indicate that a reference sample with a size of at least 2,000 is required and that little additional accuracy is gained beyond a sample size of 5,000 (Supplementary Fig. 4). The reference sample needs to be checked for cryptic relatedness and population stratification, which could cause correlations between SNPs that do not exist in the discovery set. In the present study, we included only the individuals of European descent in the ARIC cohort18 and of British Isles descent in the QIMR cohort14 and removed one of each pair of individuals with a SNP-derived relatedness estimate of >0.025 in both cohorts (Online Methods). If the expected value of the LD correlation between two SNPs is zero in the general population, the sampling variance of an observed LD correlation in a sample is proportional to the sample size (m), with var[r | E(r) = 0] = 1 / m. Thus, given a random sample of 6,654 unrelated individuals from the population, the probability of observing a LD correlation greater than 0.1 or smaller than –0.1 (r2 > 0.01) is 3.4 × 10–16. In order to investigate possible false positives resulting from errors in LD estimation, we first performed the analysis using the
