[DESeq2] multiple variables analysis

[katymeala]

So I create a DeseqDataset like this:

dds<-DESeqDataSetFromMatrix (countData=countdata,colData=coldata,design=~A+B)

here A is one factor containing levels A1. A2 ; B is the other factor containg levels B1. B2

So the default

ddsMF<-DESeq(dds)

will tell me the influence of B1 and B2
-----

IF I want to look into factor A . I have two methods

first: results(ddsMF,contrast=c('A','A1','A2'))

second: dds <- estimateSizeFactors(dds)
dds <- estimateDispersions(dds)
dds <- nbinomLRT(dds,full=design(dds), reduced = ~ B)

So I want to know : what is the difference between above two methods for 'A' ?

[Michael Love]

The Wald test and the Likelihood ratio test are just different statistical tests (like the t-test vs the rank test). With this setup, they are testing for the same thing: differences due to A controlling for B. In DESeq2, they tend to perform very similarly in these two group comparisons.

If there were more than 2 levels to A, then the LRT is testing all levels at once, while the Wald test in DESeq2 is for comparing two groups, using 'contrast'. Since A has only two levels, the Wald and LRT are testing the same thing here.

Note that for the second, you can just do:

dds <- DESeq(dds, test="LRT", reduced=~B)

[katymeala]

The Wald test and the Likelihood ratio test are just different statistical tests (like the t-test vs the rank test). With this setup, they are testing for the same thing: differences due to A controlling for B. In DESeq2, they tend to perform very similarly in these two group comparisons.

If there were more than 2 levels to A, then the LRT is testing all levels at once, while the Wald test in DESeq2 is for comparing two groups, using 'contrast'. Since A has only two levels, the Wald and LRT are testing the same thing here.

Note that for the second, you can just do:

dds <- DESeq(dds, test="LRT", reduced=~B)

Thang you！