After QC filtering the ILMN100K SNPs and removing SNPs used for imputation, we find 58,819 autosomal SNPs from the ILMN100K panel overlap with the HapMap, and 69,555 SNPs overlap with the 1000G panel. In the background of Figure 3A, we plot the correlation (R2) between ILMN100K allele counts and 1000G imputed allele dosage by minor allele frequency (MAF; estimated from ILMN100K genotypes). Most of the SNPs have high imputation accuracy, including rare SNPs. We then calculated the LOESS (locally weighted polynomial regression) smooth curve of R2 by MAF for three settings: (1) SNPs present on HapMap and imputed using the HapMap reference panel (red line, R2 calculated using HapMap imputation), (2) SNPs present on HapMap but imputed using the 1000G reference panel (blue line), and (3) SNPs not present on HapMap and imputed using the 1000G reference panel (green line). At the HapMap SNPs (settings 1 and 2), imputation using the 1000G reference achieved slightly lower quality compared with using the HapMap reference. This is probably because microarray-based genotyping is still more reliable than the low-coverage, next-generation sequencing used in
