The FDR for any particular p-value is an estimate of the percentage of probe sets expected to be false positives for the group of probe sets that have a p-value less than or equal to the selected p-value. FDR is based on the number of probe sets used in the analysis (retained after the filtering). Filtering by fraction Present or by signal or RMA value have similar effects on the FDR calculated according to Benjamini and Hochberg [1] (Fig. 5). The largest stepwise improvement in FDR occurs with the initial filtering, Present fraction > 0% or removal of an equivalent number of probe sets by an average MAS5 signal ≥ 254 or RMA value ≥ 4.25. Increasing the stringency of the filter continues to improve the FDR. However, in the IFN experiment (10 arrays in each group), increasing stringency beyond 25% Present, MAS5 signal ≥ 375, or RMA value ≥ 4.75 leads to little improvement in FDR for the more significant probe sets (p ≤ 0.001), while increasing the number of those very significant probe sets lost (Table 2). Results
