What is the substantive question you want to ask?
At first glance you have a 20-diml response on each of 42 patients, and a
manova would be normal (ambiguity intended) way to test a difference
between the two groups.
Is there any reason to expect an additive subject effect (which is what
your lme model implies) and independence of the response times given the
subject (ditto)? Might a log transformation make the first more
plausible?
On Thu, 19 Jun 2003, wayne king wrote:
> Hello list, I have two data sets with 22 subjects of one classification,
> and 20 subjects of another
>
> response times were measured on each of these subjects in response to 20
> items.
>
> Initially I had thought about examining this data in a mixed-effects
> model where the items (20 of them) were the fixed effects and the
> subjects were the random effects, something like
>
> lme(responsetime ~item, data=blah,blah,random=~subject);.
>
> I have two problems, one I have more random effects (42 20 in one group,
> 22 in another) than observations
Don't you have 20 observations per subject, 840 in total, so why is that a
problem?
> and the sheer size of the data set is
> causing a error in the lme routine
> [
> "Problem in .C("mixed_loglik",: Unable to obtain requested dynamic
> memory, while calling subroutine mixed_loglik
> Use traceback() to see the call stack" ]
You only have 840 observations, AFAICS.
> One thought is I could group the 42 subjects into their two diagnostic
> groups which would reduce the random effects in the model, but I'm
> afraid that would gloss over the important issue of variability within
> groups. Any help?
--
Brian D. Ripley, ripley@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
|