Distributions of t-statistics across various shared eQTL proportions for all nine train-test scenarios with 1000 Genomes populations for a fixed TWAS effect size and fixed number of causal eQTLs.

The labels are ordered from left to right from least admixture proportion (CEU to YRI, 0% admixture proportion) to highest admixture proportion (YRI to AA, 80% admixture proportion), with train-test scenarios from a population into itself on the right of each panel. Increasing proportions of shared eQTLs yield stronger association statistics from cross-population predictions. Fully shared eQTL architectures yield consistently high power across populations. Median t-statistics increase as populations share more haplotypes, while association tests with gene expression predicted in the same population show consistently high power.

(TIFF)