hi Ryan, i examined the design with: > is.fullrank(design) [1] TRUE > (in each group the individuals are numbered 1-100) this design was okay with voom(), which was indeed fast (as Michael and Gordon suggested), but with edgeR estimation of dispersion took too long, and in DESeq2 it failed with an error thank you, iddo On Jun 16, 2014, at 1:40 AM, Ryan <rct at="" thompsonclan.org=""> wrote: > Hi, > > The full design as you have specified it is not of full rank, so I would expect the dispersion estimation to fail with an error. This is because the individual factor is (I assume) nested within the group factor (i.e. every individual belongs to exactly one group). I think your situation is similar to a recent post on this list: > > https://stat.ethz.ch/pipermail/bioconductor/2014-May/059579.html > > In the case, again there are multiple individuals in each of two groups with before and after treatments. My answer is here: > > https://stat.ethz.ch/pipermail/bioconductor/2014-May/059587.html > > You could do the same thing for your data, except that you don't have to do the duplicateCorrelation step because you don't have technical replicates. You can use the same design for limma or edgeR. I don't know if there is a way to specify this design for DESeq2. > > -Ryan > > On 6/12/14, 6:51 AM, Iddo Ben-dov wrote: >> hi, >> >> in both edgeR and DESeq2, estimation of dispersion precedes negative binomial GLM fitting. >> >> my question is, can I use a design formula when estimating dispersion which is different from the formula used for GLM fitting? specifically, I would like to use a simplified design when estimating dispersion and a full design for GLM fitting. >> >> my motivation for doing so is that with the full design estimation of dispersion is too demanding for my computer and time. >> >> my dataset includes 400 mRNAseq profiles (~22,000 genes). there are 100 controls and 100 cases, and each was sampled twice - before and after intervention. >> >> thus, the full design is: >> ~ group*intervention + individual:group (blocking factor) >> >> as I mentioned, estimation of dispersion with the above design is not practical, and I thus would like to simplify to: >> ~ group*intervention >> >> and introduce the 'individual' blocking factor only for NB GLM fitting. >> >> is this statistically valid? >> >> appreciate any help, >> iddo >> >> _______________________________________________ >> Bioconductor mailing list >> Bioconductor at r-project.org >> https://stat.ethz.ch/mailman/listinfo/bioconductor >> Search the archives: http://news.gmane.org/gmane.science.biology.informatics.conductor >