We would like to emphasize that all methods were compared based on empirically derived significance levels, adjusting each method to an exact 5% type I error rate. Null simulations illustrated that for MV-PLINK, MV-SNPTEST, MultiPhen, MV-BIMBAM, TATES and UV-PCA these empirical significance levels were all close to the nominal level of 0.05 for p-values or between 0.01-1 for log10 BF [28]. PCHAT was slightly conservative based on our observations. Thresholds for UV analysis were around 1.7% (i.e. 5 divided by 3) in case of uncorrelated traits, which is in line with a Bonferroni correction for three independent tests. As expected, for correlated traits the adjusted threshold was less stringent and somewhere between 1.7 and 5%. We found that significance thresholds for UV-MA were highly dependent on trait correlations with increased stringency with increase in correlation. Trait correlations result in longer tails for the test statistic distribution, and therefore a more stringent threshold must be applied to keep the type I error at 5%. Thus, use of the UV-MA can potentially lead to a high number of false-positive findings if traits are highly correlated and the significance threshold is not appropriately adjusted.
