Effect of variable selection under the null hypothesis. The figure shows the Class 1 predictive accuracy (PA1) obtained varying the proportion of samples from Class 1 in the training set (), using nine classifiers. The training set contained 80 samples, and p = 40, 1000 or 10000 variables were generated from the same distribution for both classes (N(0, 1)); 40 variables were selected (G = 40). The test set was balanced ( = 0.5) and contained 20 samples. The PA were evaluated on the test set. Samples were mean centered. Details on data generation are reported in the Methods section.