[BioC] limma: decideTests, which option to chose?

[Gordon K Smyth]

> Date: Wed, 29 Jun 2005 09:54:11 +0200
> From: kerick at molgen.mpg.de
> Subject: [BioC] limma: decideTests, which option to chose?
> To: bioconductor at stat.math.ethz.ch
>
> Dear all,
>
> to the best of my knowledge these questions haven't been posed so far; I
> apologize if I didn't dig deep enough.
> I am seeking advice on which option to choose from the decideTests function in
> limma. Since the methods nestedF, heirarchical, global and separate (all using
> fdr adjustment) give pretty different numbers of genes being classified as
> significant I wonder which choice is statistically speaking the most stringend
> and which is the least stringend. Any idea how to line up all choices with
> regard to increasing stringency? nestedF < heirachical < global <
> separate  ..maybe?

"separate" is less stringent than the others.  The others can't be properly ordered because it
depends on the type of differences found in the data.  For example, the nestedF method is most
powerful for picking up genes which change in multiple conditions, but is possibly least powerful
for picking up genes which are different in only one condition.

Gordon

> practical questions:
> We performed an experiment which tries to answer seven questions(~ 7
> contrasts).
> I chose the nestedF option to get a fair amount of "differentially regulated"
> genes for each contrast to then perform fisher.test analysis on geneontology
> groups for each contrast separately.
> Then I chose the heirarchical option to have a closer look at the genes one by
> one for each contrast at hand.
> Any doubts concerning this strategy?
>
> I appreciate any comment,
> Best regards,
>
> Martin