[R-sig-ME] generalized linear mixed models: large differences when using glmmPQL or lmer with laplace approximation

[Ben Bolker]

Martijn Vandegehuchte wrote:
> First of all, thanks a lot for the info.
> 
> I know the differences seem small, but most ecological journals still
> let their opinion about ecological relevance of predictors depend
> completely on p-values... So I think I'll stick to lmer because of the
> Laplace approximation.
> 
> But I don't really know what you mean by: "If you are happy with the df
> given by lme you can use them ... this corresponds to the
> "between-within" option in SAS, Satterthwaite et al. are not available
> in R." I'm familiar to Satterthwaite's correction for the ddfm, I use it
> in SAS proc glimmix, but then I'm stuck with PQL again... But for my
> data the degrees of freedom should be large enough that it doesn't make
> that much of a difference. I just tested the same models in SAS proc
> glimmix, with and without Satterthwaite, and there's no difference. So
> if there is a way of getting the df to obtain a p-value in lmer, I would
> do so.

> 
> Then (maybe stupid) question is: how do I get the df? You mention lme,
> but can I make the same models in lme?
> 
> Thanks again,
> 

  Sorry, I meant glmmPQL -- glmmPQL basically calls lme as a back end,
so "df from lme" is the same as "df from lme".  You can either take
those df (I believe in your case it was 120 (total samples) - 6 (sites)
- 6 (est. fixed parameters) = 108, or run SAS with Satterthwaite and
see what it says df should be.  If you then have a t statistic you
can estimate its (two-tailed) p value with

pt(tstat,df,lower.tail=FALSE)*2

[you can try this on the lme values and see if you get the
right answers]

  Ben Bolker