Data source and data linkage

The SAIL databank (http://www.saildatabank.com) is a national data repository containing anonymised, person-based, linkable data in Wales for over 3 million people. The procedure for linking research study data to SAIL has been described elsewhere.^{Reference Lyons, Jones, John, Brooks, Verplancke and Ford9,Reference Rapsomaniki, Timmis, George, Pujades-Rodriguez, Shah and Denaxas17,Reference Lloyd, McGregor, John, Craddock, Walters and Linden18} In brief, data from CardiffCOGS study individuals who had consented to linkage were imported into SAIL in line with permissions already granted to SAIL relating to good practice in research governance and privacy protection. We adopted a split file approach to separate individual identifiers from the interview data. Identity matching and creation of pseudonymised linkage keys (anonymised linking fields) were performed by a trusted third party prior to linkage and further encryption of data sets using deterministic matching based on NHS number or probabilistic matching using available demographics based on the Welsh Demographic Service data-set (all individuals registered with a general practice (GP) surgery). We included participants whose data were probabilistically linked with an adequate level of matching accuracy (matching score ≥0.9).^{Reference Lloyd, McGregor, John, Craddock, Walters and Linden18}

We used the General Practice Database (GPD), containing diagnoses, symptoms, investigations, prescribed medication, referrals, coded hospital contacts and test results. At time of analysis, 77% (333/432) of GP surgeries in Wales supplied their data to SAIL. We also extracted data from the Patient Episode Database for Wales (PEDW), an NHS Wales hospital admissions data-set consisting of clinical information from all hospital admissions (in-patient and day cases) covering the entire population of Wales.

Measures for health outcomes

We used ICD-10 and Read codes for GDP and PEDW data-sets, respectively, to ascertain health outcomes.¹⁹ For schizophrenia, we adopted the codes that were validated and used in previous studies.^{Reference Lloyd, McGregor, John, Craddock, Walters and Linden18,Reference John, McGregor, Jones, Lee, Walters and Owen20,Reference Economou, Grey, McGregor, Craddock, Lyons and Owen21} We selected smoking, type 2 diabetes mellitus, ischaemic heart disease and body mass index (BMI) because of their high frequency in clinical samples of individuals with schizophrenia and their potential to contribute to increased mortality either directly, or via phenomena such as metabolic syndrome. We selected congenital disorders, intellectual disability and epilepsy as they are neurodevelopmental phenotypes with direct relevance to CNV carrier status. All of these variables also had either established SAIL algorithms or were considered to have high-quality data available in SAIL. We identified individuals with intellectual disability,^{Reference Rahman, Todd, John, Tan, Kerr and Potter22,Reference Brophy, Kennedy, Fernandez-Gutierrez, John, Potter and Linehan23} ischaemic heart disease,^{Reference Rapsomaniki, Timmis, George, Pujades-Rodriguez, Shah and Denaxas17} epilepsy,^{Reference Pickrell, Lacey, Bodger, Demmler, Thomas and Lyons24} diabetes mellitus^{Reference Eastwood, Mathur, Atkinson, Brophy, Sudlow and Flaig25} and determined smoking status^{Reference Atkinson, Kennedy, John, Lewis, Lyons and Brophy26} based on previously published works.

The list of codes used for extracting height, BMI and for identifying individuals with congenital disorders are given in supplementary Table 1 available at https://doi.org/10.1192/bjo.2020.42. For identifying schizophrenia, intellectual disability, ischaemic heart disease, epilepsy, diabetes mellitus and congenital disorders, we combined both the GPD and PEDW data-sets. We additionally extracted individuals’ height, BMI and smoking status from the GPD where available.

For comparison, we extracted lifetime diagnoses and estimated crude rates, as well as age- and gender-standardised rates, of health outcomes for the whole population and those diagnosed with schizophrenia between the ages of 17 and 84 years (as at 30 June 2016). For estimating standardised rates from SAIL, crude rates were first computed for five age groups (17–34, 35–44, 45–54, 55–64 and 65–84 years) for both genders. Standardised rates were then estimated based on the age and gender distribution in the clinical cohort. Linked data in SAIL were interrogated using structured query language (SQL DB2).

Genotyping and CNV calling

Samples were genotyped on OmniCombo or OmniExpress arrays (Illumina).^{Reference Rees, Walters, Georgieva, Isles, Chambert and Richards16} After standard quality control, imputation was performed using IMPUTE2^{Reference Howie, Donnelly and Marchini27} and the 1000 genomes (phase 3) and UK10 K reference panels.^{Reference Huang, Howie, McCarthy, Memari, Walter and Min28} Best-guess genotypes were generated with the following thresholds: minimal genotypic confidence >90%, INFO-score >0.9, minor allele frequency (MAF) >1%, and Hardy-Weinberg equilibrium P-value <1 × 10⁻¹⁰.

Full details of the CNV calling methods and quality controls measures used have been published elsewhere.^{Reference Rees, Kendall, Pardiñas, Legge, Pocklington and Escott-Price29} Illumina Genome Studio (version 2011.1) was used to process raw intensity data into log R ratios (LRR) and B allele frequencies. PennCNV (version 1.0.3.18) was then used to call CNVs based on 666 868 probes common to all Illumina arrays used.^{Reference Wang, Li, Hadley, Liu, Glessner and Grant30} CNVs were joined if separated by less than 50% of their combined length. CNVs were excluded if they were (a) called with fewer than ten probes, (b) overlapped low copy repeats by more than 50% of their length, (c) had a probe density of less than 1 probe/20 kb, or (d) had a frequency of >1%.^{Reference Purcell, Neale, Todd-Brown, Thomas, Ferreira and Bender31}

CNVs

To examine for enrichment of rare, pathogenic CNVs in physical health comorbidity, we analysed the presence of 12 CNVs robustly associated with the risk of schizophrenia,^{Reference Rees, Walters, Georgieva, Isles, Chambert and Richards16,Reference Rees, Kendall, Pardiñas, Legge, Pocklington and Escott-Price29} and 54 CNVs nominally associated (P < 0.05) with intellectual disability, autism spectrum disorder or schizophrenia (‘neurodevelopmental CNVs’).^{Reference Coe, Witherspoon, Rosenfeld, van Bon, Vulto-van Silfhout and Bosco32} Following the approach taken in our previous work, 15q11.2 duplications were excluded because of their high frequency.^{Reference Kendall, Rees, Escott-Price, Einon, Thomas and Hewitt33} CNV burden analyses were carried out using PLINK on regions of variable copy number at two size thresholds (a) ≥500 KB and (b) ≥1 MB and converted into carrier status for the purpose of regression analyses.

Polygenic risk scores

Polygenic risk scores were calculated using the largest published schizophrenia genome-wide association study meta-analysis (39 915 individuals with schizophrenia, 64 639 controls) as a training set and using the established method described by Wray et al.^{Reference Pardiñas, Holmans, Pocklington, Escott-Price, Ripke and Carrera34,Reference Wray, Lee, Mehta, Vinkhuyzen, Dudbridge and Middeldorp35} All study individuals were excluded from the training set. scores were generated using the –score function in PLINK^{Reference Purcell, Neale, Todd-Brown, Thomas, Ferreira and Bender31} for Single nucleotide polymorphisms with MAF >10%, INFO score >0.9, a low linkage disequilibrium to each other and excluding indels and the extended major histocompatibility complex region. Polygenic risk scores were calculated at nine P-value thresholds; 1 × 10⁻⁸, 1 × 10⁻⁶, 1 × 10⁻⁴, 1 × 10⁻³, 0.01, 0.05, 0.1, 0.2 and 0.5.

Rates of physical illness

Linked data in SAIL were interrogated using structured query language (SQL DB2). All crude rates and standardised rates of health outcomes were expressed as a percentage of population affected (lifetime prevalence). All standardised rate ratios (SRRs) and their 95% confidence intervals were calculated as previously described.^{Reference Armitage, Berry and Matthews36–Reference Newman38}

Ascertainment rates of behaviours and diagnoses

We evaluated the agreement on diagnoses between the clinical and electronic cohorts for each health outcome by constructing two × two contingency tables based on the paired responses from the interview and SAIL. Level of agreement was then assessed by unweighted Cohen's kappa coefficient^{Reference Cohen39} and Gwet's AC1.^{Reference Gwet40} The 95% confidence intervals of Cohen's kappa and Gwet's AC1 were estimated as described in Fleiss, Cohen & Everitt^{Reference Fleiss, Cohen and Everitt41} and Gwet,^{Reference Gwet40} respectively. Strength of agreement metrics was categorised according to previously described criteria.^{Reference Landis and Koch42}

CNV analyses

Association analyses were carried out using linear regression for average BMI (normalised using log₁₀ transformation) and average height, and Firth logistic regression for all binary traits (type 2 diabetes mellitus, smoking, ischaemic heart disease, congenital disorders, epilepsy, intellectual disability). Firth logistic regression allows for the analysis of low numbers in addition to the inclusion of covariates. All analyses included age and gender as covariates.^{Reference Wang43}

Polygenic risk analyses

We regressed a model for each polygenic risk score created from various training P-value thresholds against a base model including age, gender, the first five principal components and any additional principal components from the first 20 that were nominally associated with the phenotype of interest. We repeated these analyses with the addition of covariates reflecting symptom severity, non-response to antipsychotics, antipsychotic exposure, smoking status and genotyping platform (defined in supplementary Table 2). All statistical analyses were carried out in R and results were subject to Bonferroni correction for the eight phenotypes examined (P-value threshold 0.0063).