icc after gsem

[Guest]

Hello,
I am trying to fit a mediated gsem multilevel model, in 2 levels. I would like to obtain the icc as a postestimation test. I am aware that the 'estat icc' function does not work following gsem. Therefore, how can I get the icc?

Here is my code:
. gsem (M1[Retailer2] -> USJGr, family(ordinal) link(probit)) (M1[Retailer2] -> VHGroup, family(ordinal) link(probit)) (M1[Retai
> ler2] -> BasktSR, family(gaussian) link(identity)) (ss -> USJGr, family(ordinal) link(probit)) (ss -> VHGroup, family(ordinal)
> link(probit)) (ss -> BasktSR, family(gaussian) link(identity)), covstruct(_lexogenous, diagonal) group(Retailer2) iterate(40)
> latent(M1 ss ) nocapslatent

Thanks

[Joseph Coveney]

What's your definition of intraclass correlation in such a model? Does ss fit into it? If so, then how?

[Guest]

Joseph, firstly thank you for your response.

SS is a latent variable, from which there are two observed variables. Retailer2 is a variable which represents 17 retailers at the top level (M1). n=344 customers, nested within these 17 retailers. I would like to understand the icc = 'proportion of the variability in outcomes that is attributable to the retailers'. It is quite useful to have the icc, in order to compare models.

[Joseph Coveney]

It is quite useful to have the icc, in order to compare models.

You'll need to explain that one to me. I've never estimated intraclass correlation in order to compare models, and I don't see how ICC estimates can be used to compare (i.e., choose between candidate) models.

When I asked about your definition of ICC for this generalized SEM, I was asking how it is implemented in terms of a formula. Your definition ("proportion of the variability in outcomes that is attributable to the retailers") doesn't tell me what to do with the variances that are explicitly estimated or assumed for each outcome variable in the model, including what to do with the latent factor's variance.

[Martijn Hogerbrugge]

Does anyone else know how to get the intraclass correlation for categorical outcome variables? I was able to run a multilevel multinomial regression model using the gsem command, but as the original author of this thread posted, estat icc does not work after gsem. Can I manually calculate the ICC like for multilevel logistic models, assuming the variance at the individual level is equal to pi² / 3?