I am testing a multilevel structural equation model with a 3-wave longitudinal dataset. I have nested all three measurements within individuals and all the paths I am interested in include within-level variables. I would now like to perform group analysis to my data. So I would like to see if the coefficients are different between group 1 and group 2 of the sample. The problem is that the grouping variable is also a within-level variable and is changing over time. So its value can be either 1 or 2 across times 1-3. How can I do that with Mplus?
The reason why I need to do that is because I am using an incogruence score between A and B (both within-level variables) as a predictor in my model. I am more interested in the incogruence than the interaction between the two variables. But in order to address criticisms around the use of incogruence scores, I would need to perform my analyses for two groups: Group 1: A > B and Group 2: A < B.
If I use the grouping variable as a dummy time-varying covariate, I will control for its effect, but I will still not be able to compare the results between group 1 and group 2.
The only way to do that is to create an interaction between my incogruence score and the covariate? But this is too complicated and I am not sure what that variable would mean. Then it would be better to use an interaction term all the way from the start, right?
Kerry Lee posted on Tuesday, September 30, 2014 - 5:53 am
Dear Profs Muthen,
I am running a cohort-sequential design with four overlapping cohorts, each tested in four consecutive years. To answer different questions regarding the dependence of academic performance on executive functioning, I modelled the multivariate data using both autoregressive and latent growth models.
In each cohort, children came from different schools. Over the four years, children changed schools and the number of schools involved in the study increased dramatically (from 5 to over 40). The number of children in each school also varied from 1 to over 40. I ran a ICC analysis and about a third of the cohorts have design effects > 2 on some of the variables. I am not particularly interested in the effects of school on the substantive relations. However, given the design effects, I do need to control for clustering. With the time-varying nature of the clusters, I do not think I can just use the CLUSTER command to specify different schools at which children are enrolled in different years.
Kerry Lee posted on Tuesday, September 30, 2014 - 5:54 am
I read your recommendation above regarding using clusters as a time-varying dummy variables. I tried it and specified for the two variables of interest (one manifest and the other latent) to be regressed to the dummy (declared as nominal). The substantive relations are modelled in a cross-lagged autoregressive multigroup/longitudinal model. I got the following error:
*** ERROR in ANALYSIS command. ALGORITHM=INTEGRATION is not available for multiple group analysis. Try using the KNOWNCLASS option for TYPE=MIXTURE.
I tried to solve this by specifying the different cohorts as KNOWNCLASS and reran the model. It terminated with an error because the school variable has more than 10 categories.
Is there an alternative way to model the data? I Would very much appreciate your advice.
There are models for changing cluster membership over time but they are much more advanced and only limited versions are available in Mplus. Perhaps you could instead just use the first school attended as the cluster unit, assuming this is the most influential school membership. You can do this using Type=Complex.
Also, please note that we ask posts to be limited to just one window.
Kerry Lee posted on Wednesday, October 01, 2014 - 12:11 am
Dear Prof Muthen,
Apologies for the split post. Would you have a reference for the advanced models that you mentioned and the versions that are available in Mplus?
Kerry Lee posted on Friday, October 03, 2014 - 3:21 am
Dear Profs Muthen,
I tried using just the first school attended as the cluster unit. The model terminated normally, but with the following message:
STANDARD ERRORS OF THE MODEL PARAMETER ESTIMATES MAY NOT BE TRUSTWORTHY FOR SOME PARAMETERS DUE TO A NON-POSITIVE DEFINITE FIRST-ORDER DERIVATIVE PRODUCT MATRIX. THIS MAY BE DUE TO THE STARTING VALUES BUT MAY ALSO BE AN INDICATION OF MODEL NONIDENTIFICATION. THE CONDITION NUMBER IS -0.146D-15. PROBLEM INVOLVING THE FOLLOWING PARAMETER:Parameter 5, Group K2: [ NORW2 ]
THIS IS MOST LIKELY DUE TO HAVING MORE PARAMETERS THAN THE NUMBER OF CLUSTERS MINUS THE NUMBER OF STRATA WITH MORE THAN ONE CLUSTER.
The same model without the cluster information terminated without error. Parameter 5 is in the alpha matrix. Because I have missing data, mean-structure has to be included. I do not know which parameters the second paragraph refer to. I only have five clusters. Would you have suggestions on how to solve or trouble-shoot this problem?
The message is due to having only 5 clusters. Multilevel modeling cannot be done well unless you have at least 20 clusters, and preferably at least 30-50. If you have only 5, you should instead create 4 dummy variables to represent the differences among the 5.