[R] glm with a distribution free family

[Prof Brian Ripley]

On Sat, 4 Jun 2005, Laetitia Mestdagh wrote:

> I am trying to fit a glm with a distribution free family, link = log and 
> variance = constant*mu. I guess I have to use the quasi family but the 
> choices of variance are restricted to constant or mu or mu^2..., I don't 
> know the way to choose the variance that I need, i.e. constant*mu.

The constant in the variance specification is the dispersion parameter in 
a glm, so quasi() does cover this.  Setting variance = "constant" actually 
gives you 1 (look at the code).

If you know the constant you will need to specify it when calling e.g. 
summary() or predict(): otherwise a moment estimator will be used.

-- 
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