Finally, we investigated the effect of the expression reference panel size on performance of TWAS (Supplementary Figure 9). In general, TWAS always outperforms eGWAS when multiple variants are causal. Interestingly, power for either approaches does not increase substantially beyond 1,000 expression samples, suggesting that the expression panels analyzed in this manuscript nearly saturate the available imputation accuracy. This was further reflected in an analysis of real data where merging cohorts did not substantially change the distribution of TWAS statistics for the same gene set (Supplementary Figure 14). Although these results come with caveats (e.g. standard assumptions of additive effects and normal residuals), they suggest that the main benefit of larger expression reference panels is in increasing the total number of significant cis-heritable genes available for imputation (Figure 3).
