R help - May 2008 - Validating a mixed-effects model

Hi

I constructed a mixed-effects model from longitudinal repeated
measurements of lab values in 22 patients seperated into two groups
with the groups as fixed effect using lme. I thought about using the
jackknife procedure, i. e., removing any one subject and calculating
the fixed effect, to assess the stability of the fixed effect and
thereby validate the model. I suppose this has been done in the
following study:

http://content.nejm.org/cgi/content/full/357/19/1903
(this may be restricted access, sorry)

Is such an approach feasible?

Also in the article results are confirmed by comparing the mixed model
with a fitted least-squares regression. I understand that this can be
achieved with lmlist, but only for for models without an additional
fixed effect!?

Are there any other good approaches to validate a mixed-effects model
that will be accepted in medical peer review?

-- 
Armin Goralczyk, M.D.
--
Universit?tsmedizin G?ttingen
Abteilung Allgemein- und Viszeralchirurgie
Rudolf-Koch-Str. 40
39099 G?ttingen
--
Dept. of General Surgery
University of G?ttingen
G?ttingen, Germany
--
http://www.gwdg.de/~agoralc

Hi Armin,

Alternatively you could use premutation statistics. You could shuffle 
your subjects between groups randomly under the Null hypothesis of no 
differences between groups and each time claculating the lme model. I am 
not sure, but if you do it at each time point of your repetition at each 
draw, then you could remember the greates F value for your F-value 
distribution. This could control the multiple comparison problem. Then 
after let's say 1000 draws you have a F value distribution under the 
Null hypothesis and you could determine your critical F value from that 
distribution.

Hope that helps,

Stephan


Armin wrote:

Hi

I constructed a mixed-effects model from longitudinal repeated
measurements of lab values in 22 patients seperated into two groups
with the groups as fixed effect using lme. I thought about using the
jackknife procedure, i. e., removing any one subject and calculating
the fixed effect, to assess the stability of the fixed effect and
thereby validate the model. I suppose this has been done in the
following study:

http://content.nejm.org/cgi/content/full/357/19/1903
(this may be restricted access, sorry)

Is such an approach feasible?

Also in the article results are confirmed by comparing the mixed model
with a fitted least-squares regression. I understand that this can be
achieved with lmlist, but only for for models without an additional
fixed effect!?

Are there any other good approaches to validate a mixed-effects model
that will be accepted in medical peer review?


-- 
*Stephan Moratti, PhD/
/**/see also: http://web.mac.com/smoratti/
/*Centro de Tecnolog?a Biom?dica CBT,
Universidad Polit?cnica de Madrid,

en la actualidad (currently at) en el
Centro de Magnetoencefalograf?a Dr. Perez Modrego,
Universidad Complutense de Madrid,
Faculdad de Medicina,
Pabell?n 8,
Avda. Complutense, s/n,
28040 Madrid,
Spain,

email: moratti en gbt.tfo.upm.es <mailto:moratti en gbt.tfo.upm.es>
         moratti en med.ucm.es
Tel.:    +34 91 394 2292
Fax.:   +34 91 394 2294
*/
/*