[BioC] ebayes giving warning, should I trust the result?

[Gordon K Smyth]

> Date: Fri, 27 May 2005 17:18:18 -0400
> From: Guoneng Zhong <guoneng.zhong at yale.edu>
> Subject: [BioC] ebayes giving warning, should I trust the result?
> To: bioconductor at stat.math.ethz.ch
>
> Hi,
>
> Here is my code (thanks to all those who helped me with getting the columns
> right).
>
> library(limma)
> targets <- readTargets('targets.txt')
> f <- function(x) as.numeric(x$Flags > -99)
> RG<-read.maimages(targets$FileName,source="genepix",columns=list(Gf="F532
> Median",Gb="B532 Median",Rf="F635 Media
> n",Rb="B635 Median"))

What grounds do you have for thinking that this choice of columns is better than the default values?

> RG$printer <- getLayout(RG$genes)
> MA <- normalizeWithinArrays(RG)
> design <- modelMatrix(targets,ref="Reference")
> fit<-lmFit(MA,design)
> fit<-eBayes(fit)
>
> After the last line I get this warning:
> Warning message:
> Estimation of var.prior failed - set to default value in: ebayes(fit = fit,
> proportion = proportion, stdev.coef.lim = stdev.coef.lim)
>
> To make it even more unnerving is that when I typed "fit" to see what is in
> there, all the Ref values are NA (in the targets.txt file, I assigned Exp and
> Ref to various channels).

There shouldn't be any "Ref" values.  I suspect you have formed the design matrix incorrectly --
see below.

> Should I be worried about these NULL values and the warning message?

Yes, of course.  The warning about 'var.prior' is usually symptomatic of a degeneracy in your
data, so there is probably either a serious problem or oddity with your data or you've made a
mistake along the way.  I'm sure that, as a professional programmer, you understand that you have
to check the results line by line to make sure they are sensible.  Not checking until the end
means that you won't be able to locate where you went wrong.  For example, if you have used values
"Exp" and "Ref" in your targets.txt file, then why are you setting ref="Reference" rather than
ref="Ref" in modelMatrix()?  Have you looked at the design matrix to see it is what it should be? 
If you only have two RNA sources, then there should only be one column in your design matrix. 
Also, have you done plots of your data as read in to check that it is sensible?

Gordon

> Thanks again!
> G
>
> --
>
> Systems Programmer
> Yale Center for Medical Informatics
> fax: 203-737-5708