Message/Author 

Anonymous posted on Monday, August 26, 2002  9:18 am



In the MPLUS specification of a LCGA, does one need to explicitly specify that the variances of growth parameters and the covariances between growth parameter within class are equal to zero? I searched through your examples, but could not find a setup for the LCGA. Thank you! 


Yes, you do need to explicitly specify that the growth factor variances and covariances are zero. For example, i@0; s@0; i WITH s @0; 


Just to make sure: is that still the case ? 2002, in Mplus time, is already old... With continuous outcomes, the difference between LCGA and GMM is that in LCGA you specify : i@0; s@0; i WITH s @0; Whereas the difference between LCGA and GMM with noncontinuous outcomes is that you have to rely on numerical integration to obtain GMM. Is that it (from the Mplus input perspective)? 


The LCGA and GMM models are the same for continuous and categorical outcomes. LCGA fixes variances and covariances for the growth factors at zero while GMM has free growth factor variances and covariances. For categorical outcomes, numerical integration is required for model estimation for GMM but not for LCGA. 


Greetings and thanks for the answer, This confirm what I tought. But I think my question was not clear. In the Mplus input for noncontinuous outcomes LCGA, if we dont specify "numerical integration" then do we still have to specify i@0; s@0; i WITH s @0. In the manual examples, no such specification are given for the noncontinuous outcomes LCGA (ex 8.9, 8.10, 8.11). 


No such specification is given because it is not needed. 


Can a latent class growth analysis be performed when the indicators are dichotomous? Is it OK to construct a latent growth model with the latent intercept and slope factors as we typically think of them, when the indicators proceed in 01 patterns such as 0001 or 0111 or 0011? Or would this be inappropriate for a latent growth model? If so, is there another model that you can suggest? I am very interested in the mixture models in Chapter 8 of the Mplus manual, but I am not sure if they are appropriate for my analysis. Thank you for any guidance you can provide. 


Yes, this can be done as shown in Example 8.4 for GMM and Example 8.9 for LCGA. 


Thank you, could aspects of the models in Chapter 8 be combined? I like the model shown in EXAMPLE 8.7: A SEQUENTIAL PROCESS GMM, as we have two sequential processesbut could I adapt it: a) to have known classes as in Example 8.8, and b) to have dichotomous outcomes rather than continuous? Thank you. 


Yes, that can all be done. 


Dear Dr. Muthen, I am trying to run a 3class LCGA with a categorical outcome variable 5 categories). I want to add a categorical covariate (3 categories) to influence intercept and slope of the latent trajectories. The model looks something like that: Model: %Overall% i s  u15@0 u16@1 u17@2 u18@3 u19@4 ... i s on x1 x2; c on x1 x2; I understand that with the conditional model, I do not get the results in probability scale anymore, but that I may be able to calculate the yearly probabilities of partnership status for each class by hand  right? I found this post explaining the procedure with tech 4 with the following formulas for a latent growth model with a binary categorical variable: s_m = a+g*x and P (y=1 x) = F(tau + s_m*x_t). Does that also hold for a LCGA with multiple classes because the thresholds are equal across time and across classes by default. Or do I need to free the thresholds? Or is there a different way to calculate the conditional probabilities for each year by partnership status? Also, when I try to run tech4 in my analyses, the output just remains bank for tech4. Does tech4 require the savedata command here? Thank you! 


I assume that your 5cat outcome is ordinal. And also that i and s don't have variance parameters estimated. Then it isn't hard to compute by hand. You can work in terms of the y* latent response variable underlying the ordinal at each time point because then you can use the formulas for the mean and variance used for growth in an observed continuous outcome shown in Topic 3. Then you use that together with the thresholds for the ordinal as in regular ordinal regression  you can find this in either our book's chapter 5 or in the technical appendix 1 on our website: http://www.statmodel.com/download/techappen.pdf 


Dear Dr. Muthen, Thank you for you quick reply! Yes, variances for i and s were not estimated (I left all default settings), but the 5cat outcome is in fact nominal, not ordinal. Does that make a difference? I also looked into the materials you recommended, but I am still left confused on how to estimate the conditional probabilities. Could you please provide a little bit more guidance on that issue specifically? Thanks and kind regards! 


Are you specifying the outcome as Nominal? If specified at Categorical, it is assumed to be ordinal. The difference is important because I don't know how a growth model for a Nominal outcome should be interpreted given that a Nominal DV has several, not just one, slope. 

Back to top 