PROC GLIMMIX test for homogeneity with 2 repeated factors

[jsb]

Hi

I am working on a dataset where the interest is not in testing for the adjusted mean differences solely but on testing whether the variability differs between two independent groups.  The measurements are nested in a sense in a combination of distance and location within a distance of a circle.  Assuming the center of the circle is a reference point (of the eye) there are 3 levels of eccentricity (distance) and locations on that circle (location).  I need to correctly specify this model, considering this structure to accurately have the proper degrees of freedom and fixed effects.  Specifically, I have the following questions that I hope some could help me out with:

1. I generally use PROC MIXED but in a search focusing on variability and not means, I found that GLIMMIX is best to use as a test for homogeneity that should provide this result.  Assuming the model is specified correctly, would the indpendent factor being specified as a fixed factor in addition to group variance in a random statement be appropriate?

2. I have currently specified 3 random statements, one for ID and two each for the nested (locationxdistance combo within id scenario? This seems over specified.

Thank you for your help.

[Ksharp]

Not sure I undestand your question.

You could use GROUP= option to assume obs from the same group have the same variance .

random int/subject=school group=class;

2)You also could try COVTEST statement to test if the covariance is different between two group:

[jsb]

Thank you for your responses.  I'dl ike to specify that the residuals could differ across group.  I am trying to ask the question, does the variance differ across groups (independent factor) while maintaining the structure/design features of the datset.  So, I think I would like to determine whether the residual variance differs across group while maintaining this two-way design.  My mixed model could look something like this but I am unsure how to specify this in glimmix, given that GLIMMIX does not allow repeated statement. I understand how to do this with one repeated factor but not 2.  Is this more clear? group=independent group; rf1=first within-subject condition; rf2=second within-subject factor such that rf1xrf2 create a unique combination. I first need to specify the residual structure in glimmix correctly and then will use the covtest.

proc mixed data=data covtest;
class id group rf1 rf2;
model y=group;
random subj rf1(id) rf2(id)/ v vcorr group=group;
run;

[Ksharp]

"I understand how to do this with one repeated factor but not 2. "

If you want to specify two factors in REPEATED ,you need to specify a special covariance construct:

For GLIMMIX, could like:

random Var Year/type=un@ar(1) subject=Child residual;

And @SteveDenham  might know your question better and could give you a hand.

[SteveDenham]

Here are links to couple threads in the community from some years ago:

https://communities.sas.com/t5/SAS-Procedures/GLIMMIX-with-two-repeated-measures-variables/td-p/2326...

https://communities.sas.com/t5/Statistical-Procedures/repeated-measures-in-glimmix/td-p/128310

The first from @sld has several ways of modeling the doubly repeated measures and compares V matrices from MIXED and GLIMMIX.

The second is from me, and has a lot of untested code in the GLIMMIX call, but basically boils down to calling one effect as strictly G side, and the effect nested within that variable (say time) as an R-side (residual). Take care in identifying the subject for the R-side effect.

SteveDenham

[jsb]

I beleive I was emailing you as you were posting. I am new to posting on this site. I will look through these and get back with any questions. Thank you.

[jsb]

Hi Steve @sld ,

Thank you for the code provided.  I have a few questions for you if you could help. 

1. In your code provided, I understand year and month are within-subject factors but was treatment?  

2. The whole purpose to evaluate the variance (heterogeneity) of group.  In the GLIMMIX example below, I am able to estimate a variance matric for each rf1, but am not sure how to add a second group.  How would you recommend manipulating the code?  

3. My factors are imbalanced.  Factor levels of rf2 vary by rf1.  I do not necessarily need a 20x20 (point by point) but would like an estimated variance of rf1 and rf2 (potentially rf1xrf2) and group.  Neither of the models below get me that. Would a random statement for each be better?

4. I am familiar with post plots of lsmeans but do you know of a good post plot to display the variances aside from a table?

Thank you.

proc mixed data=data covtest;
class rf1 rf2 subj group;
model y = group|rf1|rf2 / ddfm=kr;
repeated rf1 rf2/subject=subj(group) type=un@cs r rcorr group=group;
run;

proc glimmix data=data;
    nloptions gconv=0;
    class rf1 rf2 subj group;
    model y = group|rf1|rf2 / ddfm=kr;
    random rf2 / subject=subj(group) group=rf1 type=ar(1) v residual;
    run;

proc mixed data=y covtest;
    class rf1 rf2 subj group;
    model y = group|rf1|rf2 / ddfm=kr;
random intercept/subject=subj group=group v vcorr;
repeated rf2/subject=subj*rf1 type=ar(1) group=group;
run;

[Ksharp]

https://support.sas.com/kb/40/724.html
https://support.sas.com/kb/37/107.html