Wei Tang posted on Thursday, April 16, 2009 - 10:36 pm
I¡¯m trying to model people¡¯s shopping intention using LCM. Earlier, I use 2 classes, the model runs fine. Recently, we switched to 3-class but didn¡¯t get even one reasonable result. There are several error messages. Pls help me check is there anything wrong with my code? I am a newbie for Mplus, really need some help to get started. Thanks!
FYI, my 3-class code is:
¡ VARIABLE: NAMES ARE ID INTENT AGE PPS CONVEN BBW COSTD; MISSING IS .; USEVARIABLES ARE INTENT AGE PPS CONVEN BBW COSTD; CLASSES = c (3); CATEGORICAL = INTENT;
ANALYSIS: TYPE = MIXTURE; ESTIMATOR IS MLR; ALGORITHM = EMA;
MODEL: %OVERALL% INTENT ON PPS CONVEN BBW COSTD; c ON AGE; %c#2% INTENT ON PPS CONVEN BBW COSTD; %c#3% INTENT ON PPS CONVEN BBW COSTD; ¡
I just changed the # of classes from 2 to 3 and added ¡°%c#3%¡± part under the "MODEL" part, does my code looks right? Pls help me check!
Wei Tang posted on Thursday, April 16, 2009 - 10:38 pm
Another question: in many cases, we got message like ¡°ONE OR MORE 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 VARIABLES IN THE MODEL. THE FOLLOWING PARAMETERS WERE FIXED: 4 7 8 9 20 2 19 14¡±. So what are those numbers (¡°4¡±, ¡°7¡±, ¡°8¡± etc.) mean?? Is there a way for us to know which parameters were fixed on earth in the modeling process?
You should use more random starts than the default Starts = 10 2 (see the User's Guide about this). For instance, add to the ANALYSIS command:
STARTS = 400 100;
and make sure you have several replications of the best loglikelihood value.
Add TECH1 to your OUTPUT command and you will see which parameters have been automatically fixed - this is not a problem.
Laura Tang posted on Friday, April 17, 2009 - 2:37 pm
Thanks! Yes, I was aware of the random start value and already tried to add the "STARTS = 500 20" and even "STARTS = 1500 50".But each STILL gave me a lot of error messages too (such as "THE BEST LOGLIKELIHOOD VALUE WAS NOT REPLICATED. THE SOLUTION MAY NOT BE TRUSTWORTHY DUE TO LOCAL MAXIMA. INCREASE THE NUMBER OF RANDOM STARTS.", "STARDARD ERROR MAY NOT BE TRUSTWORTHY FOR SOME PARAMETERS DUE TO A NON-POSITIVE DEFINITE FIRST-ORDER DERIVATIVE PRODUCT MATRIX." and above-mentioned fix parameter message).
So did u see any coding error in my original post? What else u would suggest to try?
The input looks fine, although the model is challenging with these class-varying slopes. Note further that AGE may have a direct effect on your DV INTENT, so not only influencing c. Also, there may not be more than 2 classes supported by these data.
I just have a general question regarding the idea behind latent mixture models. In the first chapter of the Latent Variable Mixture Modeling book, you describe mixture modeling and use a picture of a probability distribution. I am wondering if you can clarify just a bit for me. Would there actually be a mixture distribution like this for each indicator in the profile analysis? I understand that there are multiple distributions, but how are multiple indicators taken into account? Is this just a theoretical model and the mean just somehow represents a combination of the indicators for each class? Thank you for any clarification on this
If you look at e.g. the Branch 1 distributional graph, you see a mixture of 2 factor distributions. The graph below it shows how that is modeled. The model implies that each indicator distribution follows a mixture as well (as a function of the factor), but perhaps not as accentuated as the factor due to indicators also being influenced by residuals (measurement error). It is the distribution of all indicators that drives the estimation of the latent classes and thereby the determination of the factor distributions. Hope that answers your questions.
Chris Weber posted on Wednesday, December 16, 2009 - 8:03 am
I am trying to estimate a factor mixture model. The data are 14 individual cross sections taken every two years from 1976-2004. I am most interested in seeing if the factor loadings for two classes vary across time and classes, so I specified an interaction between the indicators and a variable that ranges from 1 to 14. The model should be identified (am I right?). Increasing the number of random starts doesn't seem to solve the problem.
%OVERALL% FB1 BY ID ABORT WOMEN AID JOB SPEND count d2 d3 d4 d5 d6; THE MODEL ESTIMATION DID NOT TERMINATE NORMALLY DUE TO AN ILL-CONDITIONED FISHER INFORMATION MATRIX. CHANGE YOUR MODEL AND/OR STARTING VALUES.
THE MODEL ESTIMATION DID NOT TERMINATE NORMALLY DUE TO A NON-POSITIVE DEFINITE FISHER INFORMATION MATRIX. THIS MAY BE DUE TO THE STARTING VALUES BUT MAY ALSO BE AN INDICATION OF MODEL NONIDENTIFICATION. THE CONDITION NUMBER IS 0.609D-11.