DATA: FILE IS comb1log.dat; VARIABLE: NAMES ARE age t tdur liv x1-x9; USEVARIABLES ARE t x1-x9; CLASSES = c (4); Missing are all (-999) ; CATEGORICAL = t; ANALYSIS: TYPE = MIXTURE; Type=Missing; STARTS = 20 2; estimator=mlr; algorithm = integration; Coverage=0.02; MODEL: %OVERALL% i s | x1@0x2@1x3@2x4@3x5@4x6@5x7@6x8@7x9@8; s on t; c#1 on t; c#2 on t; c#3 on t;
But I have got difficulty in understanding the output - in particular, MPLUS outputs that:
Categorical Latent Variables
C#1 ON T -1.516 -0.985 0.709 2.402 2.934
C#2 ON T 9.235 9.706 11.205 12.704 13.175
C#3 ON T -2.165 -1.991 -1.439 -0.886 -0.713
The above does no make sense to me, as classes 2 and 3 are virtually identical when looking at the graph (except that the intercept for 3 is slightly higher than for 2). Could you help me with the interpretation please?
linda beck posted on Wednesday, January 14, 2009 - 9:39 am
I would say, if two of your classes look nearly identical you should prefer a 3-class solution instead of 4 classes independently of what fit- or test criteria say. May be that would help to get more plausible effects of c on t.
Linda - thank you for the offer but unfortunately I am unable to send the data. I did, however, realise over the weekend that I should have specified T in dummy variables as it is a nominal categorical variable. That has solved the strange output problem.
Still, I have got the following questions:
1) I am assuming that the model specification above does not estimate the effect of T on variability within classes? How would I specify that?
2) How do I specify that I would like the effect of s on T to be different across different classes?
That's an easy one I guess! I had the same problem some time ago. you should mention i on t in the %obverall%-statement (if defined class invariant) and the class variant s on t in the class-specific statements (i think you have 4) for each class. it goes like this:
%c#1% s on t;
%c#2% s on t;
... The same procedure with other model parameters. Furthermore, I would increase the starts (you try to extract 4 classes!!!) up to at least 500 20, to avoid local maxima. stiterations = 20 also helps a lot to get more trustworthy solutions, especially when you have class variant effects. In addition, class variant effects can often lead (that's my experience) to LL's not replicated. Then stscale = 1 sometimes helps to replicate a LL-value. your coverage seems pretty low, which can also lead to estimates that are not trustworthy. Is this what you intended?