EWAS of fasting glucose and insulin

All statistical analyses were performed using R statistical software and the two-tailed test was considered. Insulin (FI) was natural log transformed. In the analysis (n = 4,808), we first performed EWAS of fasting glucose (FG) and insulin in each cohort separately. Linear regression analysis was used to test the association between glucose and insulin with each CpG site in the Rotterdam Study samples. Linear mixed models were used in NTR and TwinsUK accounting for the family structure. We fitted two models for each cohort: 1) the baseline model adjusting for age, sex, technical covariates (e.g. batch, chip array number and position on the array), white blood cell counts (lymphocytes, monocytes, and granulocytes) and smoking status, and 2) a second model additionally adjusting for BMI. We removed probes that have evidence of multiple mapping or contain a genetic variant in the CpG site. All cohort-specific EWAS results for each model were then meta-analysed using inverse variance-weighted fixed effects meta-analysis as implemented in the metafor R package. We produced quantile-quantile (QQ) plots of the -log₁₀ (P) to evaluate inflation in the test statistic.