Fig. 3.

Boxplots of point estimates of causal effects by various methods in simulations. The true causal effect is set at 1.00 (dotted horizontal line) for all simulations. A total of 800 genes that pass the initial PTWAS scan (with p value cutoff 0.05) are examined by each method. SMR utilizes the standard two-stage least squares algorithm based on the top eQTL SNP for each gene. For Fusion and PrediXcan, the point estimates are obtained by regressing the phenotype data on the corresponding predicted gene expression levels. Note that effect size estimation is not the designed usage of Fusion and PrediXcan. Our intention is to illustrate that methods designed for testing may not be suitable for estimation. Among all methods compared, PTWAS and SMR yield seemingly unbiased estimates. The results by PTWAS are, overall, more accurate