Message/Author 

Baoyue Li posted on Wednesday, June 26, 2013  4:13 am



Is it reasonable to first extract the lowerlevel factor scores from a 2level CFA (unstructured higherlevel covariance matrix, acceptable model fit), and then assign a mixed model to the factor (so the factor is treated as dependent variable and assumed to have a multilevel structure)? An example (not real) is: measuring IQ of students coming from different schools using some questionnaire. Since the interest is in the studentlevel common factor (IQ), we run a 2level CFA with schoollevel having an unstructured covariance matrix. If the model fits well, then we extract the factor score for each student and assume it is the IQ estimate. Next, we want to model IQ with some observed covariates to see if IQ is affected by them. Since again the students come from different schools, we run a 2level regression model. Is that reasonable? Thanks very much! 


I would prefer to do it in a single analysis. 

Baoyue Li posted on Thursday, June 27, 2013  1:13 am



what if the ICC (for the extracted factor scores) is above 0.1 and we are interested in parameters in the structural part, is it better to take into account the 2level structure in order to obtain correct standard error estimates? 


Yes. 

Baoyue Li posted on Monday, July 01, 2013  8:24 am



Dear Dr.Muthen, Thanks for the replies. Maybe I did not make this very clear. Let's say the true IQ for each student "i" in school "j" is T_{ij}, and if there exists a multilevel structure in the real case, T_{ij} could be rewritten as T_{ij}=T_{within}+T_{between}. Does the studentlevel factor score (defined in the first message) reflect the true IQ, i.e. T_{ij}, or the within part of it, i.e. T_{within}? If it reflects the latter, how can we get an estimate of the true IQ T_{ij} ? 


Mplus estimates T_{within} and T_{between} and you have to add them up to get the student's latent score (the student has a contribution from within and from between). 

Baoyue Li posted on Wednesday, July 03, 2013  1:18 am



In some cases, the factor structure may not be the same for each level, e.g. 2 factors at lower level and only 1 factor at higher level, or each level has 2 factors but are loaded on by different items across levels. I think in this case only the T_{within} could be estimated for each of the lowerlevel factors. The T_{between} of these 2 lowerlevel factors are mixed to some extent and may not be possible to split them up. 


Yes, when you don't have equality of factor loading matrices across levels, it doesn't make sense to add up the factors. 

Back to top 