Most of the experiments completed by our core facility have fewer than 10 samples per treatment group, typically 4–6, and some in the literature use fewer than 4. To understand the effects of sample size on filtering we created smaller virtual experiments (with 3, 4, 5, 6, 7, and 8 samples per treatment group) from the IFN data by using permutations of the original data. We randomly selected arrays without replacement within each treatment group; this created virtual experiments in which a difference in expression is expected. There is a large increase in power as the sample size increases (Fig. 6) especially for the most significant probe sets (p ≤ 0.001): a 15-fold increase from 3-sample experiments to 8-sample experiments. The effects of filtering in the smaller experiments were similar to the effects seen in the full experiment (Fig. 6). The more significant probe sets (p ≤ 0.001) were mostly retained when filtering was at ≤ 50% Present (Table 5). The FDR is similarly improved by filtering (Fig. 7; note the differences in scales in each panel for Figs. 6
