Predictive performance of elastic net was similar whether all SNPs from the 1000 Genomes imputation or the HapMap Phase II subset were included in the model building (Supplemental Figure 2). Models based on imputed data (both the 1000 Genomes and the HapMap subset) substantially outperformed models based on genotyped SNPs in WTCCC (Supplemental Figure 2). Thus, we chose the elastic net models built in the smaller HapMap SNP subset, relative to 1000 Genomes, in our applications of PrediXcan to reduce computation time without sacrificing performance. As reference transcriptome studies increase in sample size, we may need to switch to a more dense imputation to take advantage of increased prediction performance from rare variants.
