[R] glmmPQL in 2.3.1

[Prof Brian Ripley]

On Mon, 25 Sep 2006, Justin Rhodes wrote:

> Dear R-help,
>
> I recently tried implementing glmmPQL in 2.3.1,

I thought *I* had implemented it: are you talking about my function in 
package MASS or your own implementation?

> and I discovered a few differences as compared to 2.2.1.

You appear to be talking about contributed packages (MASS, and glmmPQL 
also depends on nlme) without giving their version numbers.

> I am fitting a regression with fixed and random effects with Gamma error 
> structure.  First, 2.3.1 gives different estimates than 2.2.1, and 
> 2.3.1, takes more iterations to converge.

We have no idea, given the lack of reproducible example.  glmmPQL does 
give the same answers as before for the book examples for which it is 
support software.  This may well be due to an underlying change in nlme.

> Second, when I try using the anova function it says, "'anova' is not 
> available for PQL fits", why?  Any help would be greatly appreciated.

Because anova implies you are using an optimization criterion, such as 
least squares or maximum likelihood, and so there is something like a 
deviance to partition.  It was not used in the book with glmmPQL supports, 
but it seems some people were using glmmPQL without reference to that book 
so I made a number of their misuses explicit errors.  This *is* in the 
NEWS and WHATS.NEWS files for MASS and VR:

- There are anova() and logLik() methods for class "glmmPQL" to stop
   misuse.

-- 
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