Skip to content

Mixed model with multiple response variables?

5 messages · Gang Chen, David Afshartous, David Duffy +2 more

#
Hi,

I have a data set collected from 10 measurements (response variables)
on two groups (healthy and patient) of subjects performing 4 different
tasks. In other words there are two fixed factors (group and task),
and 10 response variables. I could analyze the data with aov() or
lme() in package nlme for each response variable separately, but since
most likely there are correlations among the 10 response variables,
would it be more meaningful to run a MANOVA? However manova() in R
seems not to allow an error term in the formula. What else can I try
for this kind of multivariate mixed model?

Also, if I want to find out which response variables (among the 10
measurements) are statistically significant in terms of acting as
indicators for group difference, what kind of statistical analysis
would help me sort them out?

Thanks in advance,
Gang
#
Note sure how to fit the model in R, but two good references are:

Fieuws & Verbeke (2006). "Pairwise fitting of mixed models for the joint
modeling of multivariate longitudinal profiles," Biometrics, 62, 424-431.

Fieuws et al. (2008). "Predicting renal graft failure using multivariate
longitudinal profiles," Biostatistics, 9, 419-431.
On 8/5/08 3:54 PM, "Gang Chen" <gangchen6 at gmail.com> wrote:

            
#
On Tue, 5 Aug 2008, Gang Chen wrote:

            
You might look at the Oct 2007 R-News article on the subject.  But a 
flexible approach is to use the sem package.

David Duffy.
#
Hi Gang,

I suggest that you ask yourself whether or not the correlation between
the response variables is of inferential interest in the subject
matter.

If not, then analyze them separately, correct for multiple tests
somehow, and check the correlation of the residuals.  If the residuals
are correlated then a more efficient estimate would be possible using
e.g. a relation to seemingly unrelated regression.  If the residuals
are uncorrelated then I think that you can keep the separate analyses.

If you want to try to model the correlations between the response
variables in an otherwise mixed-effects framework, some nice work was
done by Daniel Hall (U of Georgia) on forestry data, published in
Biometrics, if I recall correctly.  I also tried out some ideas in a
2004 article published in the Canadian Journal of Forest Research. 

Cheers,

Andrew

  
    
#
On 06/08/2008, at 5:54 AM, Gang Chen wrote:

            
This looks like a multilevel model, with your measurements nested  
within subject. The difference to typical models is that each outcome  
will need a different variance, which I think is possible in LME.

A GEE might work (as an alternative to multilevel GEE which should)  
using the robust SE to cope with the model misspecification.

Ken