Statistical Imputation of Intelligence Scores Expands Genetic Discovery in UK Biobank

One of the most stubborn obstacles in the genetics of human intelligence is not the biology itself but who shows up to take the test. In large biobanks, cognitive assessments are voluntary, and the people who complete them tend to be more educated and less socioeconomically deprived than those who skip them. That skew, known as ascertainment bias, can quietly distort genetic associations, heritability estimates and genetic correlations, sending researchers chasing signals that reflect who participated rather than how cognition is inherited. A new study published in Nature Genetics describes a carefully validated statistical workaround that could change how such studies are done.

A team led by researchers at Amsterdam UMC and the Wellcome Sanger Institute set out to squeeze more value from the fluid intelligence measures collected in the UK Biobank, a resource of roughly half a million British volunteers. Fluid intelligence, the capacity to solve novel problems independent of acquired knowledge, was assessed through short verbal-numerical reasoning tests administered at the baseline in-person visit and in later online questionnaires. Only about 62 percent of participants ever completed one of these tests, and earlier studies typically analyzed just a subset of those. By integrating scores from five different testing occasions and then statistically imputing scores for participants who never took any test, the team expanded the effective sample from roughly 270,000 measured individuals to more than 455,000.

The imputation itself relied on a matrix-completion algorithm called SoftImpute, which fills in missing entries by exploiting low-dimensional structure in the data. The researchers began with 154 biobank variables correlated with intelligence, then deliberately pruned that set down to 82 variables whose genetic correlation profile aligned more closely with intelligence than with a residual noncognitive component of educational attainment. This refinement step proved crucial: it modestly reduced raw prediction accuracy, from a correlation of about 0.52 to 0.46, but reduced the importation of socioeconomic and other noncognitive signal into the imputed phenotype. Notably, an accuracy of 0.46 falls below the test-retest reliability of the fluid intelligence test itself, which the team measured at roughly 0.6, meaning the imputed values add noise but, according to the authors’ theoretical framework, need not add systematic bias.

That theoretical framework is one of the study’s most distinctive contributions. The authors formally derived the conditions under which pooling measured and imputed phenotypes recovers consistent population genetic effects. Two assumptions matter most: measurement error in the observed scores must be uncorrelated with genotype, and the imputed values must be calibrated to the intended cognitive construct in the unmeasured group, both in their mean and in their covariance with each genetic variant. Although these assumptions cannot be proven directly, they yield a testable consequence: the genetic correlation profile of the combined analysis should match that of an independent GWAS targeting the same construct. The team checked this against the COGENT consortium GWAS of cognitive function and found close correspondence.

The validation also exploited within-family GWAS, a technique that separates direct genetic effects from confounding sources such as genetic nurture, assortative mating and population stratification. In population-based analyses, a negative correlation between direct effects and the so-called nontransmitted coefficient has previously been interpreted as a signature of ascertainment bias. For the baseline fluid intelligence measure alone, that correlation was strongly negative at about minus 0.54. Integrating scores across testing occasions moved it closer to zero, and adding imputed values pushed it to a point estimate that was no longer significantly different from zero, exactly what the ascertainment-bias hypothesis predicted.

The payoff for common-variant discovery was substantial. Combining measured and imputed scores raised the number of independent genome-wide significant lead SNPs from 390 to 550, and meta-analyzing the result with COGENT pushed the total to 620 lead SNPs in a GWAS of 490,700 individuals. Larger sample sizes also improved gene prioritization, strengthened enrichment of associated genes in brain tissue, and monotonically improved polygenic index prediction of measured cognitive ability in three independent cohorts: the ALSPAC birth cohort, the Millennium Cohort Study and the INTERVAL study. Crucially, direct genetic effect estimates within families also increased with sample size, indicating that the gains were not an artifact of population-level confounding.

The most striking results came from the rare variant side. Using whole-exome sequencing data, the researchers tested whether ultra-rare protein-truncating variants and damaging missense variants in individual genes were associated with fluid intelligence. With measured scores alone, only five genes passed the false discovery rate threshold of 1 percent. With the combined measured-and-imputed phenotype, that number jumped to 26, a 5.2-fold increase in gene-level discovery, while effect size estimates remained essentially unchanged, suggesting the imputation introduced power rather than distortion. Fourteen of the 26 genes are already cataloged in the DDG2P database of developmental disorder genes, a significant enrichment given that only about 13 percent of tested protein-coding genes belong to that list.

Eight of the significant genes had no strong prior evidence of involvement in intelligence or neurodevelopmental conditions, yet the evidence that they matter is compelling. Across a cohort of more than 31,000 exome-sequenced probands with neurodevelopmental conditions, damaging de novo mutations in the 12 newly implicated non-DDG2P genes occurred 2.59 times more often than expected by chance. And in aggregate, protein-truncating variants in the newly discovered genes were significantly associated with lower measured cognitive ability in both ALSPAC and the Millennium Cohort Study, providing external replication that does not depend on the imputed phenotype at all. Comparisons with educational attainment showed the imputed intelligence measure outperformed that proxy, yielding 26 significant genes versus nine for educational attainment.

The findings carry a broader message about the genetics of cognition: rare variants with moderate to large effects on intelligence, estimated at roughly 0.25 to 1 standard deviations per carrier, are more pervasive than previously appreciated. The associations largely persisted even after excluding individuals with recorded diagnoses of intellectual disability, autism or epilepsy, implying that damaging variants in these genes can shape cognitive ability across the population without necessarily pushing carriers over a clinical diagnostic threshold. Polygenic background and environmental factors likely modify how such variants manifest clinically, blurring the traditional dominant-versus-recessive dichotomy into more of a gradient.

The authors are careful to frame the work as a pragmatic response to the limitations of existing biobank data rather than an endorsement of prediction as a substitute for measurement. Imputed phenotypes blur the boundary between a construct of interest and its correlates, and some residual noncognitive signal almost certainly remains in the combined measure, as reflected in slightly elevated genetic correlations with educational attainment and household income. The team argues that the most robust path forward is better phenotyping at the design stage, with routine collection of high-quality cognitive measures and deliberate strategies to minimize selection bias. Still, for researchers working with the biobanks that exist today, the study offers both a powerful tool and a principled framework for deciding when imputed traits can be trusted, and it opens a route to finding new neurodevelopmental risk genes hiding in the genomes of people who never took the test.

Subject of Research: Statistical imputation of fluid intelligence scores in the UK Biobank to reduce ascertainment bias and boost genetic discovery for cognition and neurodevelopmental conditions

Article Title: Imputation of fluid intelligence scores reduces ascertainment bias and increases power for analyses of common and rare variants

Article References: Imputation of fluid intelligence scores reduces ascertainment bias and increases power for analyses of common and rare variants. (n.d.). https://doi.org/10.1038/s41588-026-02787-5

Image Credits: AI Generated

DOI: 10.1038/s41588-026-02787-5

Keywords: fluid intelligence, UK Biobank, genome-wide association study, ascertainment bias, phenotype imputation, rare variants, polygenic scores, neurodevelopmental disorders, genetic correlations, SoftImpute, developmental disorder genes, cognitive ability