{"paper_id":"033d852f-0e77-463b-a20f-72562effd1b2","body_text":"Charting the cognitive development of children using adult ‘polygenic g \nscores’ \n \nYujing Lin 1, Robert Plomin 1 \n1 Social, Genetic and Developmental Psychiatry Centre, King's College London \n \n \nLede \nAdult ‘polygenic g scores’ increasingly predict variance in cognitive ability across \ndevelopment from about 0.4% in infancy to 5% in childhood, 10% in adolescence, \nand 12% in early adulthood. \n \n \nAbstract \n \nThe most highly predictive polygenic scores in the behavioural sciences are \nfor cognitive traits, especially general cognitive ability (g) and educational \nachievement. We combined polygenic scores derived from genome-wide association \nstudies of adult g and educational attainment, conditioned on their prediction of adult \ng, to create adult 'polygenic g scores' which we used to chart the course of cognitive \ndevelopment of 10,000 white British children from infancy through early adulthood. \nWe integrated cross-sectional regression, latent growth curve, and cross-time \ncommon factor analysis to systematically characterise cognitive development. \nPolygenic g score showed minimal prediction in infancy, modest prediction in \nchildhood, and substantial prediction by early adulthood, accounting for 12% of the \nvariance. Higher polygenic g scores were associated with faster cognitive growth in \nlatent growth models. Prediction was strongest for a cross-time common cognitive \nfactor (15%), reflecting substantial stable genetic influences across development. \nWe also examined the polygenic g score’s prediction of educational achievement, \nbehaviour problems, and anthropometric outcomes and found similar developmental \nincreases in prediction for educational achievement. \n Together, our findings demonstrated that adult polygenic g scores can be a \nuseful tool for charting the development of cognitive traits. \n \nKey words: Polygenic scores, cognitive ability, educational achievement, \neducational attainment, mental health  \n \n  \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted December 23, 2025. ; https://doi.org/10.64898/2025.12.19.695378doi: bioRxiv preprint \n\nCharting cognitive development using adult ‘polygenic g scores’ 2 \nCharting the cognitive development of children using adult ‘polygenic g \nscores’ \n \nOne of the most important outcomes of the DNA revolution for research on \nintelligence and cognitive abilities is the polygenic score (Plomin, 2019). Genome-\nwide association (GWA) studies associate DNA differences, typically common \ngenetic variants called single nucleotide polymorphisms (SNPs), with cognitive \nabilities. Many such genetic variants contribute to the substantial heritability of \ncognitive abilities, and the largest effect sizes are minuscule, accounting for less \nthan .05% of the phenotypic variance (Plomin & von Stumm, 2018; Visscher et al., \n2021). The small effect sizes make it difficult to trace pathways from genes to brain \nto behaviour. In contrast, polygenic scores aggregate thousands of these small but \nadditive effects, translating GWA discoveries into a genetic predictor of individual \ndifferences among unrelated individuals in the population.  \nDespite the availability of many polygenic scores for cognitive outcomes to \ndate, two scores consistently drive the strongest predictions (Procopio et al., 2025). \nThe first is a polygenic score for intelligence. The largest GWA analysis of adult \nintelligence (N = ~270,000) produced what has been called the IQ3 polygenic score, \nwhich predicts about 5% of the variance of adult intelligence in independent samples \n(Savage et al., 2018). GWA analyses require large sample sizes to detect the small \neffects of DNA variants. Conducting a GWA study of intelligence requires not only \nobtaining and genotyping DNA but also testing for intelligence, which is difficult with \nsuch large samples. For this reason, years of schooling (educational attainment) has \nbeen used as a proxy for intelligence (Rietveld et al., 2014). Educational attainment \ncorrelates strongly with intelligence, about 0.50 phenotypically (Deary, 2012) and \n0.75 genetically (Hill et al., 2019). It is assessed with a single self-reported item \nabout the highest level of education and, crucially, is routinely collected in most GWA \nstudies as a demographic descriptor, making it possible to assemble huge samples \nfor meta-analysis. The largest GWA analysis of educational attainment had a sample \nsize of three million that yielded a polygenic score (EA4) predicting up to 9% of the \nvariance of intelligence, with the strongest prediction observed for verbal ability \n(Okbay et al., 2022). EA4 predicts more of the variance of intelligence than IQ3, \neven though educational attainment is a coarse proxy for intelligence, because the \nEA4 sample size is more than ten times greater than the IQ3 sample size. Here, we \ncombined IQ3 and EA4 weighted by their prediction of adult g and refer to this \ncomposite as an adult ‘polygenic g score’. \nBecause inherited DNA differences are fixed at conception, polygenic scores \nremain constant during development, which is unique in developmental research: \nAdult polygenic g score can predict adult g from birth just as well as in adulthood. If \nwe turn this around, we can use the adult polygenic g score as a tool to chart the \nemergence, growth, and changes of cognitive abilities across development and to \npinpoint the earliest cognitive traits that are genetically linked to adult cognitive \nabilities. To achieve this, we need longitudinal cognitive data that follow children from \ninfancy to adulthood.  \nSuch longitudinal research takes three decades, and as a result, few data \nexist. Meta-analytic evidence indicates that while cognitive ability tends to fluctuate in \ninfancy, its stability increases rapidly across early childhood, becoming highly stable \nby adolescence (Breit et al., 2024). Although these longitudinal phenotypic \ncorrelations represent a ceiling for predicting adult g from infant measures, adult \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted December 23, 2025. ; https://doi.org/10.64898/2025.12.19.695378doi: bioRxiv preprint \n\nCharting cognitive development using adult ‘polygenic g scores’ 3 \npolygenic g scores can be used as a sharper scalpel for dissecting facets of early \ncognitive development linked genetically to adult g.  \nA recent study leveraging adult-derived PGS found that, similar to phenotypic \ncorrelations, genetic prediction follows a trajectory of increasing strength across \ndevelopment (Gustavson et al., 2025). The highest prediction reported was for EA4 \nPGS, explaining 18% of the variance in cognitive ability by age 16, which is twice as \nmuch as the usual result of 9% for adult samples (Savage et al., 2018). Twin \nanalysis showed a similar pattern of increasing heritability. The IQ3 PGS also \nshowed a parallel pattern but with lower predictive strength. Furthermore, the \nauthors used confirmatory factor analysis for polygenic prediction and Cholesky \ndecomposition for twin analysis to reveal the source of stability over time. They found \nthat the increased prediction is attributable to the amplification of stable genetic \neffects shared across ages, rather than the emergence of new, age-specific genetic \nfactors. \nOur study builds on these findings using data from the Twins Early \nDevelopment Study (TEDS; Lockhart et al., 2023), a population-based cohort with \nover 10,000 British white children. TEDS offers significantly greater power with a \nnearly tenfold increase in sample size compared to the Gustavson et al. (2025) \nreport. TEDS children were assessed using diverse measures of cognitive \ndevelopment in infancy and early childhood (2, 3, and 4 years), middle childhood (7 \nand 9 years), adolescence (12, 14 and 16 years), and adulthood (26 years), \nproviding repeated measures of g, verbal ability, and nonverbal ability.  \nWe aim to replicate prior cross-sectional and common factor findings while \naddressing several developmental aspects that were not covered in previous work. \nFirst, we create a polygenic g score as a composite predictor by weighting EA4 and \nIQ3 based on their joint prediction of adult g (Maier et al., 2018). Second, we include \nlatent growth curve modelling to chart developmental trajectories as predicted by the \npolygenic g score. Third, we look at the tails of the polygenic distribution to \ninvestigate if cognitive development diverges for the highest and lowest scorers. \nFinally, educational achievement, behaviour problems, and anthropometric outcomes \nwere also assessed longitudinally, enabling a comprehensive examination of \npolygenic g score prediction across trait domains.  \nIn our preregistration (https://osf.io/vhuyj/overview), we specified five \nhypotheses: 1) the polygenic g score significantly predicts most phenotypic \noutcomes, with educational achievement yielding the strongest correlation, 2) \nprediction for cognitive abilities increases from infancy to early adulthood; 3) the \npolygenic g score is also predictive of behaviour problems and anthropometric \noutcomes; 4) prediction is linear across the distribution, indicating that high and low \nextremes are quantitatively, not qualitatively, different from the rest of the distribution; \nand 5) there are no significant sex differences in polygenic g prediction. \n  \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted December 23, 2025. ; https://doi.org/10.64898/2025.12.19.695378doi: bioRxiv preprint \n\nCharting cognitive development using adult ‘polygenic g scores’ 4 \nMethods  \n \nSample  \nWe leveraged data from the Twins Early Development Study (TEDS), a \nlongitudinal cohort of 13,759 families with twins born between 1994 and 1996 in \nEngland and Wales (Lockhart et al., 2023b). Phenotypic data were collected across \nmultiple waves, including assessments at approximately ages 2, 3, 4, 7, 8, 9, 10, 12, \n14, 16, 18, 21, 25, and 26 years. Genotypic data were available for 10,346 \nparticipants. Ethical approval for TEDS was obtained from King’s College London \nResearch Ethics Committee (References: PNM/09/10–104 and HR/DP-20/2122060), \nand informed consent was obtained prior to each wave of data collection. \nFor cross-sectional analyses, all participants with available DNA data and at \nleast one phenotypic measure were included. To calculate a total score for a given \nmeasure, participants were required to have completed at least half of the scale. For \nmeasures composed of multiple subscales, the same rule applied to each subscale, \nand a participant was included only if at least half of the subscales were available. \nFor longitudinal analyses, we included participants with data available for at \nleast two ages (N = ~4500 to ~8000). \n \nMeasures  \nPhenotypic Measures  \nThe present study focuses on outcomes selected for their consistent \ncollection across the ages. From ages 2 to 26, we assessed cognitive abilities, \neducational achievement, behaviour problems, and anthropometric measures \nconsistently using age-appropriate measures. Additional measures specific to one \nage or one developmental stage are provided in Appendix A. Complete \ndocumentation of all measures across ages is available in the TEDS data dictionary \n(https://datadictionary.teds.ac.uk/home.htm). \nEarly Cognitive Abilities (Ages 2-4). Phenotypic measures at early ages \nwere collected via booklets sent to families. At ages 2 and 3, booklets were sent only \nto families of twins born in 1994 and 1995, as twins born in 1996 were not age-\nappropriate for the tests. At age 4, booklets were sent to all families. \nVerbal ability at ages 2 to 4 included vocabulary (what children can say) and \ngrammar (how children use words). Vocabulary was assessed via parent-reported \nchecklists. At age 2, parents completed a 100-word checklist adapted from the \nMacArthur Communicative Development Index (MCDI; Fenson et al., 1993, 2000). At \nage 3, the checklist included 45 MCDI words and 55 new words from a literature \nreview and pilot testing with an additional two questions about whether the child was \ntalking and combining words. At age 4, 48 words selected from the literature review \nand pilot testing were used. Grammar was also measured using questions derived \nfrom the MCDI. At ages 2 and 3, parents completed the 6-item word use and 12-item \nsentence complexity scales. At age 3, the word use scale was expanded to 12 items. \nAt age 4, a single 6-point global rating scale assessed language complexity from ‘not \nyet talking’ to ‘talking in long and complicated sentences.’ The verbal ability \ncomposite for the MCDI was calculated as the standardised mean of vocabulary and \ngrammar scores. More detailed descriptions of the TEDS verbal measures between \nages 2 to 4 are available in the TEDS data dictionary and previous TEDS \npublications (Dionne et al., 2003; Hayiou-Thomas et al., 2012). \nNonverbal ability was assessed using the Parent Report of Children's Abilities \n(PARCA; Saudino et al., 1998), including parent-administered tasks and parent-\n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted December 23, 2025. ; https://doi.org/10.64898/2025.12.19.695378doi: bioRxiv preprint \n\nCharting cognitive development using adult ‘polygenic g scores’ 5 \nreport questionnaires. At age 2, parent-administered tasks included matching (8 \nitems), brick building (4 items), folding (1 item), and copying (7 items) from the \nBayley Scales of Infant Development (Bayley, 1993) and a design drawing task (4 \nitems) adapted from the McCarthy Scales (McCarthy, 1972). The parent-report \ncomponent assessed conceptual knowledge (26 items). At age 3, parent-\nadministered tasks included odd-one-out (16 items), design drawing (6 items), and \nmatching (16 items), with conceptual knowledge assessed via 24 items. At age 4, \nparent-administered tasks comprised the age 3 odd-one-out and design drawing \ntasks, plus draw-a-man (1 item) and puzzles (12 items). Conceptual knowledge was \nassessed via 12 items. The nonverbal ability composite was calculated as the \nstandardised mean of parent-administered and parent-report PARCA scores, \nfollowing the TEDS data dictionary and established practice in previous TEDS \npublications (Asbury et al., 2005; Oliver et al., 2004; Petrill et al., 2001; Saudino et \nal., 1998).  \nGeneral cognitive ability at each age was calculated as the standardised \nmean of verbal and nonverbal composites. \nCognitive Abilities at Later Ages (Ages 7-25). Cognitive ability measures at \nages 7, 9, 10, 12, 16, and 25 have been described in detail in a previous TEDS \npublication and are only briefly summarised here (see Lin et al., 2025 supplementary \nmaterials).  \nAt age 7, cognitive assessments were conducted via telephone interviews. \nVerbal ability was measured using the Wechsler Intelligence Scale for Children \n(WISC-III) Similarity and Vocabulary tests (Wechsler, 1992). Nonverbal ability was \nassessed using the Conceptual Grouping Test and WISC Picture Completion Test \n(McCarthy, 1972).  \nAt ages 9 and 10, verbal ability was assessed using WISC-III as a Process \nInstrument (WISC-III-PI) Vocabulary and General Knowledge tests (Kaplan et al., \n1999). Nonverbal ability was measured using the Cognitive Abilities Test 3 (CAT3) \nfigure classification and figure analogy tests at age 9 and WISC-III-UK Picture \nCompletion and Raven's tests at age 10 (Raven et al., 1996; Smith et al., 2001). \nAssessments were administered via mailed booklets at age 9 and online at age 10 \nand all subsequent ages. \nAt age 12, verbal ability comprised language tests (syntax, semantics, \npragmatics) and reading tests (comprehension and fluency) (GOAL plc (2002), n.d.; \nHammill et al., 1994; Markwardt, 1997; Torgesen et al., 1999; Wiig et al., 1989; \nWoodcock et al., 2001). Nonverbal ability was assessed using mathematical ability \ntests from the National Foundation for Education Research (Smith et al., 2001). \nGeneral cognitive ability was independently assessed using WISC-III-PI Vocabulary, \nGeneral Knowledge, Picture Completion (Wechsler, 1992), and Raven’s Pattern test \n(Raven et al., 1996). \nAt age 14, verbal ability was assessed using a 27-item WISC-III-PI vocabulary \nmultiple-choice test (Kaplan et al., 1999). Nonverbal ability was measured using the \n30-item Raven's Standard Progressive Matrices (Raven et al., 1996). \nAt age 16, verbal and nonverbal abilities were assessed using the Mill Hill \nVocabulary test (Raven et al., 1998) and Raven's Standard and Advanced \nProgressive Matrices (Raven et al., 1996). \nBetween ages 7 and 16, except for age 12, verbal and nonverbal composites \nwere calculated as standardised means of their respective component tests, and \ngeneral cognitive ability was calculated as the standardised mean of the verbal and \nnonverbal composites. \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted December 23, 2025. ; https://doi.org/10.64898/2025.12.19.695378doi: bioRxiv preprint \n\nCharting cognitive development using adult ‘polygenic g scores’ 6 \nAge 18, assessment only included two spatial ability online measures \ndeveloped by TEDS researchers (Malanchini et al., 2020; Rimfeld et al., 2017): a \nbricks test and a navigation study. The Bricks test measured spatial ability through \nmental rotation and visualisation using both 2D and 3D stimuli across six subtests (9 \nitems each): 2D rotation, 2D rotation and visualisation, 2D visualisation, 3D rotation, \n3D rotation and visualisation, and 3D visualisation. Both individual subtest scores \nand the overall Bricks total score (mean of all six subtests) were included in the \npresent study. The Navigation test included 30 tasks across six types (5 items each): \norientation-direction, orientation-landmarks, map reading without memory, map \nreading with memory, perspective, and scanning. Each task generated accuracy, \nspeed, and total scores; only the overall total score (mean of the six task types) was \nused in analyses.  \nAt age 25, cognitive abilities were assessed using Pathfinder, a gamified web-\nbased measure developed by TEDS researchers (Malanchini et al., 2021). Verbal \nability (20 items) included the Mill Hill vocabulary, missing letter, and verbal \nreasoning tests. Nonverbal ability (20 items) included Raven’s standard progressive \nmatrices and three visual puzzle tests on analogies, grouping, and logical \nsequences. Unlike earlier ages, cognitive ability scores at age 25 were not \nstandardised; general cognitive ability scores ranged from 0-40, while verbal and \nnonverbal scores each ranged from 0-20. \nEducational Achievement. Educational outcomes were examined using \neducational achievement from primary school to university. Educational outcomes at \nages 7, 9, 10, 12, 16, 18, 21, and 26 have been described previously (Lin et al., \n2025). The present study extended these general outcomes by including subject-\nspecific grades up to age 18 and adding assessment at age 14.  \nAt ages 7, 9, 10, and 12, teachers rated achievement in English and \nmathematics (starting at age 7) and science (starting at age 9) based on the National \nCurriculum Levels (https://www.gov.uk/national-curriculum/overview). The ratings \nranged from 0-4 at age 7 and 0-9 at later ages. \nAt age 14, parents reported grades in English, mathematics, and science and \nthe grades were translated to the 0-9 National Curriculum Levels. \nAt age 16, General Certificate of Secondary Education (GCSE) is a national-\nlevel exam taken at the end of compulsory education. GCSE exam grades were \nobtained for core subjects (English, mathematics, and science), humanities, and \nlanguages. The grades ranged from 4 (G) to 11 (A*).  \nAt age 18, A-level and AS-level qualifications were assessed across English, \nmathematics, science, technology, humanities, languages, and vocational subjects. \nA-levels are two-year qualifications completed after compulsory education and \nrequired for university entry. AS-levels represent completion of the first year only. \nWhen A-level grades were unavailable, AS-level grades were used. Grades ranged \nfrom 1 (E) to 6 (A*).  \nAt age 21, university degree classification was self-reported on a scale from 1 \n(lowest pass) to 5 (first-class honours).  \nAt age 26, most twins have completed their education. Therefore, educational \nattainment (i.e., years of schooling) was used to measure educational outcomes. For \ntwins missing age 26 data, age 21 educational attainment was used (correlation \nbetween ages: r = 0.86).  \nBehaviour Problems. Behaviour problems were primarily assessed using the \nStrengths and Difficulties Questionnaire (SDQ) (Goodman, 1997), which was \nadministered consistently across ages from 2 to 26 and across multiple informants. \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted December 23, 2025. ; https://doi.org/10.64898/2025.12.19.695378doi: bioRxiv preprint \n\nCharting cognitive development using adult ‘polygenic g scores’ 7 \nThe SDQ yields five subscales: conduct problems, emotional problems, \nhyperactivity, peer problems, and prosocial behaviour. The first four problem \nsubscales were summed to create a total problems score at each age. \nBetween the ages 2 to 4, the Preschool Behaviour Questionnaire (Behar \nscales) was used to measure parent-reported behaviour problems (Behar, 1977), \nwith items converted to SDQ-comparable components by TEDS researchers \n(https://datadictionary.teds.ac.uk/pdfs/4yr/234yr_behaviour_items.pdf). From age 7 \nonward, the standard 25-item SDQ was administered with multiple informants: \nparent reports at ages 7, 9, 12, 16, and 21 (emotional and peer problem subscales \nwere not available from parents at age 16); teacher reports at ages 7, 9, and 12; and \nself-reports at ages 12, 16, 21, and 26.  \nWe also included measures of anxiety and ADHD symptoms collected at \nmultiple ages. Additional behaviour problems assessed at one or two ages were \nexamined as outcomes; these results are presented in the Appendices. \n \nSocioeconomic Status (SES). Family SES was assessed at birth and at \nages 7, 16, and 21. Each SES composite was standardised as z-scores and \ncalculated from parental employment status (coded according to the UK Standard \nOccupational Classification or SOC, \nhttps://www.ons.gov.uk/methodology/classificationsandstandards/standardoccupatio\nnalclassificationsoc/), parental highest educational qualifications, and household \nincome. \n \nGenetic Measures  \nGenotyping for the TEDS participants was conducted on one of two platforms: \nthe Affymetrix Genome-Wide Human SNP Array 6.0 and the Illumina \nHumanOmniExpressExome-8v1.2. DNA was obtained from either buccal cheek \nswabs or saliva samples collected over several waves. \nFollowing quality control, genotypes from both platforms were separately \nphased using EAGLE2 and then imputed to the Haplotype Reference Consortium \n(release 1.1) (Durbin, 2014; Loh et al., 2016; McCarthy, 2016). After imputation, \nharmonisation, and merging of the two datasets, a final set of 7,363,646 SNPs for \n10,346 twins remained for analysis. More details of the genotyping and imputation \nprocesses are described in previous TEDS publications (Lin et al., 2025; Selzam et \nal., 2018).  \nWe constructed polygenic scores using LDpred2-auto, a Bayesian method \nthat adjusts GWAS summary statistics for linkage disequilibrium (LD) using the \nHapMap3+ reference panel (Privé et al., 2021). Approximately 1.1 million SNPs \ncommon between the TEDS sample and the HapMap3+ panel were included. The \nmost recent GWAS summary statistics of educational attainment (EA4) and \nintelligence (IQ3) were used (Okbay et al., 2022; Savage et al., 2018).  \nTo maximise the prediction of general cognitive ability, we combined the EA4 \nand IQ3 polygenic scores using SMTPred, which applies an ordinary least squares \n(OLS) weighting approach (Maier et al., 2018). In our sample, EA4 was weighted \n0.08 and IQ3 0.05. The resulting combined score, termed the polygenic g score, \ncorrelated highly with IQ3 (r = 0.81) and EA4 (r = 0.87) polygenic scores in our \nsample. This polygenic g score was used as the primary predictor for all cognitive \nabilities, educational achievement, behaviour problems, and anthropometric \noutcomes in the present study. \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted December 23, 2025. ; https://doi.org/10.64898/2025.12.19.695378doi: bioRxiv preprint \n\nCharting cognitive development using adult ‘polygenic g scores’ 8 \nStatistical Analyses  \nThe present study was pre-registered at https://osf.io/vhuyj/overview. \nAnalyses were performed using RStudio 2023.09.1+494 with codes available on \nGitHub (https://github.com/YujingLinn/Cog-PGS).  \nBefore carrying out the main analyses, we conducted sensitivity tests to \nexamine potential effects of age, sex, zygosity, twin birth order, genotyping chip, and \nthe first ten genomic principal components (PCs) on all phenotypes and family \nsocioeconomic status measured across development. Sensitivity analyses of the \npolygenic g score were also conducted, except for age effects. Results are detailed \nin Appendix Table B1. \nSensitivity analyses revealed significant associations between several \ncovariates and the phenotypes. Specifically, age, sex, zygosity, genotyping chip, and \nthe first ten genomic PCs showed significant effects on about half of the phenotypes. \nTwin birth order showed no significant effects. For the polygenic g score, significant \neffects were detected for zygosity, genotyping chip, and the ninth and tenth PCs. \nThus, we adopted a conservative approach by including age, sex, genotyping chip, \nand the first ten genomic PCs as covariates in all cross-sectional analyses for \nconsistency, even though not all phenotypes were significantly associated with every \ncovariate. Age was excluded as a covariate in longitudinal analyses. Since including \nzygosity as a covariate is relatively unconventional, we performed additional \nanalyses stratifying the sample into monozygotic and dizygotic twins to examine \nwhether the polygenic g score predictions were robust across zygosity groups. \n \nPolygenic Score Prediction  \nOur main analysis was to examine associations between our polygenic g \nscore and outcomes at each age using a subsample of unrelated individuals (one \nrandomly selected twin per pair). All phenotypes were analysed, including both \ncomposite scores and their constituent components or specific test scores. \nAll continuous outcomes were standardised prior to analysis. Models included \nage, sex, genotyping chip, and the first ten PCs as covariates to control for batch \neffects and population structure. We report standardised beta coefficients for the \npolygenic g predictor. Incremental variance explained was calculated as the \ndifference in R² between the full model and a reduced model containing only \ncovariates. Confidence intervals were estimated using percentile bootstrapping with \n1000 iterations. \nTo ensure the robustness of our findings, we repeated all prediction analyses \nseparately among females, males, monozygotic twins, and dizygotic twins. \n \nCommon Factor Analyses  \nNext, we examined the polygenic g score prediction of cross-time common \nfactor of the phenotypes. We extracted cross-time common factors from repeated \nmeasures of the same construct, assuming these reflect stable underlying latent \ntraits. Using multilevel structural equation modelling (SEM) to account for twin \nstructure (i.e., family clustering), we conducted confirmatory factor analyses (CFA) \nfor g, verbal ability, nonverbal ability, SDQ, anxiety, and ADHD. Whereas cross-\nsectional analyses used one randomly selected twin per family to ensure \nindependence, SEM analyses included both twins from each pair to increase \nstatistical power, retaining the full sample while appropriately adjusting for within-\nfamily non-independence. Common factors were extracted for both total scales and \nsubscales where applicable. \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted December 23, 2025. ; https://doi.org/10.64898/2025.12.19.695378doi: bioRxiv preprint \n\nCharting cognitive development using adult ‘polygenic g scores’ 9 \nFor cognitive and educational outcomes, TEDS employed the most \nappropriate informant depending on the developmental stage. Cognitive measures at \nearly ages were administered by parents, and then the children took the tests \nthemselves. Educational achievement was mainly evaluated by teachers, while \nnational exam scores were provided by parents. Because different informants were \nused at different ages (e.g., parent-administered tests in early childhood, self-\nadministered tests in adolescence), the common factors for cognitive abilities and \neducational achievement reflect developmental changes over time as well as \npotential method variance associated with different informants. \nFor behaviour outcomes, multiple informants were used at the same age. For \nexample, SDQ was reported by parents at ages 2, 3, 4, 7, 9, 12, 16, and 21; by \nteachers at ages 7, 9, and 12; and via self-report at ages 12, 16, 21, and 26. This \nmulti-informant approach reflects developmental appropriateness: parent reports are \nmost suitable in early childhood when children cannot report reliably, teacher reports \nprovide complementary school-based perspectives during school age, and self-\nreports become increasingly valid as adolescents develop greater self-awareness \nand autonomy. \nWe therefore extracted both cross-rater common factors (combining all \ninformants) and within-rater common factors when three or more assessments from \nthe same informant were available for a given construct. \nWe then used the polygenic g score to predict these common factors and \ncompared predictive validity against the age-specific cross-sectional predictions. \nModels included age, sex, genotyping chip, and the first 10 genomic principal \ncomponents as covariates, with standardised beta coefficients and incremental R² \nreported. All analyses were repeated separately for female and male subsamples. \n \nLatent Growth Curve Model \nWe conducted latent growth curve analyses to examine how the polygenic g \nscore predicts both baseline levels (intercept) and developmental trajectories (slope) \nacross time. A positive association with the intercept indicates that higher polygenic g \nscores predict higher initial levels of the phenotype, while a positive association with \nthe slope indicates that higher polygenic g scores predict steeper increases in the \nphenotype over time. These analyses were conducted for measures with repeated \nassessments across development, including general cognitive ability, verbal and \nnonverbal abilities, educational outcomes, the subscales of the SDQ, anxiety, ADHD, \nheight, and BMI. \nFor measures with multiple potential informants, we prioritised consistency \nacross developmental stages. For most phenotypes, we used parent reports before \nage 18 and self-reports from age 18 onwards. When parent reports were unavailable \nin childhood or adolescence, we prioritised teacher reports, followed by child reports; \nhowever, in most cases, child reports were used when parent reports were \nunavailable. \nMissing data in the repeated measures were handled using Full Information \nMaximum Likelihood within the latent growth curve models (Enders & Bandalos, \n2001). This approach uses all available data points for each participant, allowing \ninclusion of individuals with partially missing timepoints. Participants were included if \nthey had data available for at least two timepoints. No additional imputation was \nperformed for other variables included in the models. \nLike the confirmatory factor analyses, all latent growth models used the full \ntwin sample with multilevel modelling to account for family clustering, maximising \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted December 23, 2025. ; https://doi.org/10.64898/2025.12.19.695378doi: bioRxiv preprint \n\nCharting cognitive development using adult ‘polygenic g scores’ 10 \nstatistical power. Sex was included as a covariate in models for the whole sample. \nWe also conducted multi-group analyses to compare intercepts and slopes between \nfemales and males, with family clustering accounted for within each sex-stratified \nsample. \n \nProfile Analysis of Extremely High and Extremely Low Polygenic g Scores  \nWe examined the developmental trajectories of participants with extreme \npolygenic g scores, defined as scores above or below three standard deviations from \nthe population mean. \nTo maximise sample size, we assigned polygenic g scores to the MZ co-twins \nof genotyped individuals, as only one twin per MZ pair was genotyped because MZ \ntwins share identical genomes. This yielded 19 participants with scores three \nstandard deviations above the mean and 14 participants with scores three standard \ndeviations below the mean. \nWe plotted the observed values of cognitive, educational, and behaviour \nproblem outcomes for these extreme groups across development (standardised for \ncomparability). Socioeconomic status was also plotted alongside the trajectories for \ncontext, rather than included as a covariate, given the potential circularity between \npolygenic g scores and SES. \nTo formally compare differences in outcomes across development between \nextreme groups, we divided the full sample into deciles based on polygenic g scores \nand conducted independent samples t-tests comparing the top and bottom deciles at \neach age. \n \nNonlinearity Tests \nFinally, to test whether the relationship between polygenic g scores and \noutcomes is linear, we conducted regression analyses including a quadratic term for \nthe polygenic g score. A significant quadratic term would indicate nonlinearity, \nrepresenting either accelerating or decelerating effects at the extremes of the \ndistribution. Conversely, a non-significant quadratic term would support linearity, \nsuggesting that high and low extremes differ only quantitatively, not qualitatively, \nfrom the rest of the distribution. \n \nMultiple Testing Correction  \nWe applied the false discovery rate (FDR) correction to account for multiple \ntesting across all regression analyses. All reported p-values are FDR-adjusted \nvalues. \n  \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted December 23, 2025. ; https://doi.org/10.64898/2025.12.19.695378doi: bioRxiv preprint \n\nCharting cognitive development using adult ‘polygenic g scores’ 11 \nResults  \nWe focus primarily on cognitive, educational, and behaviour problem \nphenotypes in this section, with other phenotypes discussed as relevant. Full results \nfor all phenotypes, including anxiety, ADHD, anthropometric outcomes, and those \nassessed at just one or two ages, are reported in the appendices. Descriptive \nstatistics for all phenotypes and the polygenic g score are shown in Appendix Table \nB2 and correlation matrices in Appendix Figure C1.  \n \n1a) \n \n1b) \n \n1c) \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted December 23, 2025. ; https://doi.org/10.64898/2025.12.19.695378doi: bioRxiv preprint \n\nCharting cognitive development using adult ‘polygenic g scores’ 12 \n \nFigure 1. Polygenic g score prediction of cognitive abilities, educational \nachievement, and behaviour problems across development. \nStandardised beta (β) coefficients with 95% bootstrapped confidence intervals \n(1000 iterations) for the prediction of developmental outcomes by polygenic g \nscore. Analyses were conducted in the unrelated sample by randomly \nselecting one twin from each pair. Panel a: General cognitive ability (g) \ncomposite, and domain-specific verbal and nonverbal ability composites \nmeasured from ages 2 to 25. Panel b: Educational achievement in English, \nmathematics, and science, and a core-subject composite. Science was first \nmeasured at age 9, while English and mathematics were also assessed at \nage 7. The core-subject composite was derived from English, mathematics, \nand science (where available) from ages 7 to 18, with general university \ngrades used at age 21 and years of schooling at age 26. For plotting \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted December 23, 2025. ; https://doi.org/10.64898/2025.12.19.695378doi: bioRxiv preprint \n\nCharting cognitive development using adult ‘polygenic g scores’ 13 \npurposes, 'core subject' serves as an overarching category representing \neducational achievement and attainment through age 26. Panel c: Strengths \nand Difficulties Questionnaire (SDQ) five subscales (conduct problems, \nemotion problems, hyperactivity, peer problems, and prosocial behaviour) and \ntotal problems score measured from ages 2 to 26. For SDQ, the rater is \nindicated by a symbol shape (parent, teacher, or child). For other outcomes, \nperformance-based measures were used, with parent ratings at early ages \nand self-reports or test-based assessments at later ages. The vertical dashed \nline at β = 0 represents no association. Asterisks denote statistical \nsignificance: * p < 0.05, ** p < 0.01, *** p < 0.001, **** p < 0.0001. Complete \nnumerical results, including sex-stratified and zygosity-stratified analyses, are \nreported in Appendix Table B3. Individual tests comprising the cognitive \ncomposites are presented in Appendix Figures C2 and C3. \n \nPolygenic g Score Prediction from Infancy to Early Adulthood  \nPolygenic g score significantly predicted most outcomes across phenotypes at \nmost ages from childhood onwards. As shown in Figure 1a, for cognitive abilities, the \nprediction was weak or absent in infancy (ages 2 to 4). From childhood, the \nprediction increased steadily and reached a peak in early adulthood (age 25) with \nstandardised beta coefficients of 0.35 (95% CI [0.31, 0.39]) for both g and verbal \nability, and 0.28 [0.24, 0.31] for nonverbal ability. \nWe also examined polygenic g score prediction for the individual verbal and \nnonverbal tests used to construct the cognitive ability composites (see Appendix \nFigures C2 and C3). Similar patterns as in the composites were identified: weak or \nabsent prediction in infancy that strengthened with age. Among verbal tests \n(Appendix Figure C2), the strongest prediction emerged for the Verbal Reasoning \ntask from the Pathfinder battery at age 25 (βstandardised = 0.32 [0.28, 0.36]). Among \nnonverbal tests (Appendix Figure C3), the strongest prediction was observed for the \nUnderstanding Number task at age 16 (βstandardised = 0.31 [0.27, 0.35]).  \nFor educational outcomes (see Figure 1b), polygenic g score associations \nwere moderate at age 7 (βstandardised ≈ 0.25), increased through mid-adolescence. \nPrediction strength peaked at age 16 for GCSE grades: English (β = 0.33 [0.30, \n0.38]), mathematics (βstandardised = 0.38 [0.34, 0.42]), science (βstandardised = 0.38 [0.34, \n0.42]), and the core-subject composite (βstandardised = 0.39 [0.35, 0.43]). Polygenic g \nscore associations declined for A-level, university grades and years of schooling. \nFor behaviour problems, we focused on SDQ measures (Figure 1c). Negative \nassociations were observed for most behaviour problem measures across ages and \nraters. Significant predictions emerged in infancy for all SDQ subscales. For conduct \nproblems and hyperactivity, associations strengthened slightly and peaked at \napproximately βstandardised = -0.12 and -0.13, respectively, in late childhood and \nadolescence, then weakened in early adulthood—becoming non-significant for \nhyperactivity but remaining significant for conduct problems. For emotional and peer \nproblems, associations were largely non-significant from infancy to adolescence but \nbecame significantly negative though weak in adulthood. We also examined \nassociations between the polygenic g score and the total problem scale. A consistent \npattern of significant weak negative associations was observed across ages and \nraters, ranging from βstandardised = -0.06 to -0.13. \nFor prosocial behaviour, the polygenic g score explained negligible variance \n(all <0.5%) after accounting for age and sex effects. Associations were generally \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted December 23, 2025. ; https://doi.org/10.64898/2025.12.19.695378doi: bioRxiv preprint \n\nCharting cognitive development using adult ‘polygenic g scores’ 14 \nweak and inconsistent across raters and ages, with some being slightly negative \nearly in life.  \nFor other behaviour problems (anxiety in Appendix Figure C4; ADHD in \nAppendix Figure C5), weak negative associations were observed from infancy \nonward. Interesting patterns emerged for height and BMI, with non-significant \nassociations until early adulthood, then positive associations with height and \nnegative associations with BMI (Appendix Figures C6 and C7). For other \nmiscellaneous outcomes, associations were largely positive for education- and \ncognitive-related phenotypes and showed mixed patterns for wellbeing-related \nphenotypes (Appendix Figure C8). \n \nPolygenic g Score Prediction of Cross-Time and Cross-Rater Latent Factor \nWe extracted cross-time latent factors to systematically examine polygenic g \nscore prediction across age and raters. For parent, teacher, and child reports of \nbehaviour problems, we extracted both cross-rater and within-rater latent factors. \nAll observed variables loaded significantly onto their respective latent \nconstructs (Appendix Table B4), with standardised loadings ranging from 0.27 to \n0.83. Model fit indices are provided in Appendix Table B5. For cognitive abilities \n(Appendix Figure C9), model fit was suboptimal compared to behaviour problems, \nyet predictions for the latent factors matched the strongest individual measure \npredictions for g (βstandardised = 0.39 [0.37, 0.41]), verbal ability (0.36 [0.34, 0.38]), and \nnonverbal ability (0.33 [0.31, 0.35). \nFor behaviour problems, model fits were mostly good. Cross-rater predictions \nshowed that the polygenic g score was negatively associated with total SDQ \nproblems (βstandardised = -0.19), conduct problems (-0.21), hyperactivity (-0.20), \nemotional problems (-0.11), peer problems (-0.08), and prosocial behaviour (-0.05; \nAppendix Figure C9). Cross-rater estimates approximated the average of within-rater \nestimates, with no clear pattern about which rater showed stronger prediction. \nWithin-rater models tended to show slightly better fit, which is expected given that \ncombining ratings across informants introduces additional method variance due to \ninformant effects.(Achenbach et al., 1987; Glaser et al., 1997). Standardised path \ndiagrams for all 39 models appear in Appendix Figure C10. \n \nPolygenic g Score Prediction of Developmental Trajectories  \nBeyond examining associations at individual ages and cross-time common \nfactors, we used latent growth curve models to examine how the polygenic g score \npredicts both baseline levels (intercepts) and rates of developmental change \n(slopes). Only longitudinal measures were included for cognitive abilities, educational \nachievement, behaviour problem and anthropometric outcomes. The full results are \npresented in Appendix Table B6 for the full sample and in Appendix Table B7 for the \nsex-stratified sample. \nAt the baseline, the polygenic g score already displayed significant positive \nassociations with cognitive, educational, and anthropometric outcomes. The \nstrongest baseline predictions were for science achievement and core-subject \nachievement (both βstandardised = 0.35), followed by cognitive composites: g and verbal \nability (both 0.10) and nonverbal ability (0.09). In contrast, all behaviour problem \nintercepts showed negative associations ranging from -0.02 to -0.17, indicating that \nchildren with higher polygenic g scores exhibited fewer behaviour problems. \nAcross development, the polygenic g score predicted steeper developmental \nincreases for most cognitive and educational outcomes, including both g and verbal \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted December 23, 2025. ; https://doi.org/10.64898/2025.12.19.695378doi: bioRxiv preprint \n\nCharting cognitive development using adult ‘polygenic g scores’ 15 \nability (βstandardised = 0.31), which were followed by science achievement (0.28), \nnonverbal ability (0.27), and core-subject achievement (0.13). Together, the positive \nbaseline and growth effects for cognitive and educational outcomes implied that \nchildren with higher polygenic g scores started with higher cognitive and educational \nperformance and continued to increase at a faster rate across development.  \nFor behaviour problems, although half of the growth effects were negative and \nnon-significant, the significant effects were primarily positive. For example, positive \nassociations were found for ADHD and its subscales (βstandardised = 0.09 to 0.11), for \nARBQ negative affect (0.06), and for SDQ hyperactivity (0.07). Peer problems \nshowed the only significant negative slope (-0.09). Together with the negative \nbaseline effects, the positive growth effects suggested that children with higher \npolygenic g scores begin with fewer behaviour problems but show slightly larger \nincreases over time, gradually moving closer to the average developmental \ntrajectories. In contrast, we found the opposite for BMI, which yielded a combination \nof a positive intercept (0.08) and a negative slope (-0.13), indicating that children \nwith higher polygenic g scores started with a higher BMI, which increased at a slower \nrate with age.  \nAdditionally, across models, consistently negative correlations between \nintercepts and slopes were identified, ranging from -0.31 to -0.68, indicating that \nindividuals with higher baseline levels tended to show slower rates of increase over \ntime. \n \nPolygenic g Score Prediction Across the Distribution \nTo examine whether polygenic g score prediction varies across the \ndistribution, we first tested for nonlinear effects by including both linear and quadratic \nterms in the same model. Across all outcomes, no significant quadratic effects were \nobserved in either the full sample or sex-stratified analyses (all incremental R² < \n0.01, all FDR-adjusted p > 0.05; Appendix Table B8). These results indicate that \nassociations are linear throughout the distribution, with no evidence of stronger or \nweaker effects at the ends of the polygenic score distribution. \nSecond, we compared phenotypic outcomes between individuals at the \ndistributional tails—those in the top versus bottom polygenic g score deciles \n(Appendix Table B9). For cognitive abilities, individuals at the top of the distribution \nconsistently scored significantly higher than those at the bottom from around age 7 \nonwards, with occasional exceptions (e.g., spatial ability in adolescence and \nadulthood). Differences were non-significant in infancy. For educational achievement, \nall comparisons between distributional extremes were statistically significant. For \nbehaviour problems, approximately half of the phenotypes showed non-significant \ndifferences between the top and bottom deciles. For example, peer problems across \nmost ages and raters, as well as height and BMI, showed no significant differences \nbetween top and bottom deciles. The magnitude of decile differences generally \ncorresponded to population-level prediction strength. Outcomes with stronger overall \nassociations (such as cross-age and cross-rater composite scores) consistently \nshowed significant differences between the top and the bottom polygenic g score \ndeciles.  \n \nPolygenic g Score Prediction for the Highest and Lowest Individuals  \nTo examine outcomes at the distributional extremes in greater detail, we \nidentified individuals with the highest (N = 19) and lowest (N = 14) polygenic g \nscores (Figure 2). Sample characteristics are presented in Appendix Table B2 and \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted December 23, 2025. ; https://doi.org/10.64898/2025.12.19.695378doi: bioRxiv preprint \n\nCharting cognitive development using adult ‘polygenic g scores’ 16 \ngroup comparisons in Appendix Table B10. Due to small sample sizes, most \ncomparisons were not statistically significant. Notably, SES differed significantly \nbetween groups at all ages. \nFor cognitive abilities, both groups initially performed within one standard \ndeviation (SD) of the population mean in infancy and early childhood (Figures 2a and \n2d). By middle childhood, the high polygenic score group had IQ-equivalent scores \nof about 110. The low polygenic score group had IQ-equivalent scores between 80 \nand 90 in middle childhood but experienced substantial attrition; by adulthood, only \none individual remained in the study. \nEducational achievement yielded similar developmental trends. The high \npolygenic score group (Figure 2b) performed about one SD above the population \nmean, while the low polygenic score group performed about one SD below the mean \n(Figure 2e). For behaviour problems, neither group differed significantly from the \npopulation mean or from each other (Figures 2c and 2f). Mean trajectory \ncomparisons for the additional phenotypic groups are presented in Appendix Figure \nC11. Individual trajectories with SES annotations for all participants with the highest \nand lowest polygenic g scores are shown in Appendix Figures C12 and C13, \nrespectively.  \n \n \nFigure 2. Mean developmental trajectories for individuals with the \nhighest and lowest polygenic g scores. Mean trajectories across \ndevelopment for individuals with polygenic g scores >145 (panels a-c) and \n<55 (panels d-f). Panel a/d: General cognitive ability (g) composite and \ndomain-specific composites (nonverbal and verbal ability). Panel b/e: \neducational achievement outcomes including English, mathematics, science, \nand core-subject composite (sum of English, mathematics, and science). \nPanel c/f: Strengths and Difficulties Questionnaire (SDQ) five subscales and \ntotal problem score. All measures are standardised to mean = 0, SD = 1, \nexcept for cognitive ability measures (mean = 100, SD = 15). Shaded areas \nrepresent 95% confidence intervals; absence of confidence intervals indicates \nthat only one individual was measured at that age. For phenotypes with \nmultiple raters, one rater per age is used: parent ratings before age 18 and \nchild/self-ratings at age 18 and older. Black horizontal line at y = 0 represents \nthe population average. \n \nPolygenic g Score Prediction Among Subsamples  \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted December 23, 2025. ; https://doi.org/10.64898/2025.12.19.695378doi: bioRxiv preprint \n\nCharting cognitive development using adult ‘polygenic g scores’ 17 \nPolygenic g score predictions showed high consistency across sexes at each \nage, the latent factor across ages, developmental trajectories, across distributions, \nand among the most extreme scorers. Estimates were typically within one to two \nstandard errors of each other, well within the range of overlapping 95% confidence \nintervals. Cognitive composites and individual cognitive tests showed nearly identical \npredictions for males and females. Educational outcomes showed similar patterns for \nboth sexes, with predictions increasing through high school and declining thereafter. \nFor behaviour problems, predictions were also comparable between sexes, with only \noccasional deviations (e.g., peer problems at age 21 showed βstandardised = -0.17 for \nfemales and -0.06 for males). On average, the differences in standardised beta \ncoefficients between sexes were smaller than 0.01.  \nZygosity comparisons likewise revealed minimal differences. Predictions were \nhighly similar for monozygotic and dizygotic twins, with an average difference in \nstandardised beta coefficients of only 0.01 across all phenotypes. The largest \nobserved difference was for English achievement at age 16 (βstandardised = 0.37 for \nmonozygotic twins vs. 0.24 for dizygotic twins), though such differences were rare \nexceptions rather than a consistent pattern. \n  \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted December 23, 2025. ; https://doi.org/10.64898/2025.12.19.695378doi: bioRxiv preprint \n\nCharting cognitive development using adult ‘polygenic g scores’ 18 \nDiscussion \nPolygenic scores derived from adult intelligence and educational attainment \ncontinue to be the strongest predictors of cognitive development. By combining IQ3 \nand EA4, we created an adult polygenic g score to chart cognitive development from \ninfancy (age 2) to early adulthood (age 26). Our polygenic g score improved \nprediction from 9% variance explained using IQ3 alone to 12% for g at age 25. \nPrediction was even higher for cross-time latent factors, reaching 15% for g. \nAdditionally, for educational achievement, the polygenic g score predicted up to 16% \nof variance at age 16, and up to 2% for behaviour problems and anthropometric \noutcomes.  \nThe polygenic g score predicts cognitive outcomes as early as age 4 \n(βstandardised = 0.07 to 0.09). Although the prediction was largely negligible in infancy, it \nincreased steadily from childhood to early adulthood, peaking at βstandardised = 0.35 at \nage 25, supporting our first two hypotheses on the significant and increased \npolygenic score prediction with age for cognitive abilities. These age-related \nincreases in PGS prediction are consistent with both twin and genomic studies. For \nexample, twin heritability estimated in TEDS increased from 41% at age 9 to 66% at \nage 17 (Haworth et al., 2010). Another study extends the finding into late midlife and \nshows that cognitive heritability remains stable since young adulthood (Lyons et al., \n2017). A similar pattern was identified in genomic studies using PGS (Allegrini et al., \n2019; Gustavson et al., 2025). This pattern is known as the Wilson effect, which \nrefers to the increasing heritability of IQ with age (Bouchard, 2013).  \nSeveral mechanisms help explain this phenomenon. First, discovery GWA \nstudies of cognitive abilities are conducted in adults; thus, the SNP effect estimates \nin these GWA studies more closely match the genetic architecture of cognitive \noutcomes in later developmental stages than in infancy or early childhood. Second, \nthe phenotypic measures become more stable and reliably assessed with age, which \nlimits how well the polygenic g score derived from GWAS of adults can predict \ncognitive abilities earlier in life. For example, the phenotypic correlation of \nintelligence between ages 2 and 25 was modest (r = 0.11), but increased \nsubstantially by late adolescence, reaching r = 0.65 between ages 16 and 25. Third, \nincreasing active and evocative gene–environment correlation amplifies genetic \neffects over time, as individuals increasingly select, evoke, and create environments \nthat align with their genetic propensities (Plomin, 1994). \nDespite potential constraints in early-age measurement, the polygenic g score \nis still one of the best predictors of cognitive development. By combining IQ3 and \nEA4, we introduced additional genetic signals relevant to cognitive abilities (Plomin & \nvon Stumm, 2018), adding 3% variance explained from 9% using IQ3 alone to 12% \nin predicting g at age 25. We achieve even higher predictive power by extracting a \ncross-time common latent factor of g, verbal ability, and nonverbal ability—15%, \n13%, and 11%, respectively. This is to be expected because genetic influences on \ncognition are stable across development (Gustavson et al., 2025). The latent factors \nhad substantial loadings from all ages (0.45–0.72), indicating that even the earliest \ncognitive assessments contribute meaningfully to this stable variance. Latent factors \naggregate information across repeated measurements, reducing age-specific \nmeasurement error and isolating the stable component of cognitive variation. \nIn addition, our latent growth curve models captured both baseline differences \nand developmental changes. Children with higher polygenic g scores showed higher \nbaseline cognitive performance and steeper growth rates, consistent with increasing \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted December 23, 2025. ; https://doi.org/10.64898/2025.12.19.695378doi: bioRxiv preprint \n\nCharting cognitive development using adult ‘polygenic g scores’ 19 \ngene–environment correlation over time. These findings illustrate that the influence \nof polygenic g is not limited to predicting outcomes at isolated ages but also \ncharacterises the overall developmental trajectory of cognitive abilities from early \nchildhood to adulthood. \nWe next examined educational achievement, behaviour problems, and \nanthropometric measures. Polygenic g score predicted most trait domains, \nsupporting the third hypothesis. The strongest prediction was for educational \nachievement, consistent with previous research (Wilding et al., 2024). Prediction of \neducational increased until the end of secondary school, peaking at 16% for GCSE \nachievement, and then declined to approximately 7% for years of schooling at age \n26. Similar developmental patterns were reported in independent research, with the \nlater decline attributed to range restriction because after secondary school, only a \nsubset of the sample entered higher education, reducing variability and attenuating \nprediction (Kvalvik et al., 2025).  \nAs with cognitive traits, polygenic g predicted both higher baseline educational \nachievement and steeper growth than cognitive abilities. In contrast, predictions for \nbehaviour problems and anthropometric outcomes were negligible to weak and \nshowed less consistency. Behaviour problems showed negative baseline \nassociations and small positive slope estimates, indicating that children with higher \npolygenic g scores began with fewer behaviour problems but exhibited slightly \ngreater increases over time. This pattern might reflect regression to the mean rather \nthan substantive developmental change. No consistent sex differences were \nobserved. \nFurthermore, prediction across the distribution of polygenic g scores was \nlinear, in line with our fourth hypothesis. The absence of significant quadratic effects \nindicates that individuals at the end of the distribution differ quantitatively instead of \nqualitatively from the rest of the sample. At the distributional extremes of the \npolygenic g scores (+/–three standard deviations), individuals who were three \nstandard deviations above the mean showed, on average, around one to two \nstandard deviations advantage in cognitive and educational outcomes in adulthood. \nBehaviour problems and anthropometric outcomes showed almost no meaningful \ndifferences between extremes. Finally, consistent with our fifth hypothesis, we found \nno evidence for sex differences in the prediction of the polygenic g score across \noutcomes. \nOur study has several limitations. First, prediction for nonverbal ability was \nweaker than for verbal ability, likely because GWA studies are more heavily \nweighted toward verbal measures. Even after incorporating EA4, some genetic \nsignals for nonverbal ability may not be fully captured (Procopio et al., 2025). Future \nGWA studies could benefit from a greater balance in verbal and nonverbal abilities. \nSecond, our g composite was constructed by averaging verbal and nonverbal \nabilities. Although this composite performed well, hierarchical phenotypic models \nwould allow us to estimate g more precisely from the underlying tests and could \nimprove prediction. Third, discovery GWA studies rely on common variants. Recent \nevidence suggests that rare genetic variants contribute substantially to cognitive \ndifferences and could improve prediction with whole-genome sequencing \n(Wainschtein et al., 2025). Finally, caution is needed when applying composite \npolygenic g scores in within-family contexts. Both IQ3 and EA4 contain family-level \ngenetic contributions, meaning prediction may be attenuated among siblings (Lin et \nal., 2025).  \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted December 23, 2025. ; https://doi.org/10.64898/2025.12.19.695378doi: bioRxiv preprint \n\nCharting cognitive development using adult ‘polygenic g scores’ 20 \nAs twin analyses suggested an increased heritability of intelligence and \ncognitive abilities from childhood to young adulthood, molecular genetic studies have \nmade use of this finding by focusing on the genetics of adults. While twin designs \nrely on phenotypic measurements at each developmental stage to estimate \nheritability, GWA- and PGS-based approaches can quantify genetic liability from \nconception onward. After decades of research, we are now able to go back to the flip \nside using adult molecular genetics outcomes to chart child cognitive development \nas it is linked genetically to adult cognitive abilities. Our results show that prediction \nbegins as early as age 4, increases steadily through development, and is especially \npowerful in estimating growth trajectories. Importantly, this predictive utility does not \ndepend on a causal explanation of individual SNP effects, but rather on the \naggregation of genetic signals associated with the phenotype (Lin et al., 2025; \nPlomin & von Stumm, 2022).  \nTogether, we show that combining two highly predictive polygenic scores \nyields substantial insights into cognitive developmental trajectories from infancy to \nearly adulthood. Outperforming other early-life predictors, polygenic scores fixed at \nconception continue to be one of the strongest predictors available for long-term \ncognitive development. \n  \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted December 23, 2025. ; https://doi.org/10.64898/2025.12.19.695378doi: bioRxiv preprint \n\nCharting cognitive development using adult ‘polygenic g scores’ 21 \nAcknowledgments \nWe gratefully acknowledge the ongoing contribution of the participants in the \nTwins Early Development Study (TEDS) and their families. TEDS is supported by the \nUK Medical Research Council (MR/V012878/1 and previously MR/M021475/1), with \nadditional support from the US National Institutes of Health (AG046938). Y.L. is \nsupported by a KCL-CSC joint scholarship awarded by King’s College London and \nthe China Scholarship Council. \nFor the purposes of open access, the author has applied a Creative \nCommons Attribution (CC BY) license to any Accepted Author Manuscript version \narising from this submission. \n \n  \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted December 23, 2025. ; https://doi.org/10.64898/2025.12.19.695378doi: bioRxiv preprint \n\nCharting cognitive development using adult ‘polygenic g scores’ 22 \nReferences \n \nAchenbach, T. M., McConaughy, S. H., & Howell, C. T. (1987). Child/adolescent \nbehavioral and emotional problems: Implications of cross-informant \ncorrelations for situational specificity. Psychological Bulletin, 101(2), 213–232. \nhttps://doi.org/10.1037/0033-2909.101.2.213 \nAllegrini, A. G., Selzam, S., Rimfeld, K., von Stumm, S., Pingault, J. B., & Plomin, R. \n(2019). Genomic prediction of cognitive traits in childhood and adolescence. \nMolecular Psychiatry, 24(6), 819–827. https://doi.org/10.1038/s41380-019-\n0394-4 \nAsbury, K., Wachs, T. D., & Plomin, R. (2005). Environmental moderators of genetic \ninfluence on verbal and nonverbal abilities in early childhood. Intelligence, \n33(6), 643–661. https://doi.org/10.1016/j.intell.2005.03.008 \nBayley, N. (1993). Bayley Scales of Infant Development. San Antonio: Psychological \nCorporation. \nBehar, L. B. (1977). The Preschool Behavior Questionnaire. Journal of Abnormal \nChild Psychology, 5(3), 265–275. https://doi.org/10.1007/BF00913697 \nBouchard, T. J. (2013). The Wilson Effect: The Increase in Heritability of IQ With Age. \nTwin Research and Human Genetics, 16(5), 923–930. \nhttps://doi.org/10.1017/thg.2013.54 \nBreit, M., Scherrer, V., Tucker-Drob, E. M., & Preckel, F. (2024). The stability of \ncognitive abilities: A meta-analytic review of longitudinal studies. \nPsychological Bulletin, 150(4), 399–439. https://doi.org/10.1037/bul0000425 \nDeary, I. J. (2012). Intelligence. Annual Review of Psychology, 63(1), 453–482. \nhttps://doi.org/10.1146/annurev-psych-120710-100353 \nDionne, G., Dale, P. S., Boivin, M., & Plomin, R. (2003). Genetic Evidence for \nBidirectional Effects of Early Lexical and Grammatical Development. Child \nDevelopment, 74(2), 394–412. https://doi.org/10.1111/1467-8624.7402005 \nDurbin, R. (2014). Efficient haplotype matching and storage using the positional \nBurrows–Wheeler transform (PBWT). Bioinformatics, 30(9), 1266–1272. \nhttps://doi.org/10.1093/bioinformatics/btu014 \nEnders, C. K., & Bandalos, D. L. (2001). The Relative Performance of Full \nInformation Maximum Likelihood Estimation for Missing Data in Structural \nEquation Models. Structural Equation Modeling: A Multidisciplinary Journal, \n8(3), 430–457. https://doi.org/10.1207/S15328007SEM0803_5 \nFenson, L., Pethick, S., Renda, C., Cox, J. L., Dale, P. S., & Reznick, J. S. (2000). \nShort-form versions of the MacArthur Communicative Development \nInventories. Applied Psycholinguistics, 21(1), 95–116. \nhttps://doi.org/10.1017/S0142716400001053 \nFenson, L., Philip, S., Reznick, J. S., Thal, D. J., Bates, E., Hartung, J. P., Pethick, \nS., & Reilly, J. S. (1993). The MacArthur communicative development \ninventories. Users guide and technical manual. San Diego: Singular \nPublishing Group. \nGlaser, B. A., Kronsnoble, K. M., & Forkner, C. B. W. (1997). Parents and Teachers \nas Raters of Children’s Problem Behaviors. Child & Family Behavior Therapy, \n19(4), 1–13. https://doi.org/10.1300/J019v19n04_01 \nGOAL plc (2002). (n.d.). GOAL Formative Assessment: Key Stage 3. London: \nHodder & Stoughton. \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted December 23, 2025. ; https://doi.org/10.64898/2025.12.19.695378doi: bioRxiv preprint \n\nCharting cognitive development using adult ‘polygenic g scores’ 23 \nGoodman, R. (1997). The Strengths and Difficulties Questionnaire: A Research Note. \nJournal of Child Psychology and Psychiatry, 38(5), 581–586. \nhttps://doi.org/10.1111/j.1469-7610.1997.tb01545.x \nGustavson, D. E., Borriello, G. A., Karhadkar, M. A., Rhee, S. H., Corley, R. P., Rhea, \nS.-A., DiLalla, L. F., Wadsworth, S. J., Friedman, N. P., & Reynolds, C. A. \n(2025). Stability of general cognitive ability from infancy to adulthood: A \ncombined twin and genomic investigation. Proceedings of the National \nAcademy of Sciences, 122(21), e2426531122. \nhttps://doi.org/10.1073/pnas.2426531122 \nHammill, D. D., Brown, V. L., Larsen, S. C., & Wiederholt, J. L. (1994). Test of \nAdolescent and Adult Language (TOAL-3). Austin, TX, Pro-Ed. \nHaworth, C. M. A., Wright, M. J., Luciano, M., Martin, N. G., de Geus, E. J. C., van \nBeijsterveldt, C. E. M., Bartels, M., Posthuma, D., Boomsma, D. I., Davis, O. \nS. P., Kovas, Y., Corley, R. P., DeFries, J. C., Hewitt, J. K., Olson, R. K., Rhea, \nS.-A., Wadsworth, S. J., Iacono, W. G., McGue, M., … Plomin, R. (2010). The \nheritability of general cognitive ability increases linearly from childhood to \nyoung adulthood. Molecular Psychiatry, 15(11), 1112–1120. \nhttps://doi.org/10.1038/mp.2009.55 \nHayiou-Thomas, M. E., Dale, P. S., & Plomin, R. (2012). The etiology of variation in \nlanguage skills changes with development: A longitudinal twin study of \nlanguage from 2 to 12 years. Developmental Science, 15(2), 233–249. \nhttps://doi.org/10.1111/j.1467-7687.2011.01119.x \nHill, W. D., Marioni, R. E., Maghzian, O., Ritchie, S. J., Hagenaars, S. P., McIntosh, \nA. M., Gale, C. R., Davies, G., & Deary, I. J. (2019). A combined analysis of \ngenetically correlated traits identifies 187 loci and a role for neurogenesis and \nmyelination in intelligence. Molecular Psychiatry, 24(2), 169–181. \nhttps://doi.org/10.1038/s41380-017-0001-5 \nKaplan, E., Fein, D., Kramer, J., Delis, D., & Morris, R. (1999). WISC-III as a Process \nInstrument (WISC-III-PI). New York, NY: Psychological Corp. \nKvalvik, E. H., Wang, Y., Walhovd, K. B., Lyngstad, T. H., & Røgeberg, O. (2025). \nBeyond years of schooling: Shifting genetic influences across educational \nmilestones in two Norwegian cohorts (p. 2025.10.08.680992). bioRxiv. \nhttps://doi.org/10.1101/2025.10.08.680992 \nLin, Y., Procopio, F., Keser, E., Kawakami, K., Rimfeld, K., Malanchini, M., & Plomin, \nR. (2025). Polygenic Score Prediction Within and Between Sibling Pairs for \nIntelligence, Cognitive Abilities, and Educational Traits From Childhood to \nEarly Adulthood. Intelligence & Cognitive Abilities, 1(1). \nhttps://doi.org/10.65550/001c.140654 \nLockhart, C., Bright, J., Ahmadzadeh, Y., Breen, G., Bristow, S., Boyd, A., Downs, J., \nHotopf, M., Palaiologou, E., Rimfeld, K., Maxwell, J., Malanchini, M., \nMcAdams, T. A., McMillan, A., Plomin, R., & Eley, T. C. (2023a). Twins Early \nDevelopment Study (TEDS): A genetically sensitive investigation of mental \nhealth outcomes in the mid‐twenties. JCPP Advances, 3(2), e12154. \nhttps://doi.org/10.1002/jcv2.12154 \nLockhart, C., Bright, J., Ahmadzadeh, Y., Breen, G., Bristow, S., Boyd, A., Downs, J., \nHotopf, M., Palaiologou, E., Rimfeld, K., Maxwell, J., Malanchini, M., \nMcAdams, T. A., McMillan, A., Plomin, R., & Eley, T. C. (2023b). Twins Early \nDevelopment Study (TEDS): A genetically sensitive investigation of mental \nhealth outcomes in the mid‐twenties. JCPP Advances, 3(2), e12154. \nhttps://doi.org/10.1002/jcv2.12154 \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted December 23, 2025. ; https://doi.org/10.64898/2025.12.19.695378doi: bioRxiv preprint \n\nCharting cognitive development using adult ‘polygenic g scores’ 24 \nLoh, P.-R., Danecek, P., Palamara, P. F., Fuchsberger, C., A Reshef, Y., K Finucane, \nH., Schoenherr, S., Forer, L., McCarthy, S., Abecasis, G. R., Durbin, R., & L \nPrice, A. (2016). Reference-based phasing using the Haplotype Reference \nConsortium panel. Nature Genetics, 48(11), 1443–1448. \nhttps://doi.org/10.1038/ng.3679 \nLyons, M. J., Panizzon, M. S., Liu, W., McKenzie, R., Bluestone, N. J., Grant, M. D., \nFranz, C. E., Vuoksimaa, E. P., Toomey, R., Jacobson, K. C., Reynolds, C. A., \nKremen, W. S., & Xian, H. (2017). A longitudinal twin study of general \ncognitive ability over four decades. Developmental Psychology, 53(6), 1170–\n1177. https://doi.org/10.1037/dev0000303 \nMaier, R. M., Zhu, Z., Lee, S. H., Trzaskowski, M., Ruderfer, D. M., Stahl, E. A., \nRipke, S., Wray, N. R., Yang, J., Visscher, P. M., & Robinson, M. R. (2018). \nImproving genetic prediction by leveraging genetic correlations among human \ndiseases and traits. Nature Communications, 9(1), 989. \nhttps://doi.org/10.1038/s41467-017-02769-6 \nMalanchini, M., Rimfeld, K., Gidziela, A., Cheesman, R., Allegrini, A. G., Shakeshaft, \nN., Schofield, K., Packer, A., Ogden, R., McMillan, A., Ritchie, S. J., Dale, P. \nS., Eley, T. C., von Stumm, S., & Plomin, R. (2021). Pathfinder: A gamified \nmeasure to integrate general cognitive ability into the biological, medical, and \nbehavioural sciences. Molecular Psychiatry, 26(12), Article 12. \nhttps://doi.org/10.1038/s41380-021-01300-0 \nMalanchini, M., Rimfeld, K., Shakeshaft, N. G., McMillan, A., Schofield, K. L., Rodic, \nM., Rossi, V., Kovas, Y., Dale, P. S., Tucker-Drob, E. M., & Plomin, R. (2020). \nEvidence for a unitary structure of spatial cognition beyond general \nintelligence. Npj Science of Learning, 5(1), 9. https://doi.org/10.1038/s41539-\n020-0067-8 \nMarkwardt, F. C. (1997). Peabody Individual Achievement Test—Revised, Circle \nPines, MN: American Guidance Service. \nMcCarthy, D. (1972). McCarthy Scales of Children’s Abilities. Psychological \nCorporation. \nMcCarthy, S. (2016). A reference panel of 64,976 haplotypes for genotype \nimputation. Nature Genetics, 48(10), 1279–1283. \nhttps://doi.org/10.1038/ng.3643 \nOkbay, A., Wu, Y., Wang, N., Jayashankar, H., Bennett, M., Nehzati, S. M., \nSidorenko, J., Kweon, H., Goldman, G., Gjorgjieva, T., Jiang, Y., Hicks, B., \nTian, C., Hinds, D. A., Ahlskog, R., Magnusson, P. K. E., Oskarsson, S., \nHayward, C., Campbell, A., … Young, A. I. (2022). Polygenic prediction of \neducational attainment within and between families from genome-wide \nassociation analyses in 3 million individuals. Nature Genetics, 54(4), 437–449. \nhttps://doi.org/10.1038/s41588-022-01016-z \nOliver, B., Dale, P. S., & Plomin, R. (2004). Verbal and nonverbal predictors of early \nlanguage problems: An analysis of twins in early childhood back to infancy. \nJournal of Child Language, 31(3), 609–631. \nhttps://doi.org/10.1017/S0305000904006221 \nPetrill, S. A., Saudino, K. S., Wilkerson, B., & Plomin, R. (2001). Genetic and \nenvironmental molarity and modularity of cognitive functioning in 2-year-old \ntwins. Intelligence, 29(1), 31–43. https://doi.org/10.1016/S0160-\n2896(00)00041-6 \nPlomin, R. (1994). Genetics and experience: The interplay between nature and \nnurture (pp. xvi, 189). Sage Publications, Inc. \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted December 23, 2025. ; https://doi.org/10.64898/2025.12.19.695378doi: bioRxiv preprint \n\nCharting cognitive development using adult ‘polygenic g scores’ 25 \nPlomin, R. (2019). Blueprint: How DNA makes us who we are. Penguin Books. \nPlomin, R., & von Stumm, S. (2018). The new genetics of intelligence. Nature \nReviews Genetics, 19(3), 148–159. https://doi.org/10.1038/nrg.2017.104 \nPlomin, R., & von Stumm, S. (2022). Polygenic scores: Prediction versus \nexplanation. Molecular Psychiatry, 27(1), Article 1. \nhttps://doi.org/10.1038/s41380-021-01348-y \nPrivé, F., Arbel, J., & Vilhjálmsson, B. J. (2021). LDpred2: Better, faster, stronger. \nBioinformatics, 36(22–23), 5424–5431. \nhttps://doi.org/10.1093/bioinformatics/btaa1029 \nProcopio, F., Liao, W., Rimfeld, K., Malanchini, M., von Stumm, S., Allegrini, A. G., & \nPlomin, R. (2025). Multi-polygenic score prediction of mathematics, reading, \nand language abilities independent of general cognitive ability. Molecular \nPsychiatry, 30(2), 414–422. https://doi.org/10.1038/s41380-024-02671-w \nRaven, J. C., Court, J. H., & Raven, J. (1996). Manual for Raven’s Progressive \nMatrices and Vocabulary Scales. Oxford: Oxford University Press. \nRaven, J., Raven, J. C., & Court, J. H. (1998). Mill Hill vocabulary scale. Oxford: \nOPP. \nRietveld, C. A., Esko, T., Davies, G., Pers, T. H., Turley, P., Benyamin, B., Chabris, C. \nF., Emilsson, V., Johnson, A. D., Lee, J. J., Leeuw, C. D., Marioni, R. E., \nMedland, S. E., Miller, M. B., Rostapshova, O., Van Der Lee, S. J., \nVinkhuyzen, A. A. E., Amin, N., Conley, D., … Koellinger, P. D. (2014). \nCommon genetic variants associated with cognitive performance identified \nusing the proxy-phenotype method. Proceedings of the National Academy of \nSciences, 111(38), 13790–13794. https://doi.org/10.1073/pnas.1404623111 \nRimfeld, K., Shakeshaft, N. G., Malanchini, M., Rodic, M., Selzam, S., Schofield, K., \nDale, P. S., Kovas, Y., & Plomin, R. (2017). Phenotypic and genetic evidence \nfor a unifactorial structure of spatial abilities. Proceedings of the National \nAcademy of Sciences, 114(10), 2777–2782. \nhttps://doi.org/10.1073/pnas.1607883114 \nSaudino, K. J., Dale, P. S., Oliver, B., Petrill, S. A., Richardson, V., Rutter, M., \nSimonoff, E., Stevenson, J., & Plomin, R. (1998). The validity of parent‐based \nassessment of the cognitive abilities of 2‐year‐olds. British Journal of \nDevelopmental Psychology, 16(3), 349–362. https://doi.org/10.1111/j.2044-\n835X.1998.tb00757.x \nSavage, J. E., Jansen, P. R., Stringer, S., Watanabe, K., Bryois, J., de Leeuw, C. A., \nNagel, M., Awasthi, S., Barr, P. B., Coleman, J. R. I., Grasby, K. L., \nHammerschlag, A. R., Kaminski, J. A., Karlsson, R., Krapohl, E., Lam, M., \nNygaard, M., Reynolds, C. A., Trampush, J. W., … Posthuma, D. (2018). \nGenome-wide association meta-analysis in 269,867 individuals identifies new \ngenetic and functional links to intelligence. Nature Genetics, 50(7), 912–919. \nhttps://doi.org/10.1038/s41588-018-0152-6 \nSelzam, S., Coleman, J. R. I., Caspi, A., Moffitt, T. E., & Plomin, R. (2018). A \npolygenic p factor for major psychiatric disorders. Translational Psychiatry, \n8(1), 1–9. https://doi.org/10.1038/s41398-018-0217-4 \nSmith, P., Fernandes, C., & Strand, S. (2001). Cognitive Abilities Test 3 (CAT3). \nWindsor: nferNELSON. \nTorgesen, J. K., Wagner, R. K., & Rashotte, C. A. (1999). TOWRE (Test of Word \nReading Efficiency); Austin, Texas: Pro-Ed. \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted December 23, 2025. ; https://doi.org/10.64898/2025.12.19.695378doi: bioRxiv preprint \n\nCharting cognitive development using adult ‘polygenic g scores’ 26 \nVisscher, P. M., Yengo, L., Cox, N. J., & Wray, N. R. (2021). Discovery and \nimplications of polygenicity of common diseases. Science, 373(6562), 1468–\n1473. https://doi.org/10.1126/science.abi8206 \nWainschtein, P., Zhang, Y., Schwartzentruber, J., Kassam, I., Sidorenko, J., Fiziev, P. \nP., Wang, H., McRae, J., Border, R., Zaitlen, N., Sankararaman, S., Goddard, \nM. E., Zeng, J., Visscher, P. M., Farh, K. K.-H., & Yengo, L. (2025). Estimation \nand mapping of the missing heritability of human phenotypes. Nature, 1–9. \nhttps://doi.org/10.1038/s41586-025-09720-6 \nWechsler, D. (1992). Wechsler intelligence scale for children (3rd Ed. UK); The \nPsychological Corporation. \nWiig, E. H., Secord, W., & Sabers, D. (1989). Test of Language Competence—\nExpanded Edition. San Antonio, TX, The Psychological Corporation. \nWilding, K., Wright, M., & von Stumm, S. (2024). Using dna to predict education: A \nmeta-analytic review. Educational Psychology Review, 36(3), 102. \nhttps://doi.org/10.1007/s10648-024-09928-4 \nWoodcock, R. W., McGrew, K. S., & Mather, N. (2001). Woodcock-Johnson III. \nItasca, IL: Riverside Publishing. \n \n.CC-BY 4.0 International licenseperpetuity. It is made available under a \npreprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in \nThe copyright holder for thisthis version posted December 23, 2025. ; https://doi.org/10.64898/2025.12.19.695378doi: bioRxiv preprint","source_license":"CC-BY-4.0","license_restricted":false}