DESeq newCountDataSet with seven groups

[yve]

Hello,

I have a question regarding the differential analysis of miRNAs with DESeq. I have seven different groups (all with replicates), one control group and six groups corresponding to different cells.
I want to do pairwise comparisons between the control group and each of the other groups. Therefore I did the following:

countData <- newCountDataSet(countTable, groups)
countData <- estimateSizeFactors(countData)
countData <- estimateDispersions(countData, method="pooled")

Next I did the following (group 1 being the control group):

res <- nbinomTest( countData, group1, group2 )
res <- nbinomTest( countData, group1, group3 )
res <- nbinomTest( countData, group1, group4 ) ...

Now my question is, if this is the right way to do this analysis or if I should estimate the size factors and dispersion using only the "needed data" for the specific analysis:

countData <- newCountDataSet(countTable[,c(group1, group2)], groups)
countData <- estimateSizeFactors(countData)
countData <- estimateDispersions(countData, method="pooled)

res <- nbinomTest( countData, group1, group2 )

Thanks in advance!

Yvonne

[Simon Anders]

There are arguments in favour of either way. Estimating the dispersion from all samples has more degrees of freedom and hence yields more precise estimates, which (due to DESeq's "maximum rule") translates into better power.

On the other hand, if replicates agree badly in one group, this will drive up the dispersion estimates for the full data and hence costs power for all comparisons, unless you do everything separately.

I would use the full data, but first check (with a sample clustering after a variance-stabilizing transformation) that there are no bad samples.