[R] Generelized Negative Binomial model in R

[Ben Bolker]

Steve Lianoglou <mailinglist.honeypot <at> gmail.com> writes:

> 
> Hi,
> 
> On Wed, Oct 12, 2011 at 11:23 AM, Akram Khaleghei Ghosheh balagh
> <a.khaleghei <at> gmail.com> wrote:
> > Hello;
> >
> > Does anybody knows that R have a function for Generelized Negative Binomial
> > model, something like "gnbreg" in "STATA" where dispersion parameter itself
> > is a function of covaraites ?
> 
> Take a look at the edgeR (and DESeq) package in bioconductor.
> 
> edgeR uses a GLMs w/ negative binomial to assess differential
> expression of genomic regions using count data (aka next generation
> sequencing data).
> 
> http://www.bioconductor.org/packages/release/bioc/html/edgeR.html
> http://www.bioconductor.org/packages/release/bioc/html/DESeq.html
> 

  You could code it fairly easily in mle2, e.g.

mle2(y~dnbinom(exp(logmu),exp(logk)),
       data=..., start=...,
       parameters=list(logmu~...,logk~...)

where the ... within parameters specify linear models for the log-mean
and log-overdispersion parameters.
  You do have to specify your own starting conditions, and it doesn't
do anything clever in terms of special-purpose optimization -- it just
uses the optimizers built into optim() [with a few other choices, e.g.
those from the optimx package]