We first identify rows j of matrix B^ that correspond to the “strongest” effects. For example, in the GTEx data we chose rows corresponding to the “top” SNP for each gene, which we defined to be the SNP with the highest value of Zjmax:= maxrb^jr/s^jr. (We used the maximum rather than the sum because we wanted to include effects that were strong in a single condition, not just effects that were shared among conditions.) For the simulated data, we ran ash separately for each condition r, computed lfsrjr for each (j,r), then chose rows j for which minr lfsrjr < 0.05.
