Anonymous posted on Tuesday, June 14, 2005 - 8:37 am
Hello, I am new to MPLUS and LCA. I am interested in performing LCA with 4 subscale scores and 2 single items scores in a sample of 445 participants. I want like to include seven categorical covariates and 4 continuous covariates. I have tested multiple solutions (a 1 class though a 6 class solution), and I have a couple of questions regarding the results.
Here is some information that you might need 1)The output contained a warning that one or more multinomial logit parameters were fixed to avoid singularity of the information matrix…but that it terminated normally. Does this mean the model did not run properly?
2)In the categorical latent variable section (relationship between the covariates and teach class) the estimates were huge (ex. 1169.266), the SEs were low (ex 1.033) and the Estimates/SE were huge for each covariate. In three of the groups all the covariates were significant
3)In the 4th groups all the SE were 0 and the Estimates/SE were 0.
I am wondering if the program corrected the problem so this is a valid (although odd) solution.
I am also wondering if there is a limit of the number of covariates.
BMuthen posted on Tuesday, June 14, 2005 - 9:29 am
1. No, it means that for some in some classes there is no variablity for some of the covariates so the regression coefficient is not defined.
2. This sounds like the same problem as in 1.
3. This sounds like the same problem as in 1 also. Although there is no absolute limit to the number of covariates, it is problematic when they have no variances in some classes.
Sanjoy posted on Wednesday, June 22, 2005 - 8:01 pm
Dear Professor....This is what I want to do
Y1 and Y2 are the main dependent variables (both of them are binary 0/1). Y1 and Y2 are very much correlated as well. X’s are the covariates. Total 14 Covaraites Our key interest is in capturing the variation across class in the regression of Y1, Y2 on X.
I want to fit LCA with covariates
In order to make a BIC comparison between different classes, I have tried to fit three different classes (class 3, class 4 and class 5). Classes 3 and 5 works well, however while trying to fit 4 class it’s not working.
MPlus output telling …. “the standard errors of the model parameter estimates may not be trustworthy for some parameters due to a non-positive definite first-order derivative product matrix. this may be due to the starting values but may also be an indication of model nonidentification. the condition number is -0.255d-11. problem involving parameter 45.”
Also there was something more …
“one or more multinomial logit parameters were fixed to avoid singularity of the information matrix. the singularity is most likely because the model is not identified, or because of empty cells in the joint distribution of the categorical latent variables and any independent variables. the following parameters were fixed: 38 39 50 36 37 33 35 34 32 52 51 30 31 28 29 47 53 49”
Following are the excerpts from the command syntax
This is a support question. Please send it to support. It is data-model dependent and cannot therefore be answered without seeing the full output and the data.
Sanjoy posted on Thursday, June 23, 2005 - 4:30 am
Oh ok ...that's great madam, I'm going to send you right now, thanks and regards
Anonymous posted on Thursday, June 23, 2005 - 5:42 am
Hello again, I posted the 1st message of this thread. I am becoming more familiar with MPlus and like the program. Just to remind you I would like to use LCA to derive the best model for 4 subscale scores and 2 single items scores. I have a sample of 445 participants. I have changed the covariates (now 6 continuous covariates). I have tested multiple solutions (a 1 class though a 6 class solution). Since all the measured variables correlate I am wondering if I relax the default of zero covariances across all models or is it solution specific? Also in example 7.22 you suggest starting values. I am wondering what strategy you would recommend for selecting these values.
Sanjoy posted on Thursday, June 23, 2005 - 9:26 am
Dear Madam ....since morning I have tried to send you the files three times, but for some reasons it failed to reach you ...I got email notice like "Sorry, unable to deliver your message to firstname.lastname@example.org for the following reason:
The correlations among the measured variables are accounted for by having several latent classes. The fact that the within class correlation is zero does not mean that the correlations among the measured variables is zero.
Usually start values are guided by theory. If you don't have a good feel for what they should be, I suggest using the default.
I was wondering if there was a difference in output if you specified covariates in the USEV statment but never regressed the covariates on the class structure (no model statement) versus regressing them on the classes (e.g., Model: c#1 - c#3 on A B C).