Figure 4

Simulated normal and log-normal data. (a) Normal distribution, 1% DEGs. As expected with independent normally distributed observations, the t-test will perform quite well, and is matched by samroc, which in this case equals the equal-variance t-test. The Bayes method has problems with these data, with SAM and Wilcoxon somewhere in between these extremes. (b) Lognormal distribution, 1% DEGs. samroc may have a slight advantage for shorter lists, whereas the Bayes method is better for longer lists, where the number of false positives is larger. The other three methods lag behind, but not by much. (c) Normal distribution, 5% DEGs. The t-test and samroc coincide; samroc is now equivalent to the equal-variance t-test, which behaves in the same way as the unequal-variance t-test in this case. (d) Lognormal distribution, 5% DEGs. The difference between methods is less when data are exponentiated. However, samroc has the edge for a wide range of cutoffs, but the Bayes method catches up when more genes are selected. The other three methods are struggling to avoid last spot. (e) Normal distribution, 10% DEGs. Again the samroc and the t-test coincide and the Bayes method has problems with normal data. SAM is also lagging behind, while the other three are very close together. (f) Lognormal distribution, 10% DEGs. samroc comes out well; Wilcoxon has the worst performance, SAM and the t-test are scarcely better, while the Bayes method is intermediate.