[R] glmmPQL in R and S-PLUS 6 - differing results

[Prof Brian Ripley]

glmmPQL calls lme, and it is that which differs between S-PLUS and R.

These are optimization problems with multiple local maxima, and like any 
complex statistical fitting problem you should not expect all programs to 
give the same answer.  The short answer is to believe all of them, and to 
check the usefulnees of the answer for your criteria.

On Thu, 19 Aug 2004, Simon wrote:

> Greetings R-ers,
> 
> A colleague and I have been exploring the behaviour of glmmPQL in R 
> and S-PLUS 6 and we appear to get different results using the same 
> code and the same data set, which worries us. I have checked the 
> behaviour in R 1.7.1 (MacOS 9.2) and R. 1.9.0 (Windows 2000) and the 
> results are the same, but differ from S-PLUS 6 with the latest Mass 
> and nlme libraries (Windows XP).

...

> Note that R and S-PLUS differ in the number of iterations. Also the 
> logLikelihoods differ considerably too. Pinheiro and Bates argue that 
> a likelihood-ratio test for fixed effects is not reliable 
> ("anticonservative"), but I think that both packages should at least 
> give the same answer!! I checked lmeControl() in R and S-PLUS and the 
> settings for lme look the same on both platforms. Which output (if 
> any) should we believe? Any insight would be greatly appreciated.

-- 
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