It's hard to answer that question because what happens behind the scenes with the special growth language depends on the type of outcome and other factors. Your second MODEL command needs the statement [prog1-prog4@0 icept linear quad]; if the outcomes are continuous. The best way to see the difference between two models is to run them and look at the results and TECH1.
Jason Bond posted on Thursday, July 05, 2007 - 2:42 pm
Bengt and Linda,
Hello. I just ran a simple quadratic LCGA model with a 5 timepoint dichotomous dependent variable (no other predictors) and saved the class probabilities variables and assigned group membership variable to an spss data file and merged it in with the original data file used for the analysis. Then I selected only those cases with a non-missing class membership value (the data file was set up as wide - with a different variable for each time point within individual so that there is only 1 row in the data file per individual) and did simple unweighted means of the dependent variables at each time point for each of the 4 values of the class membership variable. Surprisingly, the % produced from this simple descriptives procedure produced different means for the outcome variable across time for several of the classes from those represented in the Sample proportions output plot produced by Mplus. Is there a simple reason these two methods should not produced the exact same estimates? Thanks much,
The class-specific sample proportions that are plotted are derived by using the estimated posterior probabilities as weights (same as produced in Tech7). These probabilities show that a person is fractionally a member of all classes, which is not the case when you use the most likely class approach.
I am trying to conduct a LCGA model, and wondering if it is possible to have TWO criterion variables to extract classes. For example, I have both A and B variables repeatedly measured three times. And I would like to extract classes using (or based on) both A and B variables simultaneously.
Yes, this is possible. See Example 8.7 in the user's guide.
YUN HWAN KIM posted on Wednesday, November 28, 2012 - 7:29 am
I really appreciate your quick answer. But, I am not quite sure how I can use this example as my guidance. The model described in the exmaple 8.7 looks (to me) as extracting latent classes twice, but using one variable as a clustering variable. That is, once using the first four time points of data, and again using the next four time points of data (but using one variable as a clustering variable).
What I am actually wondering is if it is possible to extract latent classes once using three time points of data, but simultaneously using two different variables as clustering variables.
I am sorry if my previous explanation was not enough, or if I am not being able to appropriately catch the solution that you already suggested.
I think we understand what you mean - you want a parallel process model like what's on slide 92 of our Topic 6 handout on our web site.
That can be done in line with how ex8.7 is done. It doesn't matter that ex8.7 is a sequential model and you want a parallel process model. You just change c2 ON c1 to c2 WITH c1. In your case, y1-y4 is one process and y5-y8 another process.
If you want the same latent class variable for both processes, then you modify this example to have only one latent class variable by dropping the MODEL c1, MODEL c2.
YUN HWAN KIM posted on Wednesday, November 28, 2012 - 2:31 pm
Dr. Muthen and Muthen
I really appreciate your quick and detailed answer. I am impressed that a young scholar's voice is heard this seriously. I sincerely appreciate your advice again.
Jason Payne posted on Wednesday, August 14, 2013 - 3:05 pm
Dear Dr Muthen,
Is it possible for you to point me in the direction of some example code for an LCGA with time-varying covariates? I have been asked by replicate a PROC TRAJ model in Mplus to take advantage of the BLRT, but am struggling to specify a model with time-varying covariates. I have followed ex6.12 but cannot work out how to specify the code to measure trajectory specific effects of the t-covs.
You start with your LCGA code without tvc's and then simply add
y1 on x1; y2 on x2; ... yT on xT;
where the y's are the outcomes are the different time points and the x's are the tvc's. So here you don't have to use the random slopes of TYPE=RANDOM.
Jason Payne posted on Wednesday, August 14, 2013 - 6:07 pm
Thanks Bengt, and can I say that the support you provide to users of Mplus is first rate.
Anyway, I was wondering, using your specification above, how I identify the overall (average) impact of x on each trajectory (class) in the same way that TRAJ produces (for each class) a single estimate for x. I assume your specification gives me x1-xT estimates for each period by period regression of x on y. This is why I thought I needed to go down the Random route, but maybe im misunderstanding what TRAJ is doing as part of its tcov(x1-xT) command.
I'm not familiar with TRAJ - might its tcov(x1-xT) command imply that the effects of the x's are held equal across time? If so, you simply add a label (here "1"):
y1 on x1 (1); y2 on x2 (1); ... yT on xT (1);
If these effects are class-specific, you would mention them in each class and have different numbers for the labels.
You can try it out and see if you get the same results in the two programs. You should look for the same number of parameters and same loglikelihood.
Jason Payne posted on Thursday, August 15, 2013 - 7:22 pm
Thanks Bengt, I think this solved may problem, however, I cant tell until I solve another problem - I'm stuck with what seems to be a mis-specification of the ZIP. I have followed the example in the UG and other comments/questions on the forums to model the inflation component for i s q using:
BUT I came across Weisner et al. (2007) who said "Mplus does not offer the same flexibility as SAS Proc Traj to account for dormancy periods and exposure time".
I wasn't sure whether by "flexible" the authors meant "ease of use" rather than actual modeling capability. So that leaves me wondering whether I'll ever actually be able to replicate my ZIP LCGA with exposure adjustment in Mplus?
As a first step, you need to compare the number of parameters used in the modeling by the two programs to make sure the models are the same in that regard. You can also send your Mplus output to support together with a pdf of the TRAJ output and the Weisner reference. We can sort this out.
I am trying to determine if a LGCA is the appropriate analyses I should pursue -- I have a longitudinal dataset, and I am trying to model a nominal, categorical (three categories) variable over time. I was wondering if LGCAs are appropriate for nominal (non-ordinal data)? I was actually able to run and plot a LGCA using this data, although I'm not entirely sure what the plot is telling me, since the outcome variable is nominal. Does it reflect the predicted probability of a particular response category? If so, how do you know which response category it is plotting? Thanks in advance for your assistance!
A growth model is not appropriate for a nominal variable.
Annie Fox posted on Tuesday, January 10, 2017 - 7:27 am
I have conducted an LCGA of PTSD symptoms over time and have found 4 classes. I then used R3STEP to test whether several covariates were predictors of class membership. There was missing data on the covariates, so I imputed the covariates (10 datasets) and re-ran the LCGA using R3STEP. When I ran the imputation, I included the PTSD variables in the AUXILIARY statement so they would just be copied over to the imputed datasets. My understanding is that when I re-run the LCGA using the imputed datasets, the overall LCGA should not change (because the data for PTSD was just copied over). However, it appears that my sample size drops by 8, and therefore LCGA results are slightly different (classes are the same, but the Ns for each class are off by a few). When I open up any of the imputed datasets, they have the correct N (558). But only 550 observations are being used in the LCGA with the imputed data.