The problem that you've stated (which of these models best describes
data) applies to univariate data, for fitting models to a single
counts. For RNA-seq experiments, you are fitting glms to many data
vectors. You will get different answers for different genes, so you
to think a bit differently.
With edgeR, you would start by fitting the full model
design <- model.matrix(~E+A+E:A)
then, after some intermediate commands, use
fit <- glmFit(y, design)
lrt <- glmLRT(fit, coef=4)
to test for which genes the interaction is significant. If the test
significant, then the E*A model is necessary and best. If not, one of
simpler models is satisfactory. You could subset to genes for which
interaction is not significant, then try
design <- model.matrix(~E+A)
and proceed from there.
I interpolate some other answers below.
> Date: Wed, 20 Feb 2013 15:15:20 +0100
> From: Ari Eszter <arieszter at="" gmail.com="">
> To: bioconductor at r-project.org
> Subject: [BioC] edgeR choose the best model of GLM
> Dear edgeR Users,
> I have RNS-Seq count data by genes, 4 treatment groups (each with 2
> biological replicates) and two factors (E: C or H, and A: 13 or 25).
> I applied the GLM model of edgeR with these designs: E, A, E+A and
> I would like to decide which model and which variant describes my
> 1. Am I right that the $coefficients of a "DGELRT" object are log
> likelihood values?
No. They are coefficients from the linear model fitted, as explained
the documentation: help("glmFit").
> 2. Does anyone have a suggestion how can I apply AIC (Akaike
> Criterion) test or Likelihood ratio (LR) test on edgeR GLM results
> choose the best model?
Answered above. We have experimented with AIC, but have not found it
> 3. Is there a possibility to plot the multiple GLM regression for a
> gene? (I would like plot something like this:
I have no idea what this plot represents. In any case, I guess this
for one data vector. For RNA-seq data, you would have to view 20,000
these plots, not a very productive way to go.
> Eszter Ari
> Institut f?r Populationsgenetik
> Vetmeduni Vienna
> Veterin?rplatz 1
> 1210 Wien, Austria
The information in this email is confidential and