[BioC] limma eBayes: how to determine goodness of fit?
mrobinson at wehi.EDU.AU
Sat Jun 2 01:32:13 CEST 2007
> 1) You say, "The F-statistic will change since the variances are
> causing both the statistic to change and the degrees of freedom to
> I am afraid I don't understand this. In what circumstances does the
What I mean is that the F statistic you get from limma is moderated and is
different from the *classical* F test (which is what you'd get from the
'lm' or 'anova' commands in R). Have a read of Gordon's paper for the
> 2) I am interested in these goodness-of-fit measures because I want to
> through my data (and the eBayes output) to find genes whose behavior is
> very nicely modeled by different coefficients, like this
> a) devise a variety of model formualae based on biological
> and examinations of the data
> b) fit those models, one at a time
> c) identify different subsets of genes based on their fit to each
Can you give an example of what you mean here? I'm not sure what "Very
nicely modeled by different coefficients" and "variety of model formulae
based on biological intuition" are really referring to. Do you have a
designed experiment where you wish to look for differences between, say
treatments A, B, C, etc.?
> I remain a bit puzzled by the absence of these statistics from the eBayes
> output. Does
> their absence suggest that my 3-step procedure (2a-c, above) is not common
> practice? And
> would _that_ suggest that my 3-step procedure is not such a great idea?
> Are there better,
> more commonly used ways in limma to look for particular effects and
> of the experimental factors?
Common practice would be to make contrasts to the effect you want to
assess the significance of (if I understand what you are asking). There
are several examples in the limma user's guide:
More information about the Bioconductor