Another way of analyzing the likely true positives in smaller experiments is by examining the reproducibility of results. We assumed that probe sets deemed significant at least 50% of the time (i.e. in ≥ 500/1000 permutations) represents consistent, reproducible data. The number of consistently significant probe sets increases as the experiment size increases and approaches the number identified as true positive probe sets (Fig. 8). In the smaller virtual experiments (3 or 4 samples per treatment group) only 2–4% of the probe sets consistently significant at p ≤ 0.05 are lost when filtering by 50% Present. In larger experiments (5 or 6 samples per group) 6–9% of these consistently significant probe sets were lost when filtering by 25% Present (Fig. 8). For those probe sets consistent at p ≤ 0.01, the number lost by filtering is even smaller, almost none for 3–4 samples at 50% Present and 1–2% for 5–6 samples per group at 25% Present. Only 2 probe sets in the 3-sample permutations were found to be significant at p ≤ 0.001 at least 50% of time, whereas 65
