Alternatives for small nb of clusters
Message/Author
 Jessica Malherbe posted on Saturday, June 10, 2017 - 9:53 am
Good afternoon,

First, thank you for this really useful forum.
I am currently trying to find a solution for a paper related to educational research.
We have tested around 600 children in 35 classrooms and created a SEM on the development of language skills.
The SEM includes several latent variables and regression coefficients, and I have not been able to create a full multilevel model (the number of parameters to estimate seems to be too big compared to the number of classrooms).
What kind of alternative would you recommend? I have so far considered the following ones but I am not sure which one is the most reliable.

a) Group-mean centering all the variables, to focus on the individual level (the article does not investigate level 2 variables, but aims to control for them). Reviewers have suggested to take into account the level 2 variance instead of "removing" it.

b) Including classrooms as dummy variables, and using them as covariates on the regression paths. I am not really sure what the different with the first method would be?

c) Extracting the latent variable coefficients and using them in a second step, in multilevel regression analysis with random intercepts and/or slopes.

Type = COMPLEX did not work.

 Bengt O. Muthen posted on Saturday, June 10, 2017 - 12:37 pm
Type=Complex should work. As a second best alternative, I would go with a) which I wrote about in the paper on our website:

Muthén, B. (1994). Multilevel covariance structure analysis. In J. Hox & I. Kreft (eds.), Multilevel Modeling, a special issue of Sociological Methods & Research, 22, 376-398.

I would not choose c) because of biases in factor scores.
 Jessica Malherbe posted on Wednesday, June 14, 2017 - 2:04 am
Thank you very much!

For TYPE = COMPLEX, I have an error message stating that: "This is most likely due to having more parameters than the number of clusters minus the number of strata with more than one cluster"
(I have 55 free parameters but only 35 clusters).

I have read in some books that parameters could be fixed to avoid this, but I feel that there would be too many a priori (20 parameters to fix).
Would you agree to favour group centering in this case?

Best regards