DESeq 2 shrinkage Wald test vs. LR test

[mike_m]

This is more a conceptual question to those in the know (particularly Mike Love, Simon Anders and Wolfgang Huber if you're reading).

I'm doing a slightly quick a dirty analysis using DESeq2 to test for interactions between 3 biological conditions and expression changes over time (only between 2 time points, so one hope issues of sphericity/autocorrelation do not become an issue).

The current implementation using the Wald test will, by default, apply a beta prior to moderate the effect sizes from the GLM. I understand the default implementation as the Wald test uses the effect size estimates to calculate the p-value that the effect is different from zero. The likelihood ratio test does not require an accurate effect size estimate as I understand it, therefore I assume that this is the reason there is no default implementation for the effect size estimate moderation, or is there a mathematical reason for this?

Also, is there a way to extract the pre-moderated effect sizes in order to compare them to the moderated ones (regardless of which test is applied?).

Thanks in advance.

[Michael Love]

hi Mike,

Here's a post where I talk a bit about why we used the moderation of effect sizes only for the Wald test:

http://seqanswers.com/forums/showpost.php?p=149176&postcount=7

So it's default to have shrinkage with the Wald test, and no shrinkage with the LRT. Although both defaults can be overridden.

Note that with interaction terms in the model, the shrinkage is only applied to the interaction terms (discussed in ?nbinomWaldTest and in the Methods of the preprint).

There is a way to extract pre-moderated effect sizes in DESeq2 version 1.6, which will be released in October with Bioconductor v3.0. You simply:

results(dds, addMLE=TRUE)