Compound symmetry
Message/Author
 jan mod posted on Friday, October 02, 2015 - 11:17 am
Dear Muthen,

How can I say to Mplus to make a compound symmetry model with a dataset in long format?

kind regards
 jan mod posted on Friday, October 02, 2015 - 12:15 pm
Actually what I need it is syntax for the following situation.
I have three waves of data: time, student, school in long format.
I make two dummy variables to model both transitions: 1 to 2 and 2 to 3:
0
1
1
and
0
0
1
.
Normally when I include both dummies my time variance should be zero but it is still positive. I want a zero variance after the inclusion and I need cross level interactions between both time variables and student/school variables
 Bengt O. Muthen posted on Friday, October 02, 2015 - 2:07 pm
You get compound symmetry by using only an intercept growth factor.

In your second post you talk about transitions - do you mean that you want piecewise growth modeling?
 jan mod posted on Friday, October 02, 2015 - 2:13 pm
I read this paper that I want to replicate:
"Fostering reading comprehension in fifth grade by explicit instruction in reading strategies and peer tutoring (British journal of educational psychology (2004), 74, 37-70)

On page 56:
the author had two dummies for time and had these vary in student and class level.

Piecewise growth modeling did not work with my data.
 jan mod posted on Friday, October 02, 2015 - 2:15 pm
EDIT: The author called it a fully multivariate model with regard to the repeated measures.
 Bengt O. Muthen posted on Friday, October 02, 2015 - 11:51 pm
I will take a quick look at that paper on Monday.
 jan mod posted on Saturday, October 03, 2015 - 11:27 am
Thank you very much!!!

I tried to reconstruct the methodology of the author. Long format data, three time points,two dummies (0,1,1 and (0,0,1) for transitions. a45 and a46 are the dummies:

p:56: "(...) we allowed the effects of both dummies to vary randomly across all classes and students (Model 2), yielding what is called a fully multivariate model with regard to the repeated measures. Note that having decided that at the student level the effects of the measurement occasions could be random, it is not possible to estimate a random effect for the constant term at level one, for this coefficient is redundant in a fully multivariate model and is estimated as zero."

usevariables = a41 a45 a46 ;
cluster = a3 a1;
within= a45 a46;
between = ;
Missing are all (999) ;
ANALYSIS:
Type = threelevel random ;
Estimator is ML;
Model:
%WITHIN%
i | a41 on a45;
s | a41 on a46;
a41@0;
%BETWEEN a3%
i with s;
i with a41;
s with a41;
%BETWEEN a1%
i with s;
i with a41;
s with a41;
 jan mod posted on Monday, October 05, 2015 - 1:19 pm
Dear professor,

Did you read the paper? I'm curious how I can implement model 2 of that paper :-).

kind regards
 Bengt O. Muthen posted on Monday, October 05, 2015 - 11:22 pm
It looks like your input above is mimicking their Model 2.

But the article shows a strange way to analyze growth in my view.