Is it possible to estimate a latent difference model for multiple groups with Mplus?
The problem I see is as follows: Usually, in latent difference models the intercept of the later state isnīt estimated (because it is determined by the mean of the first state and the mean of the latent difference variable). In multi-group analyses, Mplus estimates the intercepts of the later states (and I havenīt found a way how to tell the program that they shouldnīt be estimated). Do you have an idea how I could solve this problem?
Thank you for your quick reply! But I guess, fixing the intercepts to zero isnīt the solution for this problem. Let me try to reformulate and clarify my question.
In latent difference models, the intercept of the later state isnīt estimated (perhaps because it can be calculated: [state2]-[state1]=[diff2-1] ==> [state2]=[state1]+[diff2-1]).
When I tried to estimate a model in a multi-group analysis (which could be estimated in single-group analysis without problems), Mplus said THE MODEL MAY NOT BE IDENTIFIED. The output told me that I should check the mean of the latent difference variable. Thatīs how I came to the idea to check the means and intercepts. In the preliminary results shown I noticed that in multi-group-analysis Mplus estimates the intercept of state2 (which is not the case in single-group analysis). Then I calculated if [state2]-[state1]=[diff2-1] and saw that this was not the case! How can this happen?
Yes thatīs right, with this constraint the model can be estimated. But I noticed that the results seem implausible. In single-group analysis, the intercept of the latent difference variable is positive (e.g. 0,64). In multi-group analysis, it is negative in all groups (e.g.-3,36). Thatīs why I began to wonder, wether it is possible to estimate latent difference models in multi-group analyses with Mplus.
If the model is identified, I'm sure it can be estimated in Mplus. I just have not done this type of analysis to know how exactly it should be set up. Perhaps you want to contact those who originated it.
Dear Mr. Muthén, thank you very much for your answers, thatīs great to have this support!
mdehne posted on Thursday, April 18, 2019 - 2:25 am
Dear Mr. Muthén
I am trying to conduct a multigroup latent change analysis. Therefore, I used the Steyer et al. parameterization. My question is regarding the testing for group differences in the latent change scores. I realized the model as a correlated uniqeness model as follows:
MODEL: t1 by var34 var35 (1) var34_t2@1 var35_t2 (1) var34_t3@1 var35_t3 (1); diff2_1 by var34_t2 var35_t2 (1); diff3_1 by var34_t3 var35_t3 (1);
Usually, the intercepts of the first indicators are fixed to be "0" to freely estimate the latent mean scores; these intercepts are now constrained to be equal across time. Is it correct to fix mean scores as specified above to test for group differences in the latent means/change scores?
My colleagues and I have constructed a latent change score model with two timepoints to assess differences between treatment and control groups. We would now like to examine child age (a binary variable: <=18>18 months) as a moderator. We plan to do this with a multigroup analysis, using child age as the grouping variable. However, when I run the analysis, I get this error message: THE STANDARD ERRORS OF THE MODEL PARAMETER ESTIMATES COULD NOT BE COMPUTED. THE MODEL MAY NOT BE IDENTIFIED. CHECK YOUR MODEL. PROBLEM INVOLVING THE FOLLOWING PARAMETER: Parameter 32, Group GT18: [ COGDIFF ]
THE CONDITION NUMBER IS 0.411D-17.
THE ROBUST CHI-SQUARE COULD NOT BE COMPUTED.
FACTOR SCORES WILL NOT BE COMPUTED DUE TO NONCONVERGENCE OR NONIDENTIFIED MODEL.
Four posts above I asked whether my multigroup latent change score model is correctly specified. According to our both opinion, it was, yes...
However, if I depth into the model outputs (e.g., the model estimated means), I saw that the latent change score in the group where the means are freely estimated are not "corrected" for the latent change score in the reference group. To my opinion (as in other multigroup CFA frameworks), I should be able to derive the change score by subtracting the respective marker item's mean t2-t1, and then subtracting these change scores?! On the other hand, even if I estimate the means for the change variables in both groups (i.e., by fixing the marker item's intercept at zero to identify both respective means) what accordingly leads to the same estimates as done by hand, the same problem arises. Is it possible that you take a look at my computations?