Thanks, Ben. I haven't seen your response probably due to problem in the way I send my e-mail or the way my e-mail address is subbmited to the server. Anyhow, what you mean that when lmFit identifies PLMset object it requires weights either from the PLMset object or by the users. And in the PLMset object that generated by rmaPLM does not supply weights. Is that so? Thanks, Ron II'd recommend that you use fitPLM rather than rmaPLM for these purposes.The reason why it is giving "reasonable results" in the first case andan error in the other case, is that it is using your user suppliedweights in the earlier situation and trying to derive some sensibleweights based on the PLM model fit. Unfortunately rmaPLM does not returnthe information that is used to construct these but fitPLM does.Ben On Mon, 2005-09-05 at 13:51 +0300, Ron Ophir wrote: > Hi > I am analyzing affymetrix data having the following experimental> design:> > analysis$design> Normal IOP IOPC > L91RAE230 2.CEL 0 1 0 > L92RAE230 2.CEL 0 1 0 > L93RAE230 2.CEL 0 0 1 > L94RAE230 2.CEL 0 0 1 > L95RAE230 2.CEL 1 0 0 > L96RAE230 2.CEL 1 0 0 > and following contrasts > > analysis$contrasts > IOPC.IOP IOPC.Normal IOP.Normal> Normal 0 -1 -1> IOP -1 0 1 > IOPC 1 1 0 > The preprocessing was done using affyPLM packge > analysis$data<-rmaPLM(analysis$raw)> where analysis$raw is AfyyBatch object. Then I fitted the the model> once with weighting matrix using AMP flags > fm<-as.matrix(analysis$Flags!="A") > analysis$fit<-lmFit(analysis$data,analysis$design,w=matrix(as.numeric( fm),dim(fm)[1],dim(fm)[2],dimnames=dimnames(fm)))> and got a reasonable results and once without this matrix. What was> surprising is that fitting without the weighting matrix I got the > following error message: > > analysis$bayes<-eBayes(analysis$contrasts.fit)> Error in ebayes(fit = fit, proportion = proportion, stdev.coef.lim = > stdev.coef.lim) : > No residual degrees of freedom in linear model fits> I've checked the fit coeffecient matrix (analyis$fit object) and is all NA. > Does someone have an idea why? I did not applied any weighting matrix> for this fitting. > Thanks > Ron > > > > > [[alternative HTML version deleted]] > > _______________________________________________ > Bioconductor mailing list> Bioconductor at stat.math.ethz.ch> https://stat.ethz.ch/mailman/listinfo/bioconductor

Indeed, lmFit when it identifies a PLMset object tries to derive weights using the se.chip.coef slot of the PLMset object. By default rmaPLM does not return these but fitPLM would. Looking back through your previous code it appears as though what you were trying to do is compare situations where you gave absent probesets 0 weight and everything else a weight of 1 with a situation where everything is unweighted (ie essentially equal weights). You could carry out this second situation by supplying a matrix of ones as your weights if supplying a PLMset object or alternatively carryout RMA using rma() and supply the resultant exprSet without weights to lmFit. Ben On Tue, 6 Sep 2005, Ron Ophir wrote: > Thanks, Ben. I haven't seen your response probably due to problem in the > way I send my e-mail or the way my e-mail address is subbmited to the > server. Anyhow, what you mean that when lmFit identifies PLMset object > it requires weights either from the PLMset object or by the users. And > in the PLMset object that generated by rmaPLM does not supply weights. > Is that so? > Thanks, > Ron > > II'd recommend that you use fitPLM rather than rmaPLM for these > purposes.The reason why it is giving "reasonable results" in the first > case andan error in the other case, is that it is using your user > suppliedweights in the earlier situation and trying to derive some > sensibleweights based on the PLM model fit. Unfortunately rmaPLM does > not returnthe information that is used to construct these but fitPLM > does.Ben >

>Date: Mon, 05 Sep 2005 13:51:38 +0300 >From: "Ron Ophir" <ron.ophir at="" weizmann.ac.il=""> >Subject: [BioC] LIMMA No residual using rmaPLM >To: <bioconductor at="" stat.math.ethz.ch=""> >Message-ID: <s31c4d86.078 at="" wisemail.weizmann.ac.il=""> >Content-Type: text/plain > >Hi >I am analyzing affymetrix data having the following experimental >design: > > analysis$design > Normal IOP IOPC >L91RAE230 2.CEL 0 1 0 >L92RAE230 2.CEL 0 1 0 >L93RAE230 2.CEL 0 0 1 >L94RAE230 2.CEL 0 0 1 >L95RAE230 2.CEL 1 0 0 >L96RAE230 2.CEL 1 0 0 > and following contrasts > > analysis$contrasts > IOPC.IOP IOPC.Normal IOP.Normal >Normal 0 -1 -1 >IOP -1 0 1 >IOPC 1 1 0 >The preprocessing was done using affyPLM packge >analysis$data<-rmaPLM(analysis$raw) >where analysis$raw is AfyyBatch object. Then I fitted the the model >once with weighting matrix using AMP flags >fm<-as.matrix(analysis$Flags!="A") >analysis$fit<-lmFit(analysis$data,analysis$design,w=matrix(as.numeric (fm),dim(fm)[1],dim(fm)[2],dimnames=dimnames(fm))) > and got a reasonable results and once without this matrix. What was >surprising is that fitting without the weighting matrix I got the >following error message: > > analysis$bayes<-eBayes(analysis$contrasts.fit) >Error in ebayes(fit = fit, proportion = proportion, stdev.coef.lim = >stdev.coef.lim) : > No residual degrees of freedom in linear model fits >I've checked the fit coeffecient matrix (analyis$fit object) and is all >NA. >Does someone have an idea why? I did not applied any weighting matrix >for this fitting. As Ben has already explained, you are implicitly using a weighting matrix here because your 'PLMset' object contains a slot 'se.chip.coefs' from which limma is attempting to construct probe-set weights. (Basically the weight for each expression value is the inverse squared se, with some moderation to prevent absurd values.) You can easily turn this off by using lmFit(..., weights=NULL) The philosophy is this: If the se.chip.coefs slot is empty, then limma will take the se's and hence the weights to be all equal. If however this slot is set to be a matrix with entirely NA values, then limma will assume (and this is the crucial thing) that the se's are not only missing but non-ignorable. Hence it passes NA weights to the lmFit and hence you end up with no useful data. I am open to suggestion that limma should do something different. My feeling which motivates the current behaviour is that slots in data objects are only set (or should only be set) when they really mean something. Hence they should not be ignored without specific direction to do so, even if they contain NA values. Gordon >Thanks >Ron