Our simulations are not exhaustive. Data were simulated for three traits according to an additive model, and analyzed accordingly. We did not simulate and analyze other, non-additive genetic models and/or (higher-order) interactions, nor did we study scenarios with more than three traits. In addition, priors for MV-SNPTEST and MV-BIMBAM and input parameters (e.g. number of (bagging) subsets) for PCHAT were not varied. Also, we did not study the effect of missing data. This was explored by Klei et al. for PCHAT who concluded that dropping individuals with missing data had a substantial diminishing effect on power of the test [4]. In addition, Van der Sluis et al. [15] reported that 10% missingness completely at random hardly affected the power to detect QTLs when the QTL affected all traits, but that it resulted in a higher power drop for MANOVA compared to TATES when the QTL was only associated to one of the traits. Finally, we did not simulate trait outliers in our data. O'Reilly showed that this could result in substantial inflation of the statistics for CCA for low frequency variants [12]. However, in our opinion outliers should be handled appropriately prior to association analyses.
