[R] lme for repeated measurements over time

[Armin Goralczyk]

Hi list

I am just beginning to understand the complexities of linear mixed
effects models. Maybe someone can give advise concerning the following
problem:

I have two groups of surgical patients in which repeated laboratory
measurements were taken over time after surgery. I decided that lme
would be the best model to fit the data.
I already fitted the model

lme(logratio ~ gr*I(pod-10) + I(pod^2-10) + I(pod^3-10), data=xyz,
random = ~ pod|subj)

where gr = two groups; pod = postoperative day; subj = patient;
logratio = log of value at day pod/preoperative value: log(post/pre)

but these questions remain:

1. Is lme the best model to fit the data? Other suggestions?

2. Since the ratio had no gaussian distribution I took the log which
seems to have a normal distribution. Is this OK?

3. I shifted the intercept to pod 10 because at this point the
difference of the intercept is significant different whereas the
difference at 0 is not significant. Can I do this?

4. Inspection of the data showed that a polynomial regression would be
a better fit for the data. I tried several polynomial regressions up
to pod^5. The above model had the lowest AIC, BIC and logLik. When I
use Anova to compare the models there I get the warning message:
"Fitted objects with different fixed effects. REML comparisons are not
meaningful."
What can I use instead to compare the models?

5. For random I used only pod and not pod^x. Is this correct?

6. Omitting the group factor from pod^2 and pod^3 the model had a
slightly better fit. Can I do this?

7. Can I assume that the data is heteroskedastic? How do I apply the
'weights' in the above model?

I am sorry if some questions may sound weird but I am just beginning
to understand this (for me) rather complex concept. Thanks for any
help.
-- 
Armin Goralczyk, M.D.
Dept. of General Surgery
University of Göttingen
Göttingen, Germany
http://www.chirurgie-goettingen.de