[8]β=1−(c+1/N)var(Ajk) where c is constant for a certain MAF threshold θ —for example c = 6.2 × 10−6 when θ = 0.1 and c = 0 when θ = 0.5 (Fig. 1). The regression coefficient β is less than 1.0 because of two effects. First, the term in 1/N is due to the sampling error in estimating A from only N SNPs. This corresponds to the sampling error for Aijk at a single SNP calculated above as 1. If c = 0 and N = infinite, β = 1. In this case Ajk is the genomic relationship averaged over all positions in the genome. As the causal variants are a sample of such positions, Ajk is an unbiased estimated of Gjk. Second, the term in c occurs because the causal variants are not a random sample of all SNPs but a sample with low MAF. This causes the causal variants to have lower LD with the SNPs than random SNPs do with one another. Thus, even if Ajk was calculated from an infinite number of SNPs, it would still tend
