Filtering by fraction Present improves both parametric (Welch's t-test) and non-parametric analyses (e.g. SAM). For t-tests, this type of filtering rarely removes very significant probe sets (p ≤ 0.001). Although similar results in FDR improvement can be achieved filtering by average signal or RMA value, approaches based on expression level are more likely to remove genes that are being turned on or off. Permutations of data expected to produce no true positives resulted in fewer false positives than predicted by the Benjamini and Hochberg method [1], and demonstrated that probe sets called Absent produce a disproportionate fraction of false positives.
