QLF or LRT in edgeR for paired sample designs?
1
0
Entering edit mode
@akira-imamoto-11332
Last seen 5.9 years ago
USA/Chicago/University of Chicago

This may be a naive question, but I am wondering which, glmFit/glmLRT or glmQLFit/glmQLFTest, I should use for DE gene analysis in paired designs using edgeR. In the edgeR Users Guide (rev 20.4.2016),  glmFit/glmLRT is used in Chapter 3.4.1 and 4.1 for paired designs, whereas glmQLFit/glmQLFTest is recommended over glmFit/glmLRT in general in Chapter 2.10.3. I checked a chapter published in Methods Mol Biol (vol 1418, pp391-416, 2016) about Quasi-likelihood methods in edgeR, but there was no mention paired designs in the paper. Is glmFit/glmLRT still preferred for paired designs?

edger paired samples glm QL F-test LR-test • 3.9k views
3
Entering edit mode
Aaron Lun ★ 27k
@alun
Last seen 3 hours ago
The city by the bay

Check out some previous posts on this topic:

EdgeR glmLRT vs glmQLFTest

EDGE-R exact test vs QL F-test

The experimental design doesn't determine the choice of glmLRT vs glmQLFTest. Rather, it's got to do with type I error control, for which the QL methods are better; or cases involving low counts and/or large dispersions, where the LRT is better. I use the QL methods for my routine analyses, but both of them should do the job well. The case studies were originally written using glmLRT and we've just left them as they are (I mean, it's pretty easy to swap out glmFit and glmLRT for the corresponding QL-based functions).

0
Entering edit mode

Thank you very much for the clarification as well as the links for the previous posts. Yes, it is easy to swap them. When I compared them yesterday, I found that glmQLFTest is more conservative than glmLRT in general for our RNA-Seq experiments.

My question came up as I found one of the three sets (each in the same paired design) had a lot fewer genes with glmQLFTest (approximately 200 vs 1500). In contrast, at FDR<0.05, the difference is much smaller between glmQLFTest and glmLRT (approximately 1000 vs 2000).  The other two sets had much smaller differences between glmQLFTest and glmLRT, probably due to smaller dispersion.