fed into LDpred for LD estimation, there were 1,214,408 SNPs and 14,028 individuals. At the coordination step of LDpred, we used the option “--max-freq-discrep” in order to exclude markers that have a frequency discrepancy greater than 0.1 between the summary statistics and genotype data. We also used the “--z-from-se” option so that Z statistics were obtained from the GWAS coefficient estimates and their standard errors rather than from P values (the default) because the latter led to issues in LDpred for markers with extremely small P values. For each PGI, we used the LD window recommended by Vilhjalmsson et al.30, i.e., the number of markers common between the LD reference data, cohort genotype data and summary statistics left after the remaining LDpred quality control filters (MAF > 0.01, no allele mismatch, no ambiguous alleles), divided by 3,000. The fraction of causal markers was set to 1 for each phenotype to ensure consistency across phenotypes.
