Our approach can be conceptually viewed as a test for the correlation between the genetic component of expression and the genetic component of trait (Methods). Since several recent methods have been proposed that measure genetic correlation between summary statistics37, we sought to evaluate this relationship empirically. We compared TWAS to the recently proposed cross-trait LD-score regression (LDSC) that estimates genome-wide genetic correlation between traits37. Although LDSC is not intended for local analyses due to model assumptions on polygenicity and use of block-jackknife across loci for estimating standard errors, we performed the evaluation using expression and phenotype (height) from the YFS cohort, using the results over individual data as the “gold standard” (Methods). We find that LDSC estimate of genetic correlation between height and expression from summary data is highly correlated with the gold standard (correlation=0.7, Supplementary Figure 13), but the relationship is much noisier than that of TWAS (correlation=0.99, Supplementary Figure 5, 13). This suggests that TWAS attains more power in relating expression to complex traits.
