Anonymous posted on Thursday, April 14, 2005 - 1:50 pm
Hi, I have done some multilevel latent growth curve models using familydata where spouses are nested within couples (another possibility would be, of course, to use multivariate LGM. However, beacuse MLGM seems to answer my research questions quite nicely, I ended up to use this particular method.).
Now, I have encountered one troublesome problem. There is some missing data, and due to this, in some of the clusters there is only one member (i.e., wife; Quasi-average cluster size is by this reason 1.635 rather than 2). I would like to predict different growth components with gender. If I use gender as a predictor both at the within- and between-levels, the fit of the model is quite good but, at the between-level, there is negative variance estimate for gender suggesting no between-level variation in this variable(?). However, if I use gender only as a within-level predictor, the fit of the model is poor. I am a little bit confused what to do in this kind of situation. Have I missed something important here? Would you have any advice how to proceed here?
Also, my warmest thanks to Bengt for such an inspiring workshop on multilevel modelling in Baltimore last month. That was really something enjoyable. Thank you!
BMuthen posted on Friday, April 15, 2005 - 7:33 am
I think you need to declare gender and a within variable using the WITHIN option. If you don't do that, it will be on both levels and may cause misfit on the between level. If you are already declaring it as a within level variable, I do not know why you would see what you are seeing.
Hi Professor, I'm working with MSEM using Mplus on student clustered between schools. My questions: 1- How can I tell Mplus to use MUML instead of ML because I've unbalanced clusters size? 2- If there are missing data, in order to use MUML, should I impute the data first? 3- Is it appropriate to use ML in unbalanced data and with missing data?
You should use ML which is appropriate for unbalanced and missing data. MUML is ML for balanced data but is not ML for unbalanced data.
Mukadder posted on Tuesday, March 01, 2011 - 12:14 pm
I want to ask two questions about multilevel data with the use of MPlus:
1. I want to estimate a two-level model which includes latent variables with categorical indicators. My MPlus syntax for the conventional SEM which I will then use on the two-level is as follows: VARIABLES ARE u1 u2 u3 u4 u5 u6; CATEGORICAL ARE u1-u6; ANALYSIS: ESTIMATOR IS WLSM; MODEL: f1 BY u1 u2 u3; f2 BY u4 u5 u6; f1 ON f2; Before testing the two-level model for this syntax with categorical indicators should I estimate the within-level model on the pooled-within covariance matrix as it is the third step in Bengt's 1994 article? I am confused in that f1 and f2 are output as continuous latent variables; however they have categorical indicators and that this model may not be tested on a covariance matrix.
2. My sample is an unbalanced case so I think when I step into the two-level model MUML estimation is not appropriate. Because f1 and f2 have categorical indicators. Should I use WLSM estimation or the default of MPlus?
1. The 1994 article deals with continuous outcomes. For categorical outcomes, you should explore the factor structure on between and within using TYPE=TWOLEVEL EFA.
2. You should use the WLSM estimator.
Mukadder posted on Tuesday, March 01, 2011 - 9:11 pm
I have four follow up questions, sorry if they are so simple but I wanted to be sure of some details in order to apropriately use MPlus:
1. Does TWOLEVEL EFA in MPlus allow us to include the within and between level covariates to investigate their effects on latent variables?
2. If it doesn't allow the investigation of the covariates is it plausible to estimate a TWOLEVEL CFA with covariates to have information about the within and between structure?
3. Does TWOLEVEL EFA require a model syntax in the MODEL command?
4. If we are'nt able to test a model (e.g f1 ON f2) in TWOLEVEL EFA, at the last step before we test a TWOLEVEL model is it plausible to estimate a single level model to investigate the relationships among our latent variables? Or after conducting a TWOLEVEL EFA can we proceed directly to a whole TWOLEVEL model skipping the single level analysis?
1. No. 2. Yes. Once you have determined the appropriate CFA model for each level, you can estimate a multilevel CFA model with covariates. 3. No. 4. Once you have determined the multilevel model using TWOLEVEL EFA, you should proceed to a multilevel EFA.
Mukadder posted on Wednesday, March 02, 2011 - 7:36 am
I have some confusions about the terms. 1. Doesn't TWOLEVEL EFA have the same meaning with multilevel EFA? Because performing a TWOLEVEL EFA makes it already a multilevel one?
2. Does "multilevel EFA" mean "multilevel structural equation modeling (MSEM)"? Because I think multilevel EFA would not allow me to investigate the relationships of latent variables among each other simultaneously with the relationships among the latents and covariates but MSEM does?
3. For multilevel EFA should I type in the ANALYSIS command TYPE = TWOLEVEL or is this a specific command for MSEM including two levels (e.g. students within schools)?
4. If I din't get wrong this multilevel issue, in the guidance of MPlus output is it plausible to proceed as 1)TWOLEVEL EFA with latent factors, 2)TWOLEVEL CFA with covariates, 3)TWOLEVEL structural model?