SNPs were also filtered on the basis of linkage disequilibrium (LD). If two tag SNPs are in LD with a causal variant and in LD with each other, then both SNPs will show a relationship with the phenotype, despite the fact that the two SNPs are redundant. To avoid over-counting such redundant SNPs prior studies have imposed strict LD cutoffs, such that no two SNPs included in the set of prediction SNPs can have LD r2 > .05 (e.g., (Allen et al., 2010)). We chose to evaluate three different LD cutoffs: r2 = .05, .50, and 1.0 (i.e., no cutoff). The cutoff was imposed in the following way. First, all SNPs were regressed on the phenotype and their univariate weights and p-values recorded. SNPs were then sorted according to p-value. The most significant SNP was selected, and all SNPs with LD greater than the cutoff were culled from the list of SNPs. Then the remaining second-most significant SNP was considered and all SNPs in LD with it were removed. This process was completed until the least significant SNP was considered.
