We used permutations of the IFN data in which 5 samples from each treatment group were combined to produce two new groups expected to show no difference. Welch's t-tests on the MAS5 log transformed data produced fewer nominally significant probe sets than expected by chance (Table 3): 0.04 at a nominal p = 0.05 and 0.0005 at a nominal p = 0.001. In these balanced permutations, 37% of probe sets are Absent on all arrays but 43% of the probe sets with p ≤ 0.05 are found in this "all-Absent" group, demonstrating that the Absent probe sets make a disproportionate contribution to false positives. Similar analyses of the smoking data (Fig. 5C) also produced fewer nominally significant probe sets than expected by chance.
