Figure 3
A subset of the data from Golub et al consisting of the first four samples from each group, was used to assess the performance of the t-test, SAM, samroc and the Bayesian method. The upper panel shows the ordered observed d-statistics versus the expected ordered statistics calculated from the simulation. Looking at the graph it appears that rather a lot of genes are changed. The lower panel shows the estimated FDR as a function of the cut-off δ, such that genes with |d-d expected | >δ will be called differentially expressed. Choosing a cut-off of 0.7, i.e. calling genes that fulfill |d-d expected | > 0.7 differentially expressed will only give about 3% false positives. The proportion of unchanged genes is estimated at 84%, and S0 in (1) is put to the 5% fractile of the SE's. The graph was produced by the samfdr function in SAG.