Dear Benno, It makes perfect sense to remove from your Illumina analysis probes which seem to be entirely unexpressed throughout your experiment. This is a phenomena which is especially noticeable with Illumina arrays. However, to use this strategy, you must remove probes entirely, not selectively for some arrays but not for others. The simplest and probably the best way to do this is to do a histogram of Amean values (obtainable from the output of lmFit for example), and select a value to subset your ExpressionSet object on. The use of detection pvalues or weights does not seem useful to me. Filtering low intensity Illumina probes will increase both the consensus variance values (s2.prior) and the prior degrees of freedom (df.prior) that you get from eBayes, as well as reducing the amount of multiple testing. These have competing effects, so the amount of statistical significant can increase or decrease depending on the situation. Best wishes Gordon On Thu, 31 Jan 2008, Benno Puetz wrote: > Dear Dr. Smyth, > > as I have not received any reply to a question posted on the BioC > mailing list, I hope you will excuse my asking you directly. > > I have a data set from Illumina gene expression arrays processed with > beadarray and limma. > > The data contain a lot of genes with low expression (about background) > with correspondingly low within gene variation. From my understanding > this would cause an overly aggressive shrinking and an inflated > significance. > > To counter that effect I was looking at the "Detection Pval" measure > provided by Illumina's scanning software and tried using (1-DetPVal) as > weight in lmfit() which let to the expected change in the p-values. > > As I have not found this mentioned anywhere, my question to you is now: > is this approach valid or did I miss something along the line? > > I am looking forward to your reply. > > With best regards > > Benno