In the past few releases, edgeR's glmLRT has supported an alternative significance test via the test="F" argument. However, I don't really see any documentation of the method used for this test and how it differs from the normal chi-squared likelihood ratio test. Is this alternative test recommended for normal use, and if so, when would you want to use it instead of the chi-squared or quasi-likelihood tests?
The idea with
test="F" was to account for the uncertainty in the estimates of the tagwise NB dispersion for each gene, in order to improve the performance of the GLM framework in the presence of variable dispersions. However, the math didn't pop out nicely because of the difficulties of working with adjusted profile likelihoods during NB dispersion estimation. If I remember correctly, it ended up being conservative during testing.
We think that, if variability and uncertainty in estimation needs to be considered, this would be better achieved with the quasi-likelihood framework. Indeed, I use
glmQLFTest for most of my analyses (RNA-seq, ChIP-seq and Hi-C), only using the chi-squared test in
glmLRT when I don't have any replicates.