In the “shared, unstructured” scenario, mash and mash-bmalite are the best methods, finding almost exactly the same signals (Supplementary Table 2), with metasoft-RE2 close behind as it was applied to the task it was designed for (testing unit-level significance). The strong similarity in performance between mash and mash-bmalite is explained by mash putting negligible weight on the data-driven covariance matrices, which are unnecessary in this scenario. This illustrates mash’s ability to adapt to the data, increasing power in complex scenarios without being unduly penalized in simpler scenarios.
