We have filtered out probe sets that were not called Present by the MAS5 detection call in at least 50% of the samples in one treatment group [15,16]. Requiring the probe set to meet the criterion in any one treatment group retains genes that are turned on or off; these are usually very interesting to biologists. Our method, which we call filtering by fraction Present, improves FDR. Here, we evaluate the effects of setting different thresholds and compare this approach to the use of thresholds for signal values. These analyses were performed using expression level data generated by two widely used algorithms, MAS5 and RMA (robust multichip average [17,18]). RMA is an alternative log scale measurement of expression derived from only the perfect match probes by fitting them to a linear model normalized across all arrays in the experiment. RMA does not provide a measure comparable to the MAS5 detection call. We examined the effects of the fraction Present filtering on the ability to detect changes in expression using a parametric t-test and also using a permutation-based test, significance analysis
