Imputation Accuracy

The primary metric for imputation accuracy, measuring the squared Pearson correlation between imputed dosages and true genotypes at a variant.

• Calculation: Evaluated in masked or hold-out validation sets where true genotypes are known
• Interpretation: R² = 1.0 indicates perfect imputation; values above 0.8 are generally considered high quality
• INFO score: A related metric from IMPUTE2 software that approximates R² without requiring true genotypes
• Thresholds: Variants with R² < 0.3 are typically excluded from PRS analyses to prevent noise amplification

Two distinct approaches to measuring accuracy, each with different implications for downstream modeling.

• Dosage R²: Evaluates the continuous allele dosage (0-2), preserving uncertainty information critical for probabilistic PRS methods like PRS-CS
• Best-guess concordance: Measures the agreement between hard-called imputed genotypes and true genotypes, ignoring uncertainty
• Key distinction: Dosage R² is more informative because it penalizes uncertain imputations that best-guess concordance may mask
• Recommendation: Always report dosage-based metrics for PRS applications where uncertainty propagation matters

Imputation accuracy varies dramatically across the allele frequency spectrum, requiring stratified reporting to avoid inflated aggregate metrics.

• Common variants (MAF > 5%): Typically achieve R² > 0.9 with modern reference panels
• Low-frequency variants (1% < MAF < 5%): Moderate accuracy, R² often between 0.6-0.8
• Rare variants (MAF < 1%): Substantially lower accuracy, R² frequently below 0.5
• Impact on PRS: Rare variant imputation errors can introduce systematic bias in polygenic scores, especially when using methods that do not account for imputation uncertainty

The choice of reference panel is the single largest determinant of imputation accuracy, affecting both variant coverage and phasing quality.

• Panel size: Larger panels (e.g., TOPMed with 97K+ genomes) capture more haplotypes and improve rare variant imputation
• Ancestry matching: Panels must represent the target population's genetic diversity; mismatched ancestry causes systematic accuracy degradation
• Sequencing depth: Whole-genome sequenced panels outperform whole-exome or genotyping-array-based panels
• Cross-ancestry PRS: Poor reference panel representation of non-European populations is a primary driver of PRS transferability failure

The gold-standard approach for measuring imputation accuracy uses masked variant analysis to simulate the imputation scenario.

• Procedure: Genotyped variants are intentionally removed (masked) from the dataset, imputed against the reference panel, then compared to the withheld true genotypes
• Masking rate: Typically 2-5% of variants are masked to maintain haplotype structure integrity
• Cross-validation: Multiple masking iterations provide robust accuracy estimates across the genome
• Software: Tools like SNPtest and bcftools provide built-in masking and accuracy calculation functions

Systematic filtering of poorly imputed variants is essential before PRS construction to prevent noise propagation into risk predictions.

• R²/INFO filtering: Remove variants below a study-specific threshold (commonly 0.3 or 0.5)
• Hardy-Weinberg equilibrium: Imputed variants showing extreme HWE deviation may indicate systematic errors
• Allele frequency comparison: Validate imputed MAF against external reference frequencies to detect batch effects
• Differential missingness: Check for imputation quality differences between cases and controls that could introduce bias

A specific metric output by the IMPUTE2 software that estimates imputation quality for each variant. It ranges from 0 to 1, approximating the squared correlation between imputed and true genotypes. An Info Score > 0.8 is generally considered high confidence, while scores < 0.3 indicate poor imputation quality. This metric accounts for uncertainty in the posterior genotype probabilities rather than relying solely on hard-called best-guess genotypes.

The Minimac4 imputation pipeline reports an estimated squared correlation (R²) between imputed and true genotypes. Unlike empirical correlation, this metric is estimated internally using leave-one-out cross-validation within the reference panel. It is computationally efficient and widely used in large-scale consortia. A typical filtering threshold retains variants with Minimac4 R² > 0.3 for downstream GWAS or PRS analysis.

The allele frequency spectrum and sample size of the reference panel directly constrain achievable imputation accuracy. Dense panels like the Haplotype Reference Consortium (HRC) with 64,976 haplotypes or TOPMed with 97,256 deeply sequenced genomes enable accurate imputation of rare variants (MAF < 0.1%). Sparse panels like 1000 Genomes Phase 3 struggle with low-frequency variants, producing lower accuracy scores for those sites.

Imputation algorithms do not assign a single genotype but instead calculate a probability distribution over the three possible genotypes (AA, AB, BB). The maximum posterior probability reflects the confidence in the most likely call. A variant with a 0.99 probability for 'AB' is highly certain, while one with 0.40/0.35/0.25 indicates ambiguity. Accuracy metrics aggregate these probabilities across all samples to produce a single quality score.

Two distinct validation approaches exist. Dosage R² calculates the squared Pearson correlation between imputed allelic dosages (continuous values from 0 to 2) and true genotypes—this is the standard accuracy metric. Best-guess concordance converts imputed probabilities to hard calls and measures the percentage of exact matches. Dosage R² is preferred because it retains uncertainty information and better reflects the input to downstream PRS models.

Imputation accuracy is not uniform across the allele frequency spectrum. Common variants (MAF > 5%) routinely achieve R² > 0.95 with modern reference panels. Low-frequency variants (0.5% < MAF < 5%) show moderate accuracy (R² 0.6–0.9). Rare variants (MAF < 0.5%) exhibit sharply reduced accuracy (R² < 0.5). Quality control pipelines must apply MAF-stratified accuracy filters rather than a single global threshold to avoid discarding well-imputed rare variants or retaining poorly imputed common ones.