[BioC] Limma parameters

Gordon Smyth smyth at wehi.edu.au
Tue Dec 9 00:19:26 MET 2003

Dear Stephen,

A good question. It is actually very easy mathematically to estimate the 
proportion ('p') from the eBayes model. We don't do this at the moment 
because we don't think the simply heirarchical model fits the data well 
enough for the estimate of this parameter to be biologically meaningful. We 
thought it better to use a plug-in value rather than to risk misleading 
people re the true proportion of differentially expressed genes in their 

If you like, you can simply set 'p' to a value you're more comfortable 
with. In any case, this will not change the ranking of the genes using the 
log-odds. And the moderated t-statistics provide an almost equivalent 
inferential method that does not require 'p' at all.

Getting a good biologically believable estimate of 'p' is a non-trivial 
problem which we and others are working on. If you want a data driven 
method right now try this: convert the log-odds of differential expression 
to probabilities, then vary 'p' until it matches the mean probability.


At 12:49 AM 9/12/2003, Stephen Henderson wrote:
>  I've been examining the eBayes function in Limma using some affymetrix data
>we have. The default proportion setting: 0.01 I think is probably wrong for
>the comparison of tissue groups I have.
>Is there a rational way to estimate/guesstimate the correct proportion? Or
>isn't that the point anyway?

More information about the Bioconductor mailing list