Each GRS was standardised (zero mean, unit standard deviation) over the entire dataset. Next, we employed elastic-net logistic regression50 using the R package ‘glmnet’51 to model the associations between the 19 GRSs and stroke, adjusting for sex, genotyping chip (UKB vs BiLEVE), and 10 genetic PCs. A range of models with different penalties was evaluated using 10-fold cross-validation. The best model, in terms of highest cross-validated AUC (area under receiving-operating characteristic curve), was selected as the final model and held fixed for validation in the rest of the UKB data. The final adjusted coefficients for each GRS (odds ratios) in the penalised logistic regression are shown in Supplementary Fig. 1, in comparison with the univariate estimates (based on one GRS at a time).
