Dear all, I am using limma to perform differential analysis between two categories. Some samples in my set do not fall in either category. I can see that there is two ways to approach this: 1) Set the non-category samples to zero in the design matrix e.g. > design Low High GSM89690 0 1 GSM89724 0 1 GSM89728 0 0 GSM89737 0 1 GSM89692 1 0 GSM89693 0 0 GSM89716 0 1 GSM89718 1 0 GSM89726 1 0 GSM89730 1 0 GSM89746 0 0 GSM89747 1 0 GSM89751 0 0 GSM89695 0 0 GSM89739 0 1 GSM89687 1 0 GSM89699 0 0 GSM89701 0 0 GSM89703 0 0 GSM89706 0 0 GSM89708 0 1 GSM89709 0 0 GSM89712 0 0 2) Subset the expression matrix to remove the non-category samples, which results in the following design matrix: > design2 Low High GSM89690 0 1 GSM89724 0 1 GSM89737 0 1 GSM89692 1 0 GSM89716 0 1 GSM89718 1 0 GSM89726 1 0 GSM89730 1 0 GSM89747 1 0 GSM89739 0 1 GSM89687 1 0 GSM89708 0 1 After running limma on these sets, fit <- lmFit(SeminomaOnly,design2) cont.matrix <- makeContrasts(HvsL=High-Low, levels=design2) fit2 <- contrasts.fit(fit, cont.matrix) fit3 <- eBayes(fit2) topTable(fit3,adjust="BH"), I have found that these two approaches produce different results with approach 1 producing which seems to be faulty results. Could anyone explain if and why approach 1 is wrong? It is easy enough to subset the data in this situation, but when there is multiple contrasts this might prove trickier. Many thanks Daniel Brewer -- ************************************************************** Daniel Brewer, Ph.D. Institute of Cancer Research Molecular Carcinogenesis Email: daniel.brewer at icr.ac.uk ************************************************************** The Institute of Cancer Research: Royal Cancer Hospital, a charitable Company Limited by Guarantee, Registered in England under Company No. 534147 with its Registered Office at 123 Old Brompton Road, London SW7 3RP. This e-mail message is confidential and for use by the addre...{{dropped}}

Dear List, I have a question about normalization and would be very grateful for any comments. In my experiment, 4 control samples appear to be quite different from 4 treated samples, log intensity distributions of typical control and treated samples are attached to this letter. Could you please comment on whether the quantile normalization is applicable here, as distributions look quite different? Would it be more appropriate to use VSN here? As far as I understand, VSN does not assume identical distributions for arrays, but then, may be some other VSN assumptions may be violated here. I have thought about filtering out probes that are "unexpressed" across all samples and then quantile normalize (this would make distributions closer to each other), but I am not sure whether this is a good idea as many people prefer to normalize using ALL data. With kind regards, Lev. --------------------------------- -------------- next part -------------- A non-text attachment was scrubbed... Name: Control.pdf Type: application/pdf Size: 5281 bytes Desc: 2764920795-Control.pdf Url : https://stat.ethz.ch/pipermail/bioconductor/attachments/20070808 /265fc4fa/attachment.pdf -------------- next part -------------- A non-text attachment was scrubbed... Name: Treatment.pdf Type: application/pdf Size: 5053 bytes Desc: 3071396221-Treatment.pdf Url : https://stat.ethz.ch/pipermail/bioconductor/attachments/20070808 /265fc4fa/attachment-0001.pdf