Can I conduct a second order LCA on two sets of related but not identical sets behaviors?
For example, risk behaviors and academic behaviors. I would expect them to be related to one another but not the same underlying construct.
Membership in each of the 2 first-order classes would be based on three binary indicators. For example, patterns of academic behaviors (c1) measured by 3 observed binary variables u11, u12, u13. Patterns of risk behaviors (c2) measured by and three observed binary variables u21, u22, u23.
Could I do a second order LCA to determine classes (f) of joint patterns of academic behaviors and risk behaviors?
Yes, this can be done. You have an example of a second-order LCA (actually a more complex LTA) in UG ex8.15. You have say 2 1st-order latent class variables c1 and c2 and a 2nd-order latent class variable c. You relate them by saying
Hi Dr Muthen, I am looking at the dyadic profiles of characteristics of husbands and wives (having run two independent LCAs for each group and achieved a two and a three-class solution for husbands, and I am looking at a joint LCA). In testing a second-order LCA, I understand that class invariance will have to be established across the two groups (as in multigroup LCA)? Would it matter if one has two different class solutions in this case? Next, I'll look at predictors of joint-class memberships..
Hi Dr Muthen, I am running a second order LCA based on a dyadic dataset (in wide format), looking at joint profiles or distributions amongst husbands and wives living in the same household. Would I have to transform the data from wide to long (with a variable identifying the dyad, husband/wife) and set it up using class-specific commands for the analysis? Or would it be possible to analyze the data in wide format? Many thanks
You can analyze it in wide format. The non-independence of observations is handled by multivariate modeling. This is discussed in both the Topic 7 and Topic 8 course handouts and videos on the website.
Hi, I am interested in a higher-order latent class variable indicated by two first-order latent class variables. I have specified an LTA relating the two latent class variables to each other and a higher order LCA with one class. When looking at the results of the two different models, I get different loglikelihoods and slightly different proportions across the joint distribution of the two latent class variables, when I expected them to be the same. Any insight into why they may be different?
I assume the LTA model is c1 -> c2. But that is the model that has "a higher order LCA with one class"? A latent class variable with one class is not a variable.
SRL posted on Tuesday, February 20, 2018 - 3:13 pm
To clarify, I specified two different models. I have “c2 on c1” for the LTA model. For the higher order LCA,I specified classes "c(1) c2(2) c3(4);” and compared the joint distribution of the two latent class variables for the LTA model and higher-order. They were slightly different (see below) and I’m trying to make sense of this difference and if its meaningful. I cannot get the higher order latent class variable with more than one class to converge.
SRL posted on Tuesday, February 20, 2018 - 3:14 pm
FINAL CLASS COUNTS AND PROPORTIONS FOR THE LATENT CLASS PATTERNS BASED ON THE ESTIMATED MODEL