Genotype Imputation Quality: Why a High INFO Score Is Confidence, Not Correctness
Statistical Genomics & GWAS
Genotype Imputation Quality: Why a High INFO Score Is Confidence, Not Correctness
Imputation turns a few hundred thousand genotyped SNPs into tens of millions, filling the gaps from a reference panel and stamping each filled variant with a quality score. The trap is what that score measures: it is the model’s certainty about its own guess, computed without ever seeing a true genotype — not a measurement of whether the guess is right.
Genotyping arrays are cheap because they are sparse: they directly measure a few hundred thousand to a couple million pre-selected SNPs, a tiny fraction of the variation in a genome. Imputation is what makes that sparseness workable. By matching the observed genotypes against a reference panel of densely sequenced haplotypes, it statistically infers the alleles at the millions of positions the array never measured — expanding a chip to something approaching whole-genome resolution for a fraction of the cost. It is the quiet engine underneath nearly every genome-wide association study.
Because those inferred genotypes are guesses, every imputation tool attaches a quality score — the INFO score from IMPUTE, Rsq from Minimac, conceptually similar numbers between 0 and 1. The universal convention is to filter on it: drop everything below some threshold (0.3, 0.4, 0.8 depending on the study) and analyze the rest as if it were observed. That filtering step is doing far more work than most analyses acknowledge, because of a subtlety in what the score actually is.
What the score measures — and what it doesn’t
There are two fundamentally different quantities in play, and conflating them is the central error. The first is true accuracy: the squared correlation between the imputed dosage and the real genotype — dosage r² — which you can only compute if you have the true genotypes from, say, whole-genome sequencing of the same samples. The second is the INFO/Rsq score: an estimate of imputation quality that the software produces without any access to the truth, derived entirely from the posterior genotype probabilities the model assigned.
INFO is essentially a measure of how decisive the model was. If the hidden Markov model placed almost all probability on one genotype at a site, INFO is high; if it spread probability across genotypes, INFO is low. That is a statement about the model’s internal certainty, not about correctness — and a model can be confidently wrong. The people who build these panels say so directly: although INFO-values and squared correlations are useful for characterising overall imputation success, high INFO or r² values do not guarantee that the corresponding genotypes are inferred correctly, so it is important to directly assess the accuracy of the imputed genotypes.1 The score and the accuracy are correlated in aggregate, but they are not the same number, and they come apart exactly where it matters.
They come apart literally: the deviation between Rsq and true dosage r² has been widely observed across different reference panels and software.2 The gap is not noise — it has structure, and the structure tracks the two conditions where imputation is hardest: rare variants and mismatched ancestry.
Rare variants: the accuracy cliff
Imputation works by finding stretches of reference haplotype that match the sample’s genotyped SNPs and copying in the alleles between them. This works because common variants sit on haplotype backgrounds shared by many people in the panel. A rare variant, by definition, appears on only a handful of reference haplotypes — sometimes too few for the model to place it confidently, or at all. The result is a hard floor on rare-variant performance that better common-variant imputation does not lift.
The literature is consistent and blunt about this. Public panels accurately predict common variants (MAF ≥ 5%) and low-frequency variants across diverse populations, but imputation of rare variation (MAF < 0.5%) remains rather limited,3 and in head-to-head panel comparisons accuracy decreased considerably in the rare-variant group (MAF < 1%) for every reference panel tested.4 This is precisely the frequency range where the INFO score is least trustworthy as a proxy for accuracy — so a rare variant driving a GWAS hit deserves direct validation, not a passing INFO score, before anyone believes it. Larger and better-matched panels push the floor down — a recent large panel reached r² of about 0.75 for variants with minor allele frequency as low as 2×10⁻⁴ in well-matched samples — but the floor moves, it does not disappear, and it moves only when the panel actually contains the relevant haplotypes.5
Ancestry: the panel has to contain people like your sample
Imputation can only copy in haplotypes that exist in the panel. If the sample’s ancestry is under-represented in the reference, the model has few matching haplotypes to work from, and both the accuracy and the reliability of the quality score degrade. This is the same Eurocentric-data problem that haunts polygenic scores, entering one step earlier in the pipeline: a panel built mostly from one population images that population’s haplotypes well and others poorly.
The mechanism is specific. In the hidden Markov model, a parameter governing haplotype template-switching is estimated from the panel, and when a multi-ancestry panel mixes in distant ancestries, those distant haplotypes distort that estimate — which in turn distorts the imputed dosage, the true dosage r², the reported Rsq, and the deviation between them; moreover, the distant ancestries contribute only a handful of extra high-quality variants because the model assigns their haplotypes low probability.2 The counterintuitive upshot from that analysis is that simply throwing every ancestry into one giant panel is not automatically better for a minor-ancestry sample — the quality metric itself becomes less faithful. An ancestrally matched panel raised accuracy for an under-represented population more than a larger but poorly matched one did.4 Matching the panel to the sample is not a refinement; it is the thing that makes the score mean what you think it means.
INFO is the model’s confidence in its own guess, not a measurement of the guess. Those two agree for common variants on well-represented haplotypes — and diverge exactly where imputation is hard: rare variants, and samples whose ancestry the panel barely contains. The filter that feels like a safety net is thinnest precisely where you most need one.
The downstream trap: dosages become hard calls
One more failure happens after imputation, in how the output is used. Imputation naturally produces a dosage — an expected allele count between 0 and 2 that carries the model’s uncertainty. Many downstream steps quietly convert that dosage to a “best-guess” hard genotype by rounding. Rounding throws away exactly the uncertainty the imputation worked to quantify: a dosage of 0.5 (genuine coin-flip) and a dosage of 0.05 (confident reference) both become 0, and are then analyzed as if equally certain. Association tests that take dosages directly preserve that information; pipelines that round do not. When an imputed variant matters, it is worth checking whether the analysis used dosages or hardened them — and, for the truly consequential ones, confirming the genotype directly rather than trusting any imputed value, hardened or not.1
What to check before trusting imputed genotypes
- Read INFO/Rsq as confidence, not accuracy. A high score means the model was decisive, not that it was right; the two diverge for rare and ancestry-mismatched variants.1
- Stratify quality by allele frequency. A single INFO threshold applied across all frequencies passes rare variants whose true accuracy is far lower; rare-variant hits need direct validation.34
- Match the reference panel to the sample’s ancestry. A well-matched panel beats a larger, poorly matched one, and keeps the quality metric faithful; more ancestries is not automatically better.24
- Analyze dosages, not hardened calls. Rounding to best-guess genotypes discards the uncertainty imputation exists to express; prefer dosage-aware association methods.1
- Validate the variants that carry the conclusion. For any imputed variant driving a result, confirm against sequencing rather than relying on the score alone.5
Imputation is one of the highest-leverage tricks in statistical genetics — it is why a cheap array can power a genome-wide scan. But its quality score is a self-assessment, and like any self-assessment it is most confident precisely where it has the least information to be confident about. The number in the INFO column is the model telling you how sure it is; whether it should be that sure is a separate question the score cannot answer.
The Takeaway
The INFO/Rsq score is the model’s certainty about its own guess — computed without ever seeing a true genotype, so it measures confidence, not correctness.
Where it misleads: confidence and accuracy agree for common variants on well-represented haplotypes and diverge exactly where imputation is hard — rare variants (true accuracy falls off a cliff the score doesn’t show) and samples whose ancestry the panel barely contains (distant haplotypes distort the metric itself).
What to do: read INFO as confidence, stratify quality by allele frequency, match the panel to your sample’s ancestry rather than maximizing panel size, and analyze dosages instead of rounding to hard calls.
The hard limit: a score derived from the model’s own posterior can’t certify correctness. For any imputed variant that carries a conclusion, confirm it against sequencing — the number in the INFO column can’t.
References
- Mitt M, Kals M, Pärn K, et al. Improved imputation accuracy of rare and low-frequency variants using population-specific high-coverage WGS-based imputation reference panel. European Journal of Human Genetics. 2017;25:869–876. nature.com/articles/ejhg201751
- Genotype imputation accuracy and the quality metrics of the minor ancestry in multi-ancestry reference panels. Briefings in Bioinformatics. 2024;25(1):bbad509. academic.oup.com/bib/article/25/1/bbad509
- Mitt M, et al. On the limits of rare-variant imputation from public panels (MAF < 0.5%). Eur. J. Hum. Genet. 2017;25:869–876. broadinstitute.org (Mitt 2017 PDF)
- Boonchai C, et al. A diverse ancestrally-matched reference panel increases genotype imputation accuracy in an underrepresented population. Scientific Reports. 2023;13:12439. nature.com/articles/s41598-023-39429-3
- A Genomics England haplotype reference panel and imputation of UK Biobank. Nature Genetics / PMC. 2024. ncbi.nlm.nih.gov/pmc/articles/PMC11387190