

Combination of MFA and multilevel pat... 

Message/Author 

Marco posted on Tuesday, November 29, 2005  2:22 pm



Hello, I would like to analyse clustered data with variables that have variance on the individual and group level. Unfortunately, the number of items is twice as much as the number of clusters, so a multilevel covariance structure analysis as described in the 1994paper of B.Muthén is not possible. Therefore, I'am planning a multilevel path analysis instead of the multivariate approach. My question is, whether there is a strategy in using separate MFA's (= using not the whole covmatrix, just a part of it) as a preliminary analysis? For instance, would it be reasonable to conduct separate MFA on every construct in order to evaluate the invariance of loadings over the two levels? In this way it would be impossible to determine, whether two distinct constructs on the individual level might merge into a general factor on the between level. But at least it would provide a rationale for aggregating the items to a score and using the intercepts as a 2levelpredictor in a multilevel setting. Many thanks for any suggestions! 

Marco posted on Tuesday, November 29, 2005  2:36 pm



I'm sorry, I forgot a related question: Would it be reasonable to save factor scores as an outcome of each separate MFA and use them as the variables in a path analysis? Does Mplus allow to save factor scores for latent variables on the between level? Thanks again. 

Bmuthen posted on Wednesday, November 30, 2005  9:40 am



I think your strategy of doing the prelim investigation construct by construct is a good one. Mplus can give you 2level factor scores (within and between). Using estimated factor scores as observed variables in path analyses is fraught with the same potential biasing issues as regular factor scores, but this may still be the way to go when you have so many items  and having many items make these factor score biases be small. 

Marco posted on Thursday, December 01, 2005  3:08 am



Many thanks. There´s just a technical question left: Mplus saves not just the factor scores for the (in the modelcommand) specified factors of the MFA. There are also betweenlevel factor scores named BFACT1, BFACT2, etc. Are these the estimated groupspecific intercepts of the betweenlevelitems? 


Yes. 

Back to top 

