While there is no gold standard for the IFN data, the full 10-sample dataset provides a reasonable standard to which the smaller virtual experiments can be compared. The Benjamini and Hochberg [1] FDR for the full experiment (without filtering) at p = 0.05 is 25%, the Storey [2] FDR is 17%, and the estimate by permutation is 20%. Therefore, a crude estimate of false discovery can be made by comparing the results from the small experiments to the results using all samples, assuming that any probe set that was called significant in the smaller experiments but had a p-value > 0.05 in the full experiment (without filtering) was a false positive. The number of false positives identified in this manner is slightly smaller than was found by permutations, and remains about constant for experiments of size 3 to 8 samples per treatment group (Fig. 8). The number identified as "true positives" (i.e. matching the full 10 sample experiment) at p ≤ 0.05 increases with the size of the experiment. Filtering at 50% Present removes about 50% of the false positive
