[R] p values for a GEE model

[Prof Brian Ripley]

On Tue, 11 Apr 2006, Thomas Lumley wrote:

> On Tue, 11 Apr 2006, Renaud Lancelot wrote:
>
>> 2006/4/10, Tarca, Adi <atarca at med.wayne.edu>:
>>> Hi all,
>>>
>>> I have a dataset in which the output Y is observed on two groups of
>>> patients (treatment factor T with 2 levels).
>>>
>>> Every subject in each group is observed three times (not time points but
>>> just technical replication).
>>>
>>> I am interested in estimating the treatment effect and take into account
>>> the fact that I have repeated measurements for every subject.
>>>
> <<Snip>>
>>> I would like to use also as a variant analysis a Generalized Estimation
>>> Equation Model, like
>>>
>>>                 library(gee)
>>>
>>>                 summary(gee(Y~T,id=P,data=data))
>>
>> Beware: the default within-group correlation structure is independence
>> in the gee function (see argument corstr). I think you want an
>> exchangeable correlation structure, i.e. the same within-group
>> correlation for all the measurements:
>
>
> He has a linear model with the same number of observations for each person

Not so: some have 3 and some have 2, and the two levels of T are not quite 
balanced (29/28).

> and no covariates that vary within a person. The independence and
> exchangeable working correlations will give identical answers.

They do not in this case (nor should they, I believe).

-- 
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