The Impact of Early Adversity and Educational Attainment on Genetic and Brain Morphological Predictors of Intelligence

Intelligence is a strong predictor of occupational achievement, quality of life and physical health. While variation in intelligence is strongly heritable and has been robustly associated with early environment and brain morphology, little is known about how these factors combine and interact to explain this variation in intelligence. To address this, we modelled the relationship between common genetic variation, grey matter volume, early life adversity and education and intelligence in a UK Biobank sample of 16,383 individuals using structural equation modelling. We tested the hypotheses that total grey matter volume would mediate the association between genetics and intelligence, and that early life adversity and educational attainment would moderate this relationship. Common genetic variation was estimated using a genome-wide polygenic score (PGS) of intelligence. Consistent with previous studies, we found that grey matter volume was a significant mediator of the relationship between PGS and intelligence, but only when early life adversity was included in the model. We further found that higher educational attainment in turn moderated the effects of adversity. These findings suggest an important interaction between genetic and environmental factors in the development of intellectual functioning.


Introduction
Intelligence is a strong predictor of mental and physical health, as well as mortality 1 . In the last decade, knowledge about how biological and environmental factors influence intelligence has grown rapidly. This includes knowledge about the contribution of genetic variation and environmental factors, as well as the impact of both on brain morphological differences associated with intelligence. In particular, the recent availability of large-scale data, including that of the UK Biobank, has propelled exciting developments in the area. Despite the significant advances in our understanding of the genetics of intelligence, brain structure and environment, we remain at an early stage of modelling how these factors combine and interact. Addressing this gap in knowledge is critical because maintaining intellectual ability in the general population is important for reducing disability across the lifespan.
At a genetic level, twin and family studies have confirmed the heritability of intelligence, which typically account for about 50% of the variance. It is also well established that increasing numbers of genes becoming activated throughout the course of cognitive development, thereby amplifying the contribution of genetics over environment 2,3 . Results from genome-wide association studies (GWAS) have demonstrated that intelligence is highly polygenic with hundreds of genetic loci of small effect 4 , and that these loci cluster in genes involved in regulating brain-specific gene expression 1 . While difficult to detect such small effects, the utilization of polygenic scores (PGS) has made it possible to cumulatively predict genetic variance, as well as increase the power to detect gene-environment interactions associated with intelligence. In the largest study to date 4 , an intelligence-based PGS was found to explain up to 5.2% of the variability in intelligence.
The finding that genes involved in intelligence are implicated in neuronal pathways highlight the importance of brain morphological differences and function in the variability of intelligence. A positive association between intelligence and brain size has been robustly identified by multiple studies [5][6][7] , with both phenotypes sharing a common genetic origin 8,9 . In the study by Jansen 9 , 67 intelligence genes were also associated with brain structural differences, and a genetic correlation of 0.23 was found between both traits. Similarly, the association between genetic variants and intelligence is observed to be partly mediated by grey matter volume 10 . However, few studies have examined these effects in large samples, and the precise extent to which the phenotypic relationship between genetic variation and intelligence is mediated by brain morphological changes is still unknown.
Several environmental factors have been associated with intellectual ability. In particular, many studies have reported an association between intelligence and early life adversity (ELA); that requires significant adaptation due to, for example, abuse, neglect, witnessing domestic or other violence, and chronic poverty 11,12 . In both clinical and non-clinical samples, ELA has been found to be associated with variability in brain structure, including reductions in total grey matter volume 13,14 , and across limbic and prefrontal regions [15][16][17] . However, it should be noted that most studies have not examined the extent to which genetic factors interact with the effects of ELA on brain structure and intelligence. It is also not yet known whether and how the positive environmental effects of educational attainment, can impede or halt the impact of ELA on intelligence.
In this study we used structural equation modelling (SEM) to better understand the associations between genetic variation, ELA, education, total grey matter volume (GMV) and intelligence in a UK Biobank sample of 16,383 individuals. Genetic variation was indexed using a PGS of fluid intelligence, derived from a GWAS we carried out using a non-overlapping sample of 89,748 UK Biobank participants. We tested the hypothesis that (1) total GMV would mediate the association between the PGS and intelligence. We next tested (2) whether ELA would moderate the effects on (a) PGS on total GMV (first stage), (b) total GMV on intelligence (second stage) or (c) both (moderated mediation). We did so as previous work has demonstrated a negative association between ELA and changes in brain structure and intelligence 17 and a genetic moderation of the liability for a neural effect of ELA 18 . Finally, to test for a protective role of education, we examined (3) whether educational attainment would moderate the effects of ELA on intelligence via its effects on PGS and/or total GMV.

Participants
The UK Biobank is a large epidemiological cohort study that includes data on genetic, environmental, cognitive, and brain magnetic resonance imaging (MRI) measures. A total of 502,655 participants were recruited between 2006 and 2010 in the United Kingdom. The study was approved by the National Health Service (NHS) Research Ethics Service (reference 11/NW/0382) and our access to the data was granted by the UK Biobank Access Committee  Table 1).

Cognitive Measures
Cognitive tests were administered online on the same day as the MRI scan. The four cognitive tests used in the current study were: fluid intelligence, numerical reasoning, trail-making test part B and symbol-digit substitution. Using these four cognitive tests, we generated a single latent cognitive variable of intelligence in this sample. The fluid intelligence test involved a series of 13 items assessing verbal and arithmetical deduction (Cronbach α reliability= 0.62) 19 .
The symbol-digit test, which is similar in format to the symbol digit modalities test 20 , involved matching symbols to single-digit integers and was based on the number of correct symbol-digit matches made in 60 seconds. For the numerical memory test, participants were shown a twodigit number which they had to recall after a short pause. Numbers increased by one until the participant made an error or until they reached the maximum number of 12 digits. In the trailmaking test part B, participants were presented with the numbers 1-13 and the letters A-L arranged pseudo-randomly on the screen. They were instructed to alternate between touching the numbers in numeric order and letters in alphabetical order (i.e., 1-A-2-B-3-C). Full details on the content and administration of each test have been published elsewhere 21 .

Educational Attainment
As a measure of educational attainment, UK Biobank participants were asked which of the following qualifications applied to them (with the option of selecting more than one), 1) 'college or university degree; 2) A levels or AS levels or equivalent; 3) O levels or GCSE or equivalent; 4) CSEs or equivalent; 5) NVQ or HND or HNC or equivalent; 6) Other professional qualifications, for example, nursing, teaching/none of the above; 7) prefer not to answer'. Following the approach described by Rietveld et al., 22 we created a binary variable for education to index whether participants had obtained a college or university-level degree.

Early Life Adversity
Childhood adversity items were based on the short version of the Childhood Trauma Questionnaire Short Form (CTQ-SF) 23 . The CTQ-SF is a self-report questionnaire measuring physical abuse, emotional abuse, sexual abuse, physical neglect and emotional neglect. Items that were included in the analysis consisted of the following: "When I was growing up…" "…I felt loved" (Loved as a Child): "…people in my family hit me so hard that it left me with bruises or marks" (Physical Abuse); "…I felt that someone in my family hated me" (Hated as a Child); "…someone molested me (sexually)" (Sexual Abuse). We chose to include these items as they demonstrated a correlation > .3 with at least one other adversity item. All questions were rated on a Likert 5-point scale: never true, rarely true, sometimes true, often, very often true. Additionally, there was the option: "prefer not to answer", which was recoded as a missing value.

MRI Acquisition and Analysis
MRI data were collected in a single Siemens Skyra 3 T scanner with a standard 32-channel hear coil located at UK Biobank's recruitment centre. T1-weighted MPRAGE data was acquired in the sagittal plane using a three-dimensional magnetization-prepared rapid gradientecho sequence at a resolution of 1 x 1 x 1 mm, with a 208 x 256 x 256 field of view. Global and regional brain imaging-derived phenotypes (IDPs) were processed by the UK Biobank team and made available to approved researchers. Full details of the brain imaging protocols and quality control (QC) measures have been made available 24 . For our study, we used a global brain IDP of total GMV, which had been extracted using FMRIB's Automated Segmentation Tool (FAST) 25 . Scans of individuals with severe and visual normalization problems were excluded by the UK Biobank through manual inspection (as noted in Alfaro-Almagro et al., 24 ).

Genome-Wide Association Analysis of Intelligence
We performed a GWAS of intelligence using the fluid intelligence test from the UK Biobank (discovery sample n= 89,748). We chose to conduct our own GWAS primarily to avoid sample overlap in the PGS analysis. The fluid intelligence test rather than all four available tests was chosen to maximize the number of participants (for comparison, data on all four tests were available for only 36,383 participants).
Genotypic data was collected and imputed by the UK Biobank team. Full details of these along with imputation procedures are available in a publication 26 . In addition to the QC steps performed by the UK Biobank, SNPs with a minor allele frequency (MAF) <0.01, Hardy-Weinberg equilibrium (HWE) < 1×10-6 and missingness <0.02 were excluded. Further, multiallelic SNPs were removed as well as SNPs differing in allele frequencies across the genotyping arrays (UK BiLEVE, UK Biobank axiom arrays). Samples were removed based on the following: non-European ancestry, relatedness, discordant sex info, high heterozygosity/missingness, chromosomal aneuploidies or retracted consent. Following QC, a total of 7,829,832 variants and 89,748 individuals (41,895 males and 47,853 females) were included. We used a linear regression model in the discovery sample to test for genetic association with intelligence using PLINK 2 (https://www.cog-genomics.org/plink/2.0/). For the analysis, age, gender, genotyping array, UK Biobank assessment center, socioeconomic status (as assessed by the Townsend Deprivation Index) and the top 10 principal components of genetic population structure were entered in the linear regression.

Polygenic Score Calculation
We performed a PGS analysis based on our GWAS of fluid intelligence (IQ-PGS) using PRSice-2 27 . To avoid potential bias induced by sample overlap, we generated these scores in an independent sample of 16,383 UK Biobank participants (not included in our GWAS) for whom cognitive, genomic and adversity measures were available. SNPs in high linkage disequilibrium were clumped according to PRSice-2 guidelines. Following this a total of 604,290 variants were included for analysis. An effect-size weighted PGS was then computed for each individual based on a threshold of p <0.05.

Statistical Analysis
SEM calculations were performed in R (version 4.1.2) using the Lavaan package 28 . To improve convergence, before conducting the SEM analyses, the neural, genetic, and cognitive measures were standardized. Missing data were assumed to be missing at random and analyses were Based on previous studies linking brain size and intelligence to genetic variation, we first assessed the mediating role of total GMV on the association between IQ-PGS and intellectual ability. Intelligence was defined as a latent construct using the four cognitive tests (see Section 2.2). We next evaluated whether ELA would moderate (a) the association between IQ-PGS and intelligence, (b) the association between total GMV and intelligence and/or (c) the mediation effect. Finally, we considered whether education served as a potential mitigating factor of ELA through moderating its effect on intelligence via IQ-PGS and/or GMV volume. See figure 1 for a schematic overview of the models.
[Insert Figure 1] The moderating effect of education (college/university degree vs no college/university degree) was carried out using a multi-group SEM, which allowed for the estimation of measurement invariance (factor loadings and path coefficients) across the two groups. Given the sensitivity of the Δχ2 to sample sizes, differences to reject invariance across specifications were based on the following indices and values only: ΔCFI ≤ 0.01 and ΔRMSEA ≤ 0.015 for both factor loadings and intercepts and ΔSRMR ≤ 0.03 for factor loadings and ≤ .01 for intercepts 30,31 .
Throughout models tested, we corrected for age and total intracranial volume.

Is the relationship between the IQ-PGS and intelligence mediated by total GMV?
Testing our first hypothesis, which examined the mediating role of total GMV on the association between IQ-PGS and intelligence, we found that the data fitted the model well (CFI [Insert Figure 2]

Does ELA moderate the effect of IQ-PGS and/or total GMV on intelligence?
Next, we sought to determine whether ELA moderated the pathway between either IQ-PGS or GMV on intelligence or both. The results of this model fit the data well (CFI = 0.949; TLI = 0.919; RMSEA = 0.026, SRMR= 0.071), and explained 32% of the variance in intelligence.
As in the previous model, the direct effects of IQ-PGS and total GMV on intelligence were both significant ( Figure 3). In addition, a direct negative association was found between ELA and GMV (β = -0.042, p < 0.001) and between ELA and intelligence (β = -0.069, p < 0.001).
In this moderated mediation model in which ELA was included as a moderator, GMV significantly mediated the relationship between IQ-PGS and intelligence (indirect effect β = -0.012, p < 0.001). Specifically, lower GMV mediated the effects of IQ-PGS on intelligence, but only in individuals exposed to ELA.

Does educational attainment moderate the effects of ELA on intelligence, either directly or via its effects on IQ-PGS and/or total GMV?
Our final model examined whether education moderated the effects of adversity on intelligence through its association with IQ-PGS and/or total GMV. For this, we found that the model fit the data well (CFI= 0.944, TLI= 0.914, RMSEA= 0.026, SRMR= 0.071), and the variance explained in intelligence was comparable to those with (32%) and without (31%) a college/university degree. Before testing for group differences, we tested for measurement invariance between the educational groups by (1) constraining the loadings (metric invariance) and (2) the loadings and intercepts (scalar invariance) to equality (tested with ANOVA, Table   2). In all models, acceptable fit indices were observed and we found that metric and scalar invariance were held across groups. Thus, any observed differences in structural relations were not due to differences or errors in measurement and we proceeded to the multigroup analysis using our baseline model.
[Insert Table 2] As shown in Table 3, IQ-PGS, ELA and total GMV were directly associated with intelligence, while the relationship between IQ-PGS and GMV was nonsignificant in both groups. In those who had not been to college/university, the moderated mediation effect of ELA on total GMV, IQ-PGS and intelligence remained significant (indirect effect β = -0.017, p = 0.027).
Conversely, in those who had been to college/university, ELA was not a significant moderator of these effects (indirect effect β = -0.001, p = 0.947). Using a Wald Chi-Squared test to examine differences in these paths, a statistically significant difference between groups (x 2 (1)= 5.61, p= 0.017) was found, possibly indicating that education served as a protective factor on intelligence by moderating the effect of ELA on IQ-PGS and total GM volume.

Discussion
By leveraging the rich multimodal data of 16,384 UK Biobank individuals, our study sought to model the complex relationship between IQ-PGS, brain structure, ELA and educational attainment on intelligence. Consistent with previous studies, we found that IQ-PGS and total GMV were predictive of intelligence, and that total GMV mediated this genetic contribution to intelligence. However, this mediating association was only observed in a model that simultaneously considered environmental effects. Specifically, the mediating role of total GMV was only observed in those who had been exposed to ELA. We further found evidence of a protective effect of education, such that the moderating effects of ELA were not observed in participants who had a college degree or higher.

Brain morphology and the relationship between IQ-PGS and intelligence
Previous studies have found that structural brain metrics are positively associated with intelligence 32 , and that both share a common genetic basis 8,9,10 . In our study, we focused on total GMV given its consistent, albeit modest, association (~0.15-0.35) with intelligence, and because intellectual ability is likely to involve multiple brain areas rather than one specific region 6,1 . In line with previous studies, IQ-PGS and total GM volume were both predictive of intellectual ability. However, in this first model (which did not consider environmental effects), total GMV was not a significant mediator of the relationship between IQ-PGS and intelligence.
One interpretation of this result is that the genetic and neurobiological mechanisms underlying intelligence impact on areas of brain development that are not sensitively indexed by morphological changes, i.e., synaptic architecture and neuronal efficacy and integrity.
Alternatively, the role of brain structure in mediating the genetic determinants of intelligence may be more difficult to detect if moderated by environmental factors. Consistent with this view, we only found a mediating role of total GMV on IQ-PGS and intelligence when specific environmental factors such as ELA and education were included.

Environmental exposure as a moderator of the genetic and brain-related underpinnings of intelligence
Our findings are consistent with previous reports highlighting the importance of environmental variables in modifying biological processes that influence variability in intelligence in both clinical 14,33,34 and nonclinical samples 35,36 . The significant moderating effect of ELA on the association between IQ-PGS, GM volume and intelligence reported here adds weight to the hypothesis that exposure to adversity contributes to detrimental biological brain changes that persist into adulthood and with lasting consequences for intellectual development. Indeed, previous animal studies have reported a casual effect of ELA on intelligence both directly and via structural brain changes 37 .
Importantly, however, we also found that these moderating effects of ELA were in turn moderated by educational attainment. This suggests that not all individuals exposed to ELA will show reduced intellectual performance, and that educational attainment moderates the long-term impact of these exposures. A positive correlation between educational attainment and intelligence has been widely reported in the literature. For example, a meta-analysis by Ritchie and Tucker-Drob 38 found beneficial effects of education on cognitive abilities in the order of -~1 to 5 IQ points per each year of education; an association that persisted into adulthood. Similarly, the genetic overlap between measures of intelligence and educational attainment have also been widely observed 4,39 . Interestingly, a recent study by Elliott et al.,40 found that a PGS for years of education predicted intellectual ability which was partly mediated via total brain volume. To our knowledge, however, this is the first study to report that the positive effects of high IQ-PGS on total GMV and intelligence may be reduced under the effect of ELA, which itself is moderated by educational attainment.

Limitations and Strengths
The cross-sectional design of this study prevents us from drawing any firm conclusions about causality. Indeed, for the models tested here, one could equally speculate about whether individuals with lower intellectual ability are at greater risk for exposure to ELA, and it will be important for future studies to examine these associations longitudinally. Another potential limitation is that ELA was measured using retrospective self-reports. However, it is noteworthy that retrospective reports of ELA are shown to correlate moderately well with prospective measures and demonstrate comparable effects on negative life outcomes 41 . Future studies will benefit from the inclusion of a prospective assessment of ELA, as well as considering the developmental timing and frequency of trauma exposure, which may mediate the effects on intellectual functioning 42 . Further, while we controlled for the effects of age, the UK Biobank sample is restricted to middle and older age adults which potentially limits the generalizability of the results. Another potential limitation concerns the validity and reliability of the cognitive tests used here, which were brief and administered unsupervised. Notwithstanding, a recent study by Fawns-Ritchie and Deary 21 demonstrated that these cognitive tests correlated well with validated standard tests and had a moderate to high test-retest reliability. Finally, we did not examine other environmental risk factors (e.g., low birth weight and family income) nor did we include regional specific brain MRI volumes, which may also contribute to variation in intellectual ability. As such, future research should implement a fully data driven approach to include such factors.
Our investigation of the relationship between IQ-PGS, ELA, education and brain morphology on intelligence has important strengths, including the large sample size and the inclusion of imaging and genetic data. Further, the SEM approach taken to characterize these associations expands on prior research, which has mostly used simpler regression approaches to analyze the complex relationships between biological and environmental predictors of intelligence. The main advantages of SEM are that (1) it considers measurement error among variables, (2) it allows testing complex patterns of relationships and hypotheses simultaneously, (3) it evaluates constructs that cannot be directly measured and (4) it can test invariance of effects across different groups. Further, this multivariate technique allowed us to test both multiple direct and indirect effects within an overall model.

Conclusions
Our results provide evidence that the significance of total GMV as a mediator of the genetic underpinnings of intelligence is more apparent when the effects of ELA are considered.
Secondly, the impact of ELA on the genetic influence of total GMV can have long-lasting neurobiological effects on intellectual functioning. Finally, even when early environment is sub-optimal, longer time spent in education appears to offset at least some of the negative effects of ELA. If true, these associations highlight the potential of educational attainment to moderate other deleterious environmental effects. This inference is consistent with the evidence that higher-quality early education programmes (e.g., the Head Start Programme) can help buffer or protect individuals from the negative effects of ELA later in life 43 . Important goals for future work will be to examine the impact of such inventions longitudinally given that several other factors can modify the effects of exposure to ELA. Notwithstanding, our study provides important insights into the complex interplay between genetic, biological and environmental factors on intellectual ability.       Figure 1 Schematic overview of the SEM models tested. Model one (mediation model) examines the mediation effect of total GMV on the relationship between IQ-PGS and intelligence. Model 2 (moderated mediation) examines the moderating role of ELA on IQPGS, total GMV and intelligence (direct and indirect pathways). Model 3 (multi-group SEM model) examines these associations across different educational groups: individuals without (group one) and with a college/university degree or higher (group two).