Our approach of requiring at least one group to meet the threshold insures that the gene is expressed at a detectable level in at least one biologically defined group; that seems a priori preferable to requiring a (lower) fraction present across the entire experiment, because the latter could include genes that are not reliably detected in any biological condition. To examine whether our choice led to a significant bias, we compared these two approaches. Using a fraction Present >0 (requiring just a single Present array) is, of course, identical for both methods. Comparing thresholds of 25% or 50% Present in at least one group (our method) with thresholds across the entire experiment that led to retention of similar numbers of probe sets, the number significant at any particular p-value was within 2%. filtering based on fraction Present in at least one group was much better at preserving probe sets turned on or off than using a global threshold.
