3-level CFA -- within-variance, betwe...
Message/Author
 lamjas posted on Friday, August 03, 2007 - 8:13 pm
Hi,

I am doing a 3-level CFA for school data (students nested within classrooms; classrooms nested within schools). So far, I don't see any articles doing such kind of analysis (if you have seen such articles, please let me know some references). Therefore, I am planning to follow those articles doing 2-level CFA. In the second step, it needs to calculate the within- and between-level covariance and obtain ICC values. The output I got is one within-level covariance and one between-level covariance. Is it right?

My major syntax is the following:

usevariables classid schid Q1-Q21;
cluster is schid classid;
analysis: type = meanstructure complex twolevel basic;
output: sampstat;

Thanks in advanced for you help.
 Bengt O. Muthen posted on Friday, August 03, 2007 - 8:38 pm
That's right. The icc's should be part of the output, or you can compute them yourself from the variances of those 2 matrices,

icc = between/(between + within).
 lamjas posted on Saturday, August 04, 2007 - 2:37 am
Then I don't understand. When I look at the articles doing 2-level CFA, they got the total, within-, and between- covariance/correlation matrices and calculate the ICC. But, I am doing a 3-level CFA, should I get two within- covariance/correlation matrices to identify the 3 levels?
 Linda K. Muthen posted on Tuesday, August 14, 2007 - 6:43 pm
The ICC for three-levels is the level 3 variance divided by the sum of the level 1, 2, and 3 variances.
 Harald Gerber posted on Thursday, September 24, 2009 - 11:51 am
I have the same problem. Classes are nested within schools.

my syntax:

cluster = school_id class_id;

Analysis: Type = basic complex twolevel;

But I get only one between covariance matrix and the variance is the same as in the covariance matrix of analyses with "class_id" specified as a single cluster (omitting school cluster).
I expected something like two "between covariance matrices", one for school and one for class.
Btw., is ICC only for manifest outcomes or exists there also an ICC for latent variables, like growth factors?
 Harald Gerber posted on Thursday, September 24, 2009 - 12:03 pm
Forgot to mention: The motivation for my first question above is that I'm interested in determining the additional amount of clustering of data within classes beyond what is caused by schools.
 Linda K. Muthen posted on Thursday, September 24, 2009 - 1:37 pm
The school clustering associated with COMPLEX is not modeled. The standard errors are adjusted for clustering due to school.

To compute an ICC for a latent variable, divide the between variance by the sum of the within and the between variance.