

Multilevel SEM with WLSMV estimation 

Message/Author 


Hi there, I'm running a multilevel SEM model with 10 categorical indicators of 2 latent variables (on both within and between levels), and am using WLSMV estimation. I receive a "model estimation terminated normally" note and see that the parameters seem in order (except for a nonsignificant betweenlevel regression...), but I noticed that during the bivariate estimation not every estimation 'concludes' with a Parameters and Derivatives summary. Is this an issue? Thanks in advance! 


I don't think this is a problem. If you want us to look at it further, please send the output and your license number to support@statmodel.com. 

Djangou C posted on Thursday, October 06, 2016  3:43 am



Dear Dr Muthén, I have a few questions. 1) From my understanding, MUML is a limited information estimation method that gives fast computations than the FIML estimation as it does not require numerical integration when FIML does. MUML and FIML are equivalent with balanced designs but different with unbalanced designs. In this last situation FIML can be more accurate. In MUML, like in FIML, the SE are computed using asymptotic approximation (Fisher information matrix). Is that correct? 2) Diagonally weighted least square (WLSM, Asparouhov & Muthén, 2007) is also a limited information estimation method which uses univariate FIML procedure to estimate the means of the observations and the diagonal elements of the variance covariance matrix of the observations. The offdiagonal elements are estimated using bivariate FIML procedure. These estimates are then used to obtain the asymptotic covariance matrix employed for weighting in the WLS estimation. The SE are also computed using asymptotic approximation. Is that correct? 3) In the context of LGC with an intervention. Can I compute the effect size by taking the difference between the slope mean of the intervention and the control divided by the SE of the slope? Thank you for your help. 


1) yes. 2) yes. 3) that's just one approach and perhaps one that isn't easy to understand. perhaps instead focus on mean differences at the end. 

Back to top 

