R help - Jun 2004 - GlmmPQL

Dear all,


I have two questions concerning model simplification in GlmmPQL, for for random
and fixed effects:

1. Fixed effects: I don't know if I can simply specify anova(model) and
trust
the table that comes up with the p value for each variable  in the fixed
effects formula. I have read that the only way to test for fixed effects is to
do approximate wald tests based on the standard errors of the models where I am
subsequently withdrawing one variable from the fixed effect formula at a time.
What does "aproximate" wald test mean? What is the best option?

2. Random effects: If AIC is not meaningful in GlmmPQL, how do I test for the
significance of the random effects?

3. I way to see if 1 single level of random effects is helpful in terms of
analysing the data, would be to comapre the GlmmPQL model with a glm models
without random effects, but again: what do I compare if AIC is not meaninful?
and if there is something I can compare, could I test for the significance of
that difference?

Could someone bring light to this?

Thanks,

Olga Tosas

I think you need to read the references given in ?glmmPQL and its
reference to understand what PQL actually does.  If you don't know the
theory behind a statistical method, you should try to understand it before
trying to use it.

The same applies to AIC.  Even if AIC were computable, do you know its 
theoretical limitations (and it does not apply to testing if random 
effects have zero variance, for example)?

On Wed, 9 Jun 2004, O Tosas Auguet wrote:
> 
> Dear all,
> 
> 
> I have two questions concerning model simplification in GlmmPQL, for for
> random and fixed effects:
> 
> 1. Fixed effects: I don't know if I can simply specify anova(model) and
> trust the table that comes up with the p value for each variable in the
> fixed effects formula. I have read that the only way to test for fixed
> effects is to do approximate wald tests based on the standard errors of
> the models where I am subsequently withdrawing one variable from the
> fixed effect formula at a time. What does "aproximate" wald test
mean?
> What is the best option?
> 
> 2. Random effects: If AIC is not meaningful in GlmmPQL, how do I test
> for the significance of the random effects?
> 
> 3. I way to see if 1 single level of random effects is helpful in terms
> of analysing the data, would be to comapre the GlmmPQL model with a glm
> models without random effects, but again: what do I compare if AIC is
> not meaninful? and if there is something I can compare, could I test for
> the significance of that difference?
> 
> Could someone bring light to this?
The references will.

-- 
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