[BioC] DESeq2: df error message with nbinomial LRTest

INRA - coutellec marie-agnes.coutellec at rennes.inra.fr
Sun Aug 31 12:15:18 CEST 2014

Dear all,

My question is about the right design to use to perform LRT tests with 
DESeq2. I have two crossed factors, tra (2 levels) and pop (4 levels). I 
test the interaction term as follows:

>  ddsFullCountTable <- DESeqDataSetFromMatrix(

+ countData = countE,

+ colData = cda,

+ design = ~ pop*tra)

 > dds2inter.LRT <- DESeq(ddsFullCountTable, test="LRT", full=~pop*tra, 
 > res2<-results(dds2.LRT)
 > sum( res2$padj < 0.1, na.rm=TRUE )
[1] 389

and get 389 DEGs (out of 45249 contigs). Although this is very fine with 
me, I would like to know what  exactly is listed in column 
'log2foldchange' for this interaction term (change between what and what?).

Next, I want to test for the "tra" effect, alone. I tried :

 > dds.LRT <- DESeq(ddsFullCountTable, test="LRT", full=~pop*tra, 
 > dds.LRT <- DESeq(ddsFullCountTable, test="LRT", 
full=~pop+tra+pop:tra, reduced=~pop+pop:tra)

and get the same error message:
Error in nbinomLRT(object, full = full, reduced = reduced, betaPrior = 
betaPrior,  :
   less than one degree of freedom, perhaps full and reduced models are 
not in the correct order

whereas the command below works, although the two models differs by two 
elements (which shouldn't work in my view):
 > dds.LRT <- DESeq(ddsFullCountTable, test="LRT", 
full=~pop+tra+tra:pop, reduced=~pop)

I understand it is somehow meaningless to eliminate a factor while 
keeping it through its interaction in a model. Thus, I plan to:
1) take out the 389 "interaction" DEGs
2) reestimate size factors and dispersion with an additive design 
(design=~pop+tra), from which I would be able to test the effect of 
"tra" and of "pop" very easily.

Can anyone tell me if this is correct, or if I alternatively should keep 
"pop*tra" as design to estimate dispersion and then use "pop+tra" as 
full model and "pop" as reduced model in my LRT.

Many thanks in advance,

Marie-Agnès Coutellec
INRA/Agrocampus-Ouest UMR0985 Ecology and Ecosystem Health
Ecotoxicology and Quality of Aquatic Environments
65 rue de Saint-Brieuc
35042 Rennes cedex - FRANCE
phone: +33 223 485 248

	[[alternative HTML version deleted]]

More information about the Bioconductor mailing list