[BioC] Limma nestedF

[James W. MacDonald]

Hi Noelle,

noel0925 at sbcglobal.net wrote:
> Hi All,
> 
> I am wondering if someone can explain when it is appropriate to use
> the nestedF method of the decideTests function in Limma?
> 
>> From the manual I am aware that this adjusts down genes then across
>> contrasts and that this strategy is recommended for complex
>> experiments with many contrasts (like mine) since it may be
>> "desirable to select genes firstly on the basis of their moderated
>> F-statistic, and subsequently decide which of the individual
>> contrasts are significant for the selected gene.
> 
> I am interested in identifying genes that are differentially
> expressed in many contrasts- in particular across genotypes that are
> fairly similar, across treatments, and in various genotype:treatment
> combinations (ie interaction effects). I expect that the genotype
> effect will alter the same genes for some of the samples. I also
> expect that the treatment effect will alter some of the same genes
> across all the genotypes.
> 
> Is this an appropriate situation in which to use a nestedF test (I
> will also correct for multiple testing using "BH")? Is it correct to
> form a contrast matrix for all my contrasts of interest (including
> interaction effects) and test them all simultaneously using
> decideTests(x, method="nestedF", method.adjust="BH", p=0.05)?

I am not sure the nestedF method is appropriate for this situation, 
because you have interaction terms. When there is an interaction term in 
the model, the usual thing to do is to check for significance of the 
interaction term, and if it isn't significant, then you would drop it 
from the model and check for significance of the main effects terms. The 
nestedF method won't do this - it will treat all the contrasts as if equal.

Another issue that arises with microarray analyses is the usefulness of 
main effects in general. As Gordon mentions in the Limma User's Guide, 
the main effects are usually not that interesting for microarray data, 
and are often nonsensical.

As an example, say you have wild type and knockout mice that you treated 
with a control vehicle and a drug. Usually one would be interested in 
finding those genes that are affected differently in the KO and WT mice 
when treated with the drug (i.e., the interaction). It is usually not 
interesting to ask which genes are differentially expressed in treated 
vs untreated mice regardless of KO/WT status (treatment main effect), or 
which genes are differentially expressed in WT vs KO mice regardless of 
treatment status (mouse main effect).

Using the nestedF approach with various interactions and main effects 
looks (to me) like a 'shotgun' approach to the analysis. I think you 
would be much better served to approach the analysis in a stepwise 
manner, testing each particular hypothesis separately.

Best,

Jim



> 
> I have done so and as can be expected this returned a far larger
> number of DE genes (compared to decideTests(method = "separate",
> method.adjust="BH", p=0.05). In fact, the number of genes called
> significant by this approach for some of my contrasts is quite alot ~
> 5,000/24,000  (only ~ 100 for others) and perahps more than I perhaps
> want use in for example other analyses such as GeneSet enrichment or
> GO analyses, or heat maps.
> 
> If I wish to only consider a smaller number of genes, is it more
> correct in a statistical sense to use a more stringent p-value cutoff
> after performing decideTests(x, method= "nestedF",
> method.adjust="BH", p=0.0001) or to consider the contrasts separately
> and use a larger p-value cutoff? Clearly, some of these tests are not
> independent, so I am inclined to go with the nestedF approach and a
> more stingent cutoff.
> 
> I have read through other postings, and found this one especially
> helpful: 
> https://stat.ethz.ch/pipermail/bioconductor/2006-March/012182.html 
> but am still uncertain about my approach. Any comments would be
> appreciated.
> 
> Thanks in advance, Noelle
> 
> [[alternative HTML version deleted]]
> 
> _______________________________________________ Bioconductor mailing
> list Bioconductor at stat.math.ethz.ch 
> https://stat.ethz.ch/mailman/listinfo/bioconductor Search the
> archives:
> http://news.gmane.org/gmane.science.biology.informatics.conductor


-- 
James W. MacDonald, M.S.
Biostatistician
Affymetrix and cDNA Microarray Core
University of Michigan Cancer Center
1500 E. Medical Center Drive
7410 CCGC
Ann Arbor MI 48109
734-647-5623


**********************************************************
Electronic Mail is not secure, may not be read every day, and should not be used for urgent or sensitive issues.