[R] negative binomial overdispersion question

[Prof Brian Ripley]

On Tue, 22 Nov 2005 kartik at pdx.edu wrote:

> Hello,
>
> I'm a grad student in the Intelligent Transportation Systems lab at Portland
> State Univ. in Portland, OR, USA.  I'm trying to learn the basics of R to run a
> negative binomial in the near future, and so I ran a test regression on roadway
> crash data obtained from "Statistical and Econometric Methods for
> Transportation Data Analysis" by Washington et al (p. 250).  I ran the test
> (glm.nb from library MASS) and got the same output as in the text for all the
> parameters except the overdispersion parameter; the text lists 0.516, but R
> gave me an overdispersion parameter output of 1.9365.  (I've attached the
> raw dataset, in Excel .csv format, if anyone wants to try the test.  If the
> attachment does not go through, please email me and I'll forward it on to you.)
>
> Has anyone had similar problems with the overdispersion parameter output from R?

MASS documents exactly what it does on p.206, but in no parametrization 
does the negative binomial have an `overdispersion' parameter in the sense 
used for glms (e.g. in McCullagh and Nelder).  glm.nb certainly does not 
report an `overdispersion parameter output', so you are not reading it 
carefully enough.

Did you consult the book whose support software you are using?  That is 
where the detailed documentation is.  My guess is that your reference has 
a parameter 1/theta, as 1/1.9365 is very close to 0.516.

-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595