The impact of educational attainment, intelligence and intellectual disability on schizophrenia: a Swedish population-based register and genetic study

The study addresses how cognitive traits relate to schizophrenia (SCZ) risk, etiology, and heterogeneity. SCZ is a severe, heritable psychiatric disorder with both common and rare genetic variants contributing to risk. Cognitive impairment is central to SCZ, associated with poorer outcomes, and lower premorbid cognitive ability and intellectual disability (ID) are known risk factors or comorbidities. Prior genomic studies show complex relationships: negative genetic correlation between intelligence and SCZ, near-zero genome-wide correlation with educational attainment (EDU) in recent GWAS despite epidemiological associations, and contributions from rare variants implicated in ID and neurodevelopmental disorders. The purpose is to integrate population-level epidemiology with comprehensive genetic data to clarify associations among EDU, premorbid cognitive ability, ID, and SCZ; quantify their genetic correlations; assess common and rare variant burdens in relation to cognition within SCZ; and empirically identify cognitive-related subgroups of SCZ with differential outcomes and genetic profiles. This work aims to elucidate SCZ heterogeneity and inform patient stratification.

Background work indicates: (1) SCZ has substantial heritability (twin/pedigree 0.60–0.80; SNP-heritability ~0.24) with risk influenced by both common and rare variants. (2) Intelligence and SCZ share loci with a negative SNP-based genetic correlation (rg ≈ −0.21), and bidirectional causal signals have been suggested by Mendelian randomization. (3) Earlier reports of positive genetic correlation between SCZ and EDU may reflect overlap with bipolar disorder, but the most recent EDU GWAS shows near-zero rg with SCZ. (4) Rare high-impact variants (e.g., SETD1A pLoF) associate with SCZ and developmental delay/ID; pLoF burden in constrained genes relates to SCZ and EDU, with enrichment in SCZ cases with comorbid ID and intermediate cognitive effects in unaffected CNV carriers compared to CNV carriers with SCZ. This literature motivates a joint epidemiological-genomic evaluation of cognitive traits and SCZ and exploration of SCZ subtypes indexed by cognition and ID.

Design: Two Swedish samples were analyzed. (1) A national register-based cohort (birth years 1958–1993) including SCZ cases (≥2 inpatient/outpatient diagnoses of SCZ or schizoaffective disorder; exclusions per predefined criteria) and all other Swedish residents as controls, with follow-up from 01-01-1973 to 12-31-2013. (2) A genotyped subsample (S3) drawn from the national cohort, comprising 4992 SCZ cases and 6009 controls with comprehensive genomic data and register linkages. Cognitive measures: Educational attainment (EDU) from national LISA database coded per ISCED and standardized as Z-score by birth year and sex; premorbid cognitive ability as IQ Z-scores from the Conscription Register (males, assessed at ages 18–19; individuals with SCZ at exam excluded); ID defined from National Patient Register diagnoses. Genetic data: Common variant polygenic risk scores (PRS) computed using external GWAS (excluding Swedish samples) for SCZ, bipolar disorder (BIP), intelligence (IQ), and EDU at p-value threshold ≤ 0.05; copy number variants (CNVs) defined as frequency < 0.01, size ≥ 100 kb, ≥15 probes; CNV burdens quantified by total deletion size (KB), number of CNVs, and number overlapping known pathogenic regions; rare exonic burden from whole-exome sequencing as counts of ultra-rare disruptive/damaging SNVs/indels absent from ExAC in constrained genes. In total, 4288 cases and 5305 controls had all genetic profiles. All burden metrics were standardized. Epidemiological analyses: Cox regression models estimated associations of EDU, premorbid IQ, and ID with incident SCZ. Models adjusted for sex, birth year (categorical), parental EDU (average if both available), maternal/paternal age at birth, and winter birth. Separate models evaluated each cognitive trait; joint models included (a) EDU and ID; and (b) EDU, premorbid IQ, and ID (males only). Entry at 01-01-1973; censoring at emigration, death, or 12-31-2013. Genetic epidemiology: Using the Multi-Generation Register, an extended twin-family SEM estimated pedigree heritabilities and bivariate genetic correlations (rg) between SCZ and each cognitive trait, adjusting for sex and birth year. SNP-heritabilities and SNP-rg were compared to literature estimates. Genetic burden–cognition analyses: In S3, linear regressions tested associations between genetic burdens (PRS, CNV deletion size, rare exonic burden) and EDU/premorbid IQ, adjusting for ancestry PCs and genotyping waves. Interaction by SCZ status was tested; significant burdens were entered into joint models stratified by case/control. Clustering: Among SCZ cases with complete data, unsupervised clustering was performed using cognition-related features: EDU, parental EDU, ID, age at first SCZ diagnosis, and number of BIP contacts. After regressing out birth year and sex from continuous covariates, Gower’s distance was computed, UMAP (n_neighbors=50, fixed seed) projected to 2D, and DBSCAN (eps=1, MinPts=50) identified clusters. Internal replication was assessed by random 1:1 split. Adverse outcomes compared across clusters via Cox models: suicidality, first hospitalization >200 days (median among top decile), and mortality; clozapine use summarized descriptively. The same clustering and between-cluster genetic burden comparisons (logistic regression) were performed in S3 (excluding CNV duplications due to null results with cognition). Multiple testing used Bonferroni within related test sets. Analyses used R (v4.0.3), OpenMx for SEM, and cluster/umap/fpc packages.

Samples: National cohort included 14,230 SCZ cases and 3,816,264 controls (lifetime prevalence 0.37%); S3 subsample had 4992 cases and 6009 controls. Profound case–control differences existed for EDU, premorbid IQ, and ID in both samples. Epidemiology (Cox HR per SD): Strong associations between poorer cognition and higher SCZ risk. Separate models: premorbid IQ HR=0.54 [0.52–0.55], EDU HR=0.43 [0.42–0.44], ID HR=13.81 [12.90–14.79] (all P<1×10^-300). Joint model 1 (EDU + ID): EDU HR=0.47 [0.46–0.48]; ID HR=7.54 [6.99–8.14] (both P<1×10^-300). Joint model 2 (males: premorbid IQ + EDU + ID): premorbid IQ HR=0.65 [0.63–0.67], EDU HR=0.65 [0.63–0.67], ID HR=12.56 [10.78–14.65] (all highly significant). Genetic epidemiology: Pedigree heritabilities: SCZ h2=0.70 [0.63–0.77]; premorbid IQ h2=0.65 [0.62–0.68]; EDU h2=0.37 [0.34–0.39]; ID h2=0.84 [0.77–0.91]. Pedigree genetic correlations with SCZ: premorbid IQ rg=−0.11 [−0.15, −0.07]; ID rg=0.50 [0.47, 0.52]; EDU rg=0.09 [−0.04, 0.22] (ns). These approximated literature SNP-rg: SCZ–IQ −0.21; SCZ–severe neurodevelopmental disorders 0.28; SCZ–EDU 0.02. Genetic burden and cognition (S3): Association with SCZ case status was negative for IQ-PRS (OR=0.88 [0.84–0.92], P=1.38×10^-5) and positive for EDU-PRS (OR=1.08 [1.04–1.13], P=1.12×10^-2). Within SCZ cases, EDU-PRS and IQ-PRS positively associated with EDU and premorbid IQ; the EDU-PRS effect on EDU was weaker in cases than controls (beta 0.13 vs 0.19; interaction P=3.89×10^-4). Rare exonic burden inversely associated with cognition in cases (EDU beta=−0.06, P=4.91×10^-6; premorbid IQ beta=−0.09, P=0.002) but not in controls. SCZ-PRS showed no association with cognition within cases. Clustering: Four reproducible SCZ clusters emerged: (1) medium EDU (56–57% of cases), (2) low EDU (≈25–29%), (3) high EDU with higher parental EDU and more BIP contacts (≈7–12%), and (4) ID (≈6%). High-EDU cluster had fewer adverse outcomes: in national training set, mortality 3.5% vs 8.3–11.4% in other clusters; long hospitalization (>200 days) 8.6% vs 16.3–28.0% (highly significant differences). In S3, clusters differed in genetic burdens: the high-EDU cluster showed higher EDU-PRS, no deficit in IQ-PRS, and no excess rare exonic or pathogenic CNV burden compared to controls, whereas the ID cluster showed greater CNV deletion size and pathogenic CNV counts and higher rare exonic burden.

The findings demonstrate a robust, multi-level relationship between cognitive traits and SCZ. Epidemiologically, lower premorbid cognitive ability, lower educational attainment, and the presence of ID are strong risk factors for SCZ. Pedigree-based genetic correlations between SCZ and cognition align with SNP-based correlations, indicating convergent genetic architecture: negative shared genetics with intelligence and positive with severe neurodevelopmental disorders, while the SCZ–EDU genetic correlation is near zero despite strong phenotypic association. Within SCZ, cognitive performance relates more to polygenic influences on cognition (IQ-PRS, EDU-PRS) than to SCZ-PRS, suggesting partially distinct mechanisms underpin cognition variation versus illness liability. Rare coding variant burden appears to depress cognition among cases, consistent with rare variant contributions to neurodevelopmental impairment. Data-driven clustering using cognition-related variables identifies clinically meaningful SCZ subgroups with distinct outcomes and genetic profiles; notably, a high-EDU subgroup shows better prognosis and cognitive-favoring polygenic profiles without excess rare variant burden, supporting the utility of cognitive indices for stratifying heterogeneity and potentially informing tailored interventions.

By integrating national registry epidemiology with genetic epidemiology and burden analyses, the study clarifies relationships between SCZ and cognitive traits. It confirms strong negative associations of premorbid cognitive ability and positive associations of ID with SCZ risk, and reveals complex, largely null genetic correlation between EDU and SCZ despite phenotypic links. Within SCZ, cognition tracks with polygenic scores for cognitive traits and inversely with rare exonic burden. Unsupervised clustering yields reproducible cognitive-indexed subgroups with differential outcomes and genetic architectures; a high-EDU subgroup has fewer adverse outcomes and cognition-favoring polygenic profiles. These results underscore the value of cognitive variables in dissecting SCZ heterogeneity and motivate future research into subtype-specific mechanisms and treatments. Future directions include replication in diverse settings, expansion of genotyped cohorts, incorporation of longitudinal cognitive measures and broader cognitive batteries, and testing whether cognitive-indexed subtypes predict treatment response.

Key limitations include: (1) Limited power for rare CNV analyses; rare events explain small variance and require very large samples. (2) National Patient Register likely captures a subset with more severe ID, limiting generalizability. (3) Potential diagnostic surveillance bias inflating associations between ID and SCZ due to increased clinical contact. (4) Selection bias in the genotyped S3 subsample (survivorship and consent capacity). (5) Clustering relies on complex methods and proxy variables; risk of overfitting cannot be excluded. (6) Generalizability may be limited outside Sweden’s universal healthcare system. (7) Time-varying confounders (e.g., socioeconomic status, comorbidities) affecting EDU were not modeled and may bias associations. (8) EDU is an imprecise proxy for cognitive ability; broader, longitudinal cognitive assessments would improve inference.