Jay Tanzman <[EMAIL PROTECTED]> wrote in message 
news:<[EMAIL PROTECTED]>...
> We recently submitted a paper to a journal reporting the results of a diet study
> that used a 3-treatment, 3-period crossover design.  Subjects consumed all their
> meals (which were prepared by study personnel) under tightly controlled
> conditions in a metabolic kitchen under the supervision of study personel. 
> Consequently, there was no variation in diet composition among subjects within a
> particular study diet.  The data were analyzed using a mixed linear
> variance-components model with fixed effects for diet and period and a random
> effect for subjects.
> 
> A reviewer has asked us to assess compliance with the diets by comparing the
> fatty acid content of subjects' blood on each diet with the fatty acid contents
> of the diets "at the individual level using 'correlations'."  These variables
> can be assumed to be normally distributed, at least after transformation.  My
> questions are these:
> 
> 1. Is there an appropriate correlation coefficient to do the above, given this
> study design?  
> 
> 2. Would it be valid to compute the proportion of within-subject variation due
> to the diet, like this:
> 
>      Var(Diet) / [Var(Period) + Var(Diet) + Var(Residual)] 

I will probably be less of help than Donald, but
this "proportion of within-subject variation"  sounds somewhat like
the
Eta^2 printout when you do a repeated measures ANOVA.  Eta^2 is a
measure of "association" within-subjects.   Correlation is also
sometimes thought of as "magnitude of effect", a form of effect.


> 
> 3. Assuming #2 is OK, can I "partial out" the period effect by leaving
> Var(Period) out of the denominator?
> 
> Thanks in advance.
> 
> -Jay Tanzman


I am  also interested in this, 
For instance to assess adherence to treatment at the _group_ level,
crosstab procedures have several measures of association, all acting
very similar to measures of correlation since they are between 0 and
1.

For within-subject, Repeated Measures ANOVA, Eta^2 describes the
"association", or % change in serum that is due to the change in
treatment.
Eg., a high Eta would mean that almost all of the change in serum is
related to treatment, suggesting adherence.
You would conclude:  "compliance was good, the serum marker increased
dramatically, (Eta^2=0.90, p<0.001) and then show the within-measures
ANOVA graph that illustrates compliance"

I would also be interested in comments on "reliability" and ICC.  One
site described ICC as a "misnomer" since it isn't really a
correlation:
http://www.acponline.org/journals/ecp/janfeb01/primer.htm


> 2. Would it be valid to compute the proportion of within-subject variation due
> to the diet, like this:
> 
>      Var(Diet) / [Var(Period) + Var(Diet) + Var(Residual)] 
> 
> 3. Assuming #2 is OK, can I "partial out" the period effect by leaving
> Var(Period) out of the denominator?
> 
> Thanks in advance.
> 
> -Jay Tanzman
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to