene” — any variant more common in an East Asian population would associate with chopstick use in a mixed sample, at genome-wide significance, and mean nothing.
Crucially, mixing cases and controls in different ancestral proportions is enough. If your cases happen to be drawn slightly more from one ancestral background than your controls — through recruitment, geography, or the structure of the biobank — then every ancestry-differentiated variant in the genome acquires a small association with case status. Across millions of markers, that produces a genome-wide smear of inflated test statistics, and some of them clear significance.
The corrections — and why each has an edge
The good news is that stratification is detectable and, usually, correctable. The bad news is that every correction rests on assumptions that can fail, and the failures are subtle.
Genomic control was the first standard. It estimates the genome-wide inflation factor λ from the bulk of the test statistics — which, being mostly null, should not be inflated — and divides it out, effectively forcing λ = 1.3 It is simple and it works in simple cases, but it has two well-known edges. First, it corrects the bulk of the distribution, but the tail — the part you actually care about — may not be well calibrated, and errors in estimating λ can leave uncorrected stratification.3 Second, and counterintuitively, it has fallen out of favor because many traits are polygenic enough that with large samples we genuinely expect λ > 1 from real signal, making genomic control over-conservative — which is why many now use LD score regression to distinguish confounding from polygenicity.3 A single number cannot tell inflation-from-ancestry apart from inflation-from-real-polygenic-signal; LD score regression can, by exploiting how the two scale differently with local LD.
Principal component analysis is the modern default. It infers the axes of ancestry directly from the genotype data and includes the top principal components as covariates in the regression, modeling ancestry differences between cases and controls so the correction is specific to each variant’s frequency variation across ancestral populations.4 It is powerful and general. But it assumes the confounding is captured by a handful of continuous linear axes, and subtle stratification not captured by common-SNP PCs will slip through — in an analysis of 36 GWAS summary datasets, several showed a significantly inflated LD score regression intercept even though they had included PC covariates.5 PCs handle the large, obvious axes of ancestry; fine-scale and recent structure can hide beneath them.
Correcting for stratification is not a checkbox that makes confounding go away. Genomic control can be both incomplete (in the tail) and over-conservative (for polygenic traits) at once; PCA can miss structure too subtle for its top axes. “We adjusted for ten PCs” is the beginning of the argument that a hit is real, not the end of it.
When the correction makes it worse
The most important recent lesson is that the correction itself is not always safe. In ancestrally heterogeneous — admixed — populations, adjusting for global-ancestry principal components can induce spurious associations rather than remove them. Adjusting for principal components can create spurious associations in admixed populations, because global ancestry and local ancestry at a given locus can relate to the phenotype in different ways, and a global-ancestry covariate can distort the local signal it was meant to protect.2 This is the deepest form of the trap: a method applied precisely to prevent false positives can generate them, in exactly the under-represented populations that most need reliable analysis. The correction is not neutral; it encodes an assumption about how ancestry relates to the trait, and where that assumption is wrong, the fix is a new source of error.
What to check before believing a hit
- Look at the genomic inflation and the QQ plot first. A λ well above 1 with an early lift-off from the diagonal signals confounding; interpret no single hit until the genome-wide calibration is understood.3
- Use LD score regression to separate confounding from polygenicity. The LDSC intercept distinguishes inflation from stratification versus inflation from real polygenic signal, which λ alone cannot.5
- Adjust with PCs (or a mixed model), but don’t assume it is sufficient. Top PCs catch coarse ancestry; check for residual inflation after adjustment rather than treating the covariates as a guarantee.4
- Treat admixed cohorts with special care. Global-ancestry PC adjustment can induce false associations there; consider local-ancestry-aware methods and validate carefully.2
- Ask whether the trait plausibly differs across ancestry for non-genetic reasons. If it does, the priors on stratification confounding are high and the bar for believing an ancestry-differentiated hit should rise accordingly.1
Population stratification is the original cautionary tale of statistical genetics, and it endures because it is not a bug to be patched but a structural feature of studying correlation in a species with population history. An association test cannot, by itself, tell you why an allele and a trait travel together — only that they do. Ancestry is the ever-present alternative explanation, and the whole apparatus of QQ plots, inflation factors, principal components, and LD score regression exists for one purpose: to make it hard for ancestry to masquerade as biology. It is never quite impossible.
The Takeaway
If ancestry differs between cases and controls, a GWAS will report ancestry as biology — producing associations that are strong, reproducible, and entirely spurious.
The mechanism: a variant only needs to differ in frequency across ancestral groups while the trait also differs across those groups (often for non-genetic reasons). It need not be functional or near any causal locus — the “chopstick gene” problem. Uneven ancestral mixing of cases and controls inflates test statistics genome-wide.
What to do: read the QQ plot and λ before any single hit, use LD score regression to separate confounding from real polygenicity, adjust with principal components or a mixed model — and then check for residual inflation rather than trusting the covariates.
The hard limit: the correction isn’t neutral. Genomic control can be incomplete in the tail and over-conservative for polygenic traits at once; in admixed populations, PC adjustment can induce the spurious associations it was meant to remove. “We adjusted for PCs” starts the argument that a hit is real; it doesn’t end it.
References
- Hajiloo M, Sapkota Y, Mackey JR, et al. ETHNOPRED; and the definition of population stratification as systematic allele-frequency difference causing spurious GWAS associations. BMC Bioinformatics. 2013;14:61. ncbi.nlm.nih.gov/pmc/articles/PMC3618021
- Adjusting for principal components can induce spurious associations in genome-wide association studies in admixed populations. bioRxiv. 2024. biorxiv.org/content/10.1101/2024.04.02.587682
- Correcting for population structure in GWAS — genomic control, λ, and LD score regression. Mathieson, methods notes. 2020. mathii.github.io — correcting for population structure
- Price AL, Patterson NJ, Plenge RM, et al. Principal components analysis corrects for stratification in genome-wide association studies. Nature Genetics. 2006;38:904–909. nature.com/articles/ng1847
- Correcting subtle stratification in summary association statistics (LDSC intercept inflation despite PC covariates). bioRxiv. 2016. biorxiv.org/content/10.1101/076133