The design and methods of the European Prospective Investigation into Cancer and Nutrition (EPIC)-InterAct case-cohort study have previously been described . A total of 340,234 EPIC participants, who were over the age of 40 and free of known diabetes at baseline, in eight of the ten EPIC study countries (26 centres) were followed up for 3.99 million person-years (median follow-up 10.9 years), during which 12,403 incident cases of diabetes were ascertained and verified. The mean (SD) age at diagnosis was 62.3 (7.8) years in men and 62.6 (8.2) years in women . Ascertainment of incident diabetes involved multiple sources of evidence including self-report, linkage to primary care registers, secondary care registers, medication use (drug registers), hospital admissions and mortality data, with a minimum of two data sources being required to confirm the diagnosis. Cases in Denmark and Sweden were not ascertained by self-report but identified via local and national diabetes and pharmaceutical registers, and hence all ascertained cases were considered to be verified. Information from any follow-up visit or external evidence with a date later than the baseline visit was used. Follow-up was censored at the date of diagnosis, 31 December 2007 or the date of death, whichever occurred first. A centre-stratified subcohort of 16,835 (4.9% of the entire EPIC cohort) individuals was selected at random . We excluded 548 individuals with known prevalent diabetes and 133 with unknown diabetes status at baseline. We also excluded 422 cases and 352 subcohort participants (of whom 34 were incident cases) with insufficient sample volume for GAD65 antibody measurement, resulting in 15,802 subcohort participants and 11,981 incident cases being included in the analysis.
All study participants gave informed consent, and the investigation has been carried out in accordance with the Declaration of Helsinki as revised in 2008.
GAD65 antibody measurement
Blood samples were drawn at the time of participation in EPIC, at which time all participants were free of known diabetes. Blood plasma was prepared and stored at −196°C in liquid nitrogen at the coordinating centre at the International Agency for Research into Cancer (IARC) in Lyon, France, or in liquid nitrogen in local biorepositories except for Umeå, where −80°C freezers were used .
The samples had been subject to at least two freeze-thaw cycles before being analysed for GAD65 antibody. Recombinant S-GAD65 was produced in an in vitro coupled transcription and translation system with SP6 RNA polymerase and nuclease-treated rabbit reticulocyte lysate (Promega, Madison, WI, USA) as previously described . The WHO standard  was included and used to express immunoglobulin binding levels in relative units.
To determine the cut-off for GAD65 antibody positivity, we used a competition assay employing recombinant human GAD65 (rhGAD65) (Diamyd Medical, Stockholm, Sweden) as previously described . A total of 900 serum samples were randomly selected across countries in the EPIC-InterAct study population. The samples were incubated with radiolabelled GAD65 in the absence or presence of rhGAD65 (200 ng/ml) or BSA (200 ng/ml). Samples in which binding to radiolabelled GAD65 was reduced by 50% in the presence of rhGAD65, but not BSA, were considered to be positive for GAD65-specific antibodies. We used 50% as a cut-off for successful competition as an approximation of the IC50. Given the low sample volume, we chose to use the competitor at the optimal concentration (200 ng/ml) found in previous experiments, in which we titrated the amount of rhGAD65 necessary to give maximal competition. In a variation of the traditional receiver operating characteristic (ROC) analysis, we plotted GAD65 antibody levels of samples that were competed by ≥50% against GAD65 antibody levels of samples that were competed by <50%. The area under the ROC curve was 0.97, indicating excellent predictive ability of the GAD65 antibody measurement. At a cut-off level of ≥65 U/ml, the measurement had 99% specificity and 85% sensitivity (electronic supplementary material [ESM] Fig. 1). Thereafter, all samples in the subcohort (n = 15,802) and incident cases (n = 11,981) were analysed for GAD65 antibodies in a radiobinding assay (RBA) as previously described .
Measurement of covariates
Weight, height, and waist and hip circumferences were measured with participants not wearing shoes and in light clothing or underwear, as described previously . BMI was calculated as weight/height squared (kg/m2). Waist circumference was measured either at the narrowest circumference of the torso or at the midpoint between the lower ribs and the iliac crest. Hip circumference was measured horizontally at the level of the largest lateral extension of the hips or over the buttocks. Anthropometric data were mostly self-reported in the Oxford centre, and waist and hip circumferences were not measured in the Umeå centre (n = 1845).
Standardised information on highest educational level (none, primary, technical, secondary or further education) and smoking status (current smoker, never a smoker or former smoker) was collected by questionnaire at baseline . Physical activity was assessed using a brief questionnaire covering occupation and recreational activity, from which a validated physical activity index (inactive, moderately inactive, moderately active or active) was derived .
Genetic analysis and GRS
Samples were processed for array-based genotyping if they had sufficient DNA that could be successfully genotyped on TaqMan (Thermo Fisher Scientific, Waltham, MA, USA) or Sequenom (San Diego, CA, USA) platforms and had sex chromosome genotypes concordant with self-reported sex. Samples that failed one genotyping round for reasons that did not relate to sample quality (e.g. signal intensity outliers or plates/arrays with an unusually high failure rate) were repeated. Samples were genotyped on the Illumina 660 W-Quad BeadChip, the Illumina HumanCoreExome-12 or the Illumina HumanCoreExome-24 (San Diego, CA, USA). Samples genotyped on the Illumina 660 W-Quad BeadChip were randomly selected from the available samples with the number of individuals selected per centre being proportional to the percentage of total cases in that centre. The Danish samples were not available for genotyping at this stage. Genotyping was carried out at the Wellcome Trust Sanger Institute. Most of the remaining non-Danish samples were genotyped on the Illumina HumanCoreExome-12 at Cambridge Genomic Services in the University of Cambridge Department of Pathology. Finally, the Danish samples and repeat samples due to poor genotyping were genotyped on the Illumina HumanCoreExome-24 also at Cambridge Genomic Services. Sample quality control criteria varied slightly by array but included call rate (<95.4% in Illumina 660, <98% in core exome arrays), X chromosome heterozygosity concordance with self-reported sex, outliers for heterozygosity and concordance with previous genotyping results.
From the genome-wide array data, we calculated a type 1 diabetes GRS as a weighted average of 33 SNPs, including five HLA variants. The relevant SNPs and their individual associations with GAD65 antibody positivity are described in ESM Table 1 . We also calculated a type 2 diabetes GRS as a weighted average of 68 SNPs from a DIAbetes Genetics Replication And Meta-analysis (DIAGRAM) consortium publication .
Baseline characteristics of the analysis sample were summarised by GAD65 antibody status (negative/positive), separately within the subcohort and incident diabetes cases, using means and standard deviations for continuous variables (except for GAD65 antibody level, which had a skewed distribution so the median and interquartile range were used) and percentages for categorical variables.
The association between GAD65 antibody status (positive/negative) and incident diabetes was estimated using Prentice-weighted Cox regression, which is appropriate for estimating association in a case-cohort study. We fitted models within each country and the estimated HRs were combined across countries using random effects meta-analysis. We fitted three models including the following covariates: Model 1—age (as underlying time scale), sex and centre; Model 2—also including physical activity, smoking status and education; Model 3—also including family history of diabetes. In order to study the effects of high vs low GAD65 antibody levels, GAD65 antibody-positive individuals were further subdivided into those with GAD65 antibody equal to or above, and those with GAD65 antibody below, 167.5 U/ml, which is the median antibody level in the GAD65 antibody-positive group in this study. Prentice-weighted Cox regression was also used to test possible multiplicative interactions of GAD65 antibody status with: (1) sex; (2) BMI category; (3) waist/hip ratio (WHR) category (sex-specific tertiles); and (4) type 1 diabetes GRS tertile. The interactions with anthropometry measures were tested because of prior studies suggesting that adiposity could moderate the association of autoimmunity with diabetes [9, 10]. In this instance we fitted models to the overall dataset with adjustment for country, due to insufficient data within each country to obtain country-specific estimates. HRs and 95% CIs within each subgroup were calculated.
The associations of the type 1 diabetes GRS and the type 2 diabetes GRS with GAD65 antibody status (positive/negative) were estimated separately in the subcohort and the incident diabetes cases, using logistic regression adjusted for age and sex, since the prevalence of autoimmunity is associated with age and sex. Models were fitted within each country and estimated odds ratios combined across countries using random effects meta-analysis. ORs (and 95% CIs) per risk allele of each of the individual SNPs contributing to the type 1 diabetes GRS were also calculated in the subcohort using the same method.
The association between the type 1 diabetes GRS and incident diabetes, by GAD65 antibody status, was estimated using Prentice-weighted Cox regression as described above.