[BioC] limma: eBayes / topTable and the choice of the "proportion" parameter for models with multiple contrasts

Gordon Smyth smyth at wehi.edu.au
Sat Dec 3 03:12:54 CET 2005

Dear David,

At 03:16 AM 2/12/2005, Dr. D. P. Kreil wrote:
>Dear Sir/Madam,
>I would be grateful for your comments regarding the interpretation of the 
>"proportion" parameter of limma's eBayes function.
>The documentation says: "assumed proportion of genes which are 
>differentially expressed".
>How does that relate to more complex linear models with multiple 
>coefficients/contrasts? In topTable, using the "coef" parameter, a linear 
>model coefficient or contrast can be specified. In that context, 
>differential expression directly makes sense. The eBayes routine, however, 
>does not take a "coef" parameter, so what is the meaning of 
>"differentially expressed" there for a linear model, in general?

The meaning is that the same proportion is assumed to apply for every 
contrast. This may admitedly be unrealistic.

>Is that 1 minus the proportion of genes that are non-differentially 
>expressed in all contrasts? In that case, the more complex the model, the 
>smaller "proportion" should be set, correct?

No. I can't see any reason to think so.

>   If trying to crudely estimate the parameter from the data, what would 
> you recommend?

I recommend not estimating it, for reasons that I explain in Section 6.4 of 
Smyth (2004) (http://www.statsci.org/smyth/pubs/ebayes.pdf). You're 
certainly free to vary the proportion based on your biological 
understanding of the process for each fit or contrast.

Best wishes

>I am very much looking forward to your comments / thoughts about 
>interpreting and/or setting the "proportion" parameter. (I know it doesn't 
>affect the ranking. We are nevertheless trying to get some meaning out of 
>the B statistic.)
>With many thanks for your help in adance,
>yours sincerely,
>(Boku University, Vienna)

More information about the Bioconductor mailing list