[R] formula for degrees of freedom for nonlinear mixed model in nlme

[David Winsemius]

On Jun 11, 2009, at 8:47 PM, Ben Bolker wrote:

> David Winsemius wrote:
>>
>>
>> The FAQ 7.35 links to this posting:
>>
>> https://stat.ethz.ch/pipermail/r-help/2006-May/094765.html
>>
>>
>
> Actually, this is a different question from the usual "why don't I
> get denominator df?" question -- it is "how are these calculated"
> (since the poster is using nlme, not (n)lmer).  The answer in this
> case is "check Pinheiro and Bates 2000" -- I don't remember the
> page number exactly, looks like it's page 91 -- see Google books:
> <http://tinyurl.com/ntygq3>
>
> If the denominator df don't agree with your intuition, you can
> always recompute p-values with the appropriate den df:
> (two-tailed) 2*pt(abs(t.score),df=dendf,lower.tail=FALSE)

Point well taken. So neither my citation nor page 91 (which also deals  
with LME models) are on point to the OP's question. P & B say that  
inference regarding  covariates in nlme models (which i believe was  
the question posed) are generally Wald statistics, and so don't really  
have denominator degrees of freedom, only numerators DFs.  The  
numerator is then the observation count minus the model degrees of  
freedom. See  365-368 which deals with the sort of nlme model about  
which the OP is asking. A method for generating an anova analysis is  
also demonstrated on page 374.

David Winsemius, MD
Heritage Laboratories
West Hartford, CT