R devel - Jun 2015 - Estimating overdispersion when using glm for count and binomial data

Dear All

I recently proposed a simple modification to Wedderburn's 1974 estimate 
of overdispersion for count and binomial data, which is used in glm for 
the quasipoisson and quasibinomial families (see the reference below).

Although my motivation for the modification arose from considering 
sparse data, it will be almost identical to Wedderburn's estimate when 
the data are not sparse. It should therefore work uniformly better than 
the current estimate in glm.

As I am not a regular to R mailing lists, I wasn't sure of the best 
means of proposing this modification. I therefore asked Paul Murrell, 
and he suggested I post it here.

The modification is very simple and would take at most a couple of 
lines of code. The reference below gives details regarding its 
asymptotic properties, as well as simulation results that illustrate the 
benefits of using it for sparse data.

I am happy to give more details if needed.

David Fletcher
Department of Mathematics and Statistics
University of Otago
Dunedin
New Zealand

D.J. Fletcher (2012) Estimating overdispersion when fitting a 
generalized linear model to sparse data. Biometrika 99:230?237 
(http://biomet.oxfordjournals.org/content/99/1/230.abstract?etoc)

David Fletcher <dfletcher <at> maths.otago.ac.nz> writes:
> 
> Dear All
> 
> I recently proposed a simple modification to Wedderburn's 1974 estimate
> of overdispersion for count and binomial data, which is used in glm for 
> the quasipoisson and quasibinomial families (see the reference below).
> 
 [snip]
> 
> The modification is very simple and would take at most a couple of 
> lines of code. The reference below gives details regarding its 
> asymptotic properties, as well as simulation results that illustrate the 
> benefits of using it for sparse data.
> 
  [snip]
> 
> D.J. Fletcher (2012) Estimating overdispersion when fitting a 
> generalized linear model to sparse data. Biometrika 99:230?237 
> (http://biomet.oxfordjournals.org/content/99/1/230.abstract?etoc)
  This looks really useful.  Base R is very conservative; despite the
fact that it would be much more easily adopted in base R, I think it
is much more likely to find a home in an add-on package such as aods3
or glm2 than in base R ...

  cheers
    Ben Bolker