In each scenario, the 20 simulated phenotypes were either a) summed and the sum score was regressed on the GV, b) subjected to a 1-factor model to calculate Thompson's factor scores [23], which were regressed on the GV, c) subjected to a MANOVA with the GV as covariate (canonical correlation analysis), d) subjected to MultiPhen (regressing the GV on all 20 phenotypes in a multivariate ordinal regression model), or e) individually regressed on the GV (using logistic or ordinal regression where appropriate). The last procedure yielded 20 p-values per simulated GV, which were then combined into 1 overall trait-based p-value PT using TATES. In addition, we compared the performance of TATES to that of various other published procedures for combining p-values, limiting our comparison to procedures that, like TATES, do not require permutation, i.e., Fisher's combination test, Lancaster's weighted Fisher test, the Z-transform test, and the original Simes procedure (see Text S1). All data simulations and subsequent analyses were repeated 2000 times. We counted the number of times that the GV effect was detected given α = .05.
