3-step multilevel LCA PreviousNext
Mplus Discussion > Latent Variable Mixture Modeling >
 Lies Missotten posted on Wednesday, October 19, 2016 - 5:43 am
I have conducted latent class analysis on 8 continuous variables (4 conflict behaviors of adolescents and 4 conflict behaviors of mothers) to identify profiles of mother-adolescent conflict behaviors.
I used the 3 step method (BCH command) to validate my 4-class solution to several distal outcomes (some reported by adolescents and some reported by mothers).

One of the reviewers states that my data structure is hierarchical (mother-adolescent dyads) and that my observations are not independent. He suggests that I should do some kind of multilevel analyses to take this dependency into account. He doesn't question the dependency issue for the LCA but only for the comparison of the means of the outcome variables.

My question is:
Is it possible to conduct the 3-step method in a multilevel manner?

And if not, is there an elegant way to account for this dependency of the observations?
I used MLR as estimator.
 Bengt O. Muthen posted on Wednesday, October 19, 2016 - 5:44 pm
I assume that the latent class variable influences both the mother and child outcomes. If so, the correlation between mother and child is accounted for. Because of your "wide" analysis format, yous observational unit is the dyad, not its members and the non-independence is modeled (by the latent class variable).
 Lies Missotten posted on Friday, October 21, 2016 - 1:43 pm
Thank you for your swift response.

I would like to ask for somewhat more clarification just to be sure I understand you correctly.

Do you mean that the BCH command in the 3 step method controls for the correlations between the distal outcomes? I thought that the BCH command only generates individual tests of equality of means for each outcome variable separately (like in an ANOVA).
Could you help me understand this?

I am also not quite sure that i understand what you mean exactly with the last sentence "Because of your "wide" analysis format, your observational unit is the dyad, not its members and the non-independence is modeled (by the latent class variable)".
How does putting information from perceptions of both children’s and their mothers' conflict resolution styles in one LCA take into account the non-independence between these measures (as children and mothers are part of the same dyad)?
 Bengt O. Muthen posted on Friday, October 21, 2016 - 3:59 pm
The correlations between the distal outcomes is accounted for by them all being influenced by the latent class variable.

Wide format analysis means that if mothers have 5 variables and children have 5 variables, you are analyzing 10 variables, that is, there are 10 columns in your data (so wide). Because the latent class variable influences all 10 variables, the mother and child correlation is accounted for by the latent class variable (because it influences both types of variables).
 J.D. Haltigan posted on Tuesday, April 03, 2018 - 2:21 am
Hi Bengt:

Just wanting to follow-up on the 2nd part of your above post. Is it the case (am I understanding correctly) that in an x-sectional LCA/LPA of, say, 5 parent and 5 child interactive behavioral indicators, that the latent class variable in a way accounts for the clustered nature of the data at the dyad level? Would this change if one wanted to do a FMM and model a latent cont. variable within each class?

I ask as I have been unable to find any applications of LCA to dyadic analyses almost all of which is done in the context of lead-lag multi-level type models.

Thanks for insights.
 Bengt O. Muthen posted on Tuesday, April 03, 2018 - 10:53 am
Yes, the latent class variable influencing the two sets of variables accounts for the "clustering", that is, the correlation between the two sets of variables.

Adding a factor gives an additional vehicle for accounting for this correlation.
 J.D. Haltigan posted on Friday, April 20, 2018 - 4:59 am
Many thanks, Bengt. To clarify one further point: in the case of the x-sectional dyadic data example I provided above--parent and child behaviors in a standardized paradigm (5 indicators parent, 5 indicators child) would I need to also used TYPE=COMPLEX etc. (cf. UG chap 9) and create a 'dyad' grouping variable to account for this clustering or will the analyses itself (LPA) be enough assuming the LI assumption is met?

It's not really complex survey data but it would seem relevant.
 Bengt O. Muthen posted on Friday, April 20, 2018 - 2:57 pm
The clustering can be handled in the modeling or correcting SEs and chi-2 by Type=Complex. If you analyze the 5+5 variables in one model and the model uses latent variables(factors or classes) that influence both sets of 5 variables, then you have accounted for the clustering in the model and don't need complex. You are essentially doing a twolevel analysis but in wide format.
 J.D. Haltigan posted on Friday, April 20, 2018 - 10:33 pm
Thanks Bengt. The way you frame it in your last sentence very insightful for me.
 Heidi Lyons posted on Tuesday, May 07, 2019 - 11:50 am
Has anyone been able to successfully do a multi-level dyadic LCA? I understand the second comment above regarding the wide file. I have also read the Henry and Muthen (2010) article, and that makes sense. My question is can I put these two ideas together? I have husbands and wives and would like to include couple level variables like a variable that would measure the age difference of the couple. When I try to do the random effects model (without any covariates), I get this warning *** WARNING One or more individual-level variables have no variation within a cluster for the following clusters. I assume this means when a husband and a wife have the same response on one of my level 1 variables I use to estimate the LC I get this error. Is there any way to secessfully model this?
 Bengt O. Muthen posted on Tuesday, May 07, 2019 - 5:19 pm
Why not analyze it in single-level wide format? See our Short Course video and handout for Topic 8, slides 52-56.
 DavidBoyda posted on Saturday, September 28, 2019 - 5:11 am
May I ask how the Bch method in the 3step derives its means in the Equality test of means across the classes? Is this a regression?

I have 3 classes predicting the distal outcome (GHQ12) and the range is 0-36. The sample average is 19.8 /SD 5.8.

However, the means for the distal outcomes of the GHQ are all less than 5.0 for three classes which seem low.
 Bengt O. Muthen posted on Saturday, September 28, 2019 - 12:35 pm
Yes, the means are obtained from a regression (intercept plus contributions from predictors).

Perhaps you are looking at the intercepts of the distals, not their means.
Back to top
Add Your Message Here
Username: Posting Information:
This is a private posting area. Only registered users and moderators may post messages here.
Options: Enable HTML code in message
Automatically activate URLs in message