with previous publications, PGS were generated using the basic p-value clumping and thresholding (P+CT) method (LD clumping r2 threshold of 0.1, clump window of 500kb, 10 p-value thresholds). We also generated PGS using SBayesR,20 which is one of several methods that has been found to improve accuracy of PGS compared to C+T by better choice of the SNPs and their weights (derived from the GWAS effect sizes) through modelling of the genetic architecture. Of these methods we chose SBayesR because it ranked high in a study comparing methods,21 requires no tuning sample to estimate hyper-parameters and is computationally less demanding. We used the software recommended LD reference sample (sbr_ldmatrix.band.mldm) to infer the expected correlation structure between SNP association statistics. We also used the SBayesRC, which is an extension of SBayesR which uses functional information in the SNP weighting algorithm.71 For comparisons with the C+T PGS using genome-wide significant SNPs (p < 5×10−8) we also constructed a PGS based on the genome-wide significant SNPs and their weights (bJ) estimated from a conditional/joint COJO analysis (that allows multiple SNPs within an LD block to be selected if they show association additional to the lead SNP).12 For benchmarking comparisons, we also calculated PGS
