[BioC] edgeR Quasi-likelihood with tagwise dispersion?

[Gordon K Smyth]

Dear Ryan,

Yes, glmQLFTest() can be substituted for glmLRT() in an edgeR analysis 
pipeline, but it assumes that trended not tagwise dispersion has been 
estimated.  This is explained in the published paper Lund et al (2012).

There has been no detailed description of the usage of glmQLFTest() in the 
edgeR document so far, but there will be before the next release in a few 
weeks' time.

Yes, glmQLFtest() does its own tagwise dispersion moderation.  This is the 
reason for the function.  glmQLFTest() actually does its own trending as 
well, so that it is quite robust to the NB dispersions that it gets as 
input.  The help page says:

"[glmQLFTest] calls the limma function squeezeVar to conduct empirical 
Bayes smoothing of the genewise multiplicative dispersions."

I don't think that you will get meaningfull results from inputing tagwise 
dispersions into glmQLTest(), i.e., it doesn't make sense to re-shrink 
values that are already shrunk.

Because of the possible misunderstanding, glmQLFTest() now has a new 
calling sequence, so that it can stand alone from glmFit().  This permits 
the function to ensure that it gets the correct NB dispersion estimates.

Best wishes
Gordon

---------------------------------------------
Professor Gordon K Smyth,
Bioinformatics Division,
Walter and Eliza Hall Institute of Medical Research,
1G Royal Parade, Parkville, Vic 3052, Australia.
http://www.statsci.org/smyth

On Mon, 18 Mar 2013, Ryan C. Thompson wrote:

> Hi Gordon,
>
> I was looking in the devel code for edgeR and I noticed this new check 
> in glmQLFTest:
>
>> if(!is.null(disptype)) if(disptype=="tagwise") stop("glmfit should be
> computed using trended dispersions, not tagwise")
>
> which refuses to run the quasi-likelihood F-test on a fit for tagwise 
> dispersion. Could you elaborate on this? Does glmQLFTest implement its 
> own different kind of tagwise behavior that is different from that of 
> estimateGLMTagwiseDisp?
>
> What would happen if I ran glmQLFTest after estimateGLMTagwiseDisp 
> (using the release version, which does not disallow it)? Would the 
> results be statistically meaningful? I have been doing this sometimes 
> because I assumed that glmQLFTest was supposed to be a "drop-in" 
> replacement for glmLRT, using all the same dispersions and such, due to 
> this quote from the glmQLFTest help text in the current release of 
> edgeR:
>
>> [glmQLFTest] behaves the same as ‘glmLRT’ except that it replaces
> likelihood ratio tests with quasi-likelihood F-tests for coefficients in 
> the linear model.
>
> So, what is the correct usage of glmQLFTest with the estimate*Disp 
> family, and has this changed between the current release and devel 
> versions of edgeR?
>
> Thanks,
> -Ryan Thompson
>

______________________________________________________________________
The information in this email is confidential and intend...{{dropped:5}}