Different methods have been used to pre-filter data to remove data for probe sets that are believed to be less reliable, but the effects of such pre-filtering are rarely analyzed quantitatively. Filtering by signal (expression level) [7] removes probe sets with signal close to background; the choice of how close to background is arbitrary. Removal of probe sets that are called Absent on all arrays has been reported [8]. Some use post-hoc methods by eliminating significant probe sets with low fold changes [9]; again, the choice of fold-change is arbitrary and there is no theoretical null distribution. Others use combinations of these strategies: trimming upper and lower signals plus fold change filter [10]; confidence score based on fold change, p-value, signal and percent Present [11]; signal, fold change and percent Present filters [12]; minimum signal and fold change [13]. Most of these studies did not provide a reference for their selection of filtering criteria or provide evidence that these strategies were helpful. Aston et al. [12] chose their fold-change threshold based on fold changes seen in previous microarray studies of
