-GSEM- Adding higher-level predictors to a multilevel random intercept model

[Anton Ivanov]

Hello!

I have a question on generalized structural equation modeling. Say, I fit the following two-level random intercept model:

gsem (x1 -> y, ) (x2 -> y, ) (M1[hospital_id] -> y, ), covstruct(_lexogenous, diagonal) vce(robust) latent(M1 ) nocapslatent

where x1/x2 and y are at the doctor level and hospital_id is a latent variable denoting random intercept for upper (hospital) level.

Now, in my data set I also have several hospital level predictors (say, b1 and b2). Is it plausible to add them to the model and estimate like:

gsem (x1 -> y, ) (x2 -> y, ) (M1[hopsital_id] -> y, ) (b1 -> y, ) (b2 -> y, ), covstruct(_lexogenous, diagonal) vce(robust) latent(M1 ) nocapslatent

I recall that in multilevel estimation dependent variable (y) must be at the lowest level (doctor in my case), so I am not sure if I can estimate effects of higher-level b1 and b2 on y.

Thank you in advance for help.

[Anton Ivanov]

Let me provide a bit more information on the issue. My goal is to fit a -gsem- model similar to the following -xtmixed- one:

xtmixed y x1 x2 b1 b2 || hospital_id: x1 x2 b1 b2, vce(robust)

[Joseph Coveney]

Example 38g Random-intercept and random-slope models (multilevel) of the SEM / GSEM volume of the user's manual has some guidance on how to fit random-slopes multilevel models using gsem.

[Anton Ivanov]

Joseph, I am thankful to you for providing such a useful example.