Table 1

Effect of the presence of DEG when applying the classic permutation strategy to the PLGEM-STN statistic.

FPR vs. significance level
 estimated vs. observed FDR


including DEG
 excluding DEG
 including DEG
 excluding DEG

# of genes in data set
 slope
 intercept
 adj. R^2
 slope
 intercept
 adj. R^2
 slope
 intercept
 adj. R^2
 slope
 intercept
 adj. R^2
22300
 1.690
 3.070
 0.856
 0.871
 -0.114
 0.938
 0.187
 -0.383
 0.314
 1.090
 -0.692
 0.826
10000
 1.710
 2.470
 0.881
 0.888
 -0.083
 0.931
 0.224
 -0.247
 0.433
 1.040
 -0.555
 0.824
5000
 1.460
 1.270
 0.880
 0.877
 -0.147
 0.934
 0.093
 -0.228
 0.067
 1.080
 -0.425
 0.836
2500
 1.590
 1.180
 0.888
 0.864
 -0.155
 0.939
 0.092
 -0.176
 0.135
 1.110
 -0.348
 0.853
1500
 1.670
 0.935
 0.908
 0.876
 -0.166
 0.948
 0.078
 -0.122
 0.170
 1.130
 -0.182
 0.880
1000
 1.720
 0.689
 0.874
 0.944
 0.002
 0.956
 0.038
 -0.122
 0.082
 0.991
 -0.226
 0.897
500
 1.900
 0.263
 0.864
 0.857
 -0.255
 0.956
 0.062
 -0.030
 0.307
 1.160
 0.038
 0.909
200
 2.490
 -0.059
 0.875
 0.905
 -0.378
 0.946
 0.064
 -0.012
 0.426
 1.050
 0.233
 0.914

Eight data sets with different percentages of DEG were constructed from the Latin Square data set by keeping the 62 known spiked-in probe sets, but randomly removing increasing amounts of the remaining probe sets reaching the total indicated in the first column. The null distributions of the PLGEM-STN statistics were evaluated through the classic permutation strategy either including or excluding DEG. A wide range of significance levels was used to select DEG and correlation between the FPR and the significance level or between estimated and observed FDR was evaluated through linear regressions in log-log plots. Table reports slopes, intercepts and adjusted r2 of linear models. See text for details on estimation of FDR.

Pavelka et al. BMC Bioinformatics 2004 5:203   doi:10.1186/1471-2105-5-203