We compared MTAG to three multi-trait methods that can be applied to an arbitrary number of GWAS summary with unknown overlap12,13 (Supplementary Note). Unlike MTAG, these methods do not provide trait-specific SNP effect estimates but instead test whether the SNP is associated with none of the traits. We generate a (conservative) MTAG-based test of the same null hypothesis by using the minimum of the trait-specific MTAG P values, Bonferroni-adjusted for the number of traits. In two-trait simulations, we find that MTAG has greater power when the correlation in true effect sizes or GWAS estimation error is non-zero, especially when the traits’ GWASs are higher powered. In real-data applications to (i) DEP, NEUR, and SWB, and (ii) six anthropometric traits, MTAG identifies more loci. We test the anthropometric loci in GIANT consortium results and find that the loci identified by MTAG and missed by the other methods replicated at a higher rate than the loci identified by one of the other methods and missed by MTAG.
