Dear Gordon and Limma users, I am using limma 2.12.0 (also tested 2.14.3) to analyze two-color arrays which have a common reference and I think I am not getting the expected output from lmFit() and eBayes(). Data from these arrays is not very good and I used weights to flag spots which did not meet the quality criteria I had set. Then: >RG.b <- backgroundCorrect(RG, method="normexp",offset=100) >MA.norm <- normalizeWithinArrays(RG.b, method="loess", weights=RG.b$weights, bc.method="none") >MA.norm.AQ <- normalizeBetweenArrays(MA.norm,method="Aquantile") >MA <- MA.norm.AQ >design <- modelMatrix(targets,ref="Ref") >fit1 <- lmFit(MA, design,weights=MA$weights) >fit2 <- eBayes(fit2) >write.fit(fit2,file="none.txt",adjust="none") >write.fit(fit2,file="fdr.txt",adjust="fdr") If I am not mistaken, in previous limma versions whenever you had only one measurement per row and the rest were "NA", no p.values were calculated (which makes sense to me). I believe the same happened if only one weight for a spot was 1 ( being the others equal to 0). I have narrowed down the problem to an MAList with 2 arrays (dye-swap) and checked that weights and design are well set. With these versions (2.12.0 and 2.14.3) I have tried: - weight 0 all spots which are not good - write NA in MA$M and MA$A for all spots which are not good - weight 0 and write NA in MA$M and MA$A for all spots which are not good And the "problem" persists: p.values are calculated when there is only one measurement and one NA or weights are 0. Example of the third case: >MA$M[63:64,] Array1 Array2 [1,] -0.5069863 0.8276455 [2,] -1.2127786 NA >MA$A[63:64,] Array1 Array2 [1,] 10.35034 10.14791 [2,] 10.32554 NA >MA$weights[63:64,] Array1 Array2 [1,] 1 1 [2,] 1 0 Results I get: Gene logFC AveExpr t P.Value adj.P.Val B 1 0.67 10.25 3.48 0.004 0.1225 -1.91 2 1.21 10.33 4.43 0.001 0.0428 -0.12 Am I getting what is expected? What could possibly be causing this behaviour? Thank you very much for your time and consideration, Sergi Beltran Universitat de Barcelona sergiba at yahoo.com #### Might be worth reading: http://thread.gmane.org/gmane.science.biology.informatics.conductor/17 082/focus=17100 #### R version 2.6.0 (2007-10-03) x86_64-redhat-linux-gnu locale: LC_CTYPE=en_US.UTF-8;LC_NUMERIC=C;LC_TIME=en_US.UTF-8;LC_COLLATE=en_US .UTF-8;LC_MONETARY=en_US.UTF-8;LC_MESSAGES=en_US.UTF-8;LC_PAPER=en_US. UTF-8;LC_NAME=C;LC_ADDRESS=C;LC_TELEPHONE=C;LC_MEASUREMENT=en_US.UTF-8 ;LC_IDENTIFICATION=C attached base packages: [1] tools stats graphics grDevices utils datasets methods [8] base other attached packages: [1] R2HTML_1.59 doBy_3.0 fields_4.1 [4] spam_0.13-3 geneplotter_1.16.0 lattice_0.17-8 [7] annotate_1.16.1 xtable_1.5-2 AnnotationDbi_1.0.6 [10] RSQLite_0.6-8 DBI_0.2-4 Biobase_1.16.3 [13] sma_0.5.15 limma_2.12.0 loaded via a namespace (and not attached): [1] cluster_1.11.10 grid_2.6.0 Hmisc_3.4-3 KernSmooth_2.22-22 [5] RColorBrewer_1.0-2 rcompgen_0.1-17 [[alternative HTML version deleted]]

On Wed, 4 Jun 2008, sergiba at yahoo.com wrote: > Dear Gordon and Limma users, > > I am using limma 2.12.0 (also tested 2.14.3) to analyze two-color arrays which > have a common reference and I think I am not getting the expected output from > lmFit() and eBayes(). > > Data from these arrays is not very good and I used weights to flag spots which > did not meet the quality criteria I had set. > If I am not mistaken, in previous limma versions whenever you had only one > measurement per row and the rest were "NA", no p.values were calculated (which > makes sense to me). I believe the same happened if only one weight for a spot > was 1 ( being the others equal to 0). You are mistaken. limma has always been able to give p.values, even when only a single observation is available, provided the contrast is estimable. For these genes, the prior residual standard deviation is used. Best wishes Gordon