deseq2_confects support for likelihood ratio test output?

[sharbhaw]

We are wondering if there is a possibility of adding likelihood ratio test support in your package because we are trying to run a more complicated model than the Wald Test. We tried to use the deseq_2confects function and got an error with the likelihood test setup.

[pfh]

Hi sharbhaw,

This is something I definitely want to provide, but haven't worked out how to yet. The confects method currently only works when testing a single coefficient or contrast. For more complicated tests such as ANOVA style tests, there is a question of what is the actual effect size (there are more options here than when only defining a null hypothesis). I'm also still trying to puzzle out how a test-relative-to-a-threshold would work in more than one dimension.

[sharbhaw]

I wonder if it would be possible to support LRT results in some simple use cases
that have a binary comparison / effect size? For example
we have used DESeq2::nbinomLRT with full = ~ batch + genotype, reduced = ~ batch,
where our effect size of interest is genotype-wt vs genotype-mutant. The other
model we have used is full = ~ genotype, reduced = ~ 1 (we have some experience that
the LRT is more reliable than the Wald test).

[pfh]

In this case you could try edger_confects, which uses the edgeR quasi-likelihood version of TREAT.

(This is quasi-likelihood rather than likelihood, which I think is better as it takes into account uncertainty in the dispersion. I could fairly easily support the likelihood version of TREAT from edgeR, but I think it's too liberal on account of not taking into account uncertainty in the dispersion. Willing to listen to arguments on this though.)