Figure 2

Performance of the four different methods using simulations. Results shown in plots (A-D) are averaged over 1000 simulated datasets and results shown in plots (E-F) are averaged over 100 resampled datasets at each sample size. Plot (A) shows the cumulative numbers of differentially variable CpG probes containing outlier observations when ranking by the four different methods. There are no differentially variable CpGs simulated but 200 outliers are included in the data. Plot (B) shows ROC curves when 200 outliers and 1000 differentially variable CpGs are present in the simulated data. Plot (C) shows the control of the false discovery rate (FDR) of the four methods at a 5% nominal FDR cut-off (horizontal dashed black line) over ten different sample sizes. This simulation contains 1000 CpGs which are roughly five times more variable in Group 2 compared with Group 1. Plot (D) shows the power to detect differentially variable features at ten different sample sizes when Group 2 is roughly five times more variable than Group 1. Plot (E) shows the control of the FDR using resampled kidney cancer datasets at 11 different sample sizes. Plot (F) shows power to detect differentially variable features using resampled kidney cancer datasets at 11 different sample sizes.