Warnings in LCA output PreviousNext
Mplus Discussion > Latent Variable Mixture Modeling >
Message/Author
 Svenja Schmidt posted on Thursday, October 30, 2008 - 9:39 am
Dear Professors Muthen

I got MPlus to run and figured that as I have lots of data it would be a good idea to aggregate them through a LCA. The model runs fine for 2 classes, when using more than two classes it omits the warning "WHEN ESTIMATING A MODEL WITH MORE THAN TWO CLASSES, IT MAY BE NECESSARY TO INCREASE THE NUMBER OF RANDOM STARTS USING THE STARTS OPTION TO AVOID LOCAL MAXIMA." So my first question would be: Is this just a thoughtful reminder of the program or does it actually indicate any trouble with my data?

When I increase the number of classes in order to find the best fitting LCA model, once I get to 5 classes the program omits warnings like "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 -0.122D-10. PROBLEM INVOLVING PARAMETER 118." How do I deal with this warning (preferably in syntax terms as I have no clue about statistics)?
Thanks for your help!
 Linda K. Muthen posted on Thursday, October 30, 2008 - 10:55 am
This is a reminder that it is necessary to replicate the best loglikelihood to be sure you have not reached a local solution.

I would need to see the full output to answer this. You may be extracting too many classes. Please send your output and license number to support@statmodel.com.
 Keng-Han Lin posted on Friday, February 26, 2010 - 12:01 pm
Hi,

My random start setting is as "STARTS = 500 10; STITERATIONS = 20;" for a 4-class model. But I got a error message saying "WARNING: THE BEST LOGLIKELIHOOD VALUE WAS NOT REPLICATED. THE SOLUTION MAY NOT BE TRUSTWORTHY DUE TO LOCAL MAXIMA. INCREASE THE NUMBER OF RANDOM STARTS."
Therefore, I tried setting as "STARTS = 2000 200; STITERATIONS = 40;", however, the best loglikelihood value still wasn't replicated.
Should I keep trying more random start which is a very time-consuming calculation.
Or this is maybe I'm extracting too many classes from the data?

Thank you so much.

Keng-Han
 Bengt O. Muthen posted on Saturday, February 27, 2010 - 6:26 am
Starts = 2000 200 is already a fairly large search for the best LL (although I would use a ratio closer to 4:1, and not change the default STITER, so starts = 2000 500, STITER=20), so there is probably a good reason why you don't get replicated best LLs. The reasons can be asking for too many classes as you say, too complex a model, or numerical integration that makes the LL calculation have less precision. For the latter our web handout for Topic 5 shows how to compare the top LL solutions with respect to their estimates to see if the solutions are really different.
 Keng-Han Lin posted on Monday, March 01, 2010 - 7:49 am
Hi,

The best LL is still not replicated given the setting, starts = 2000 500, STITER=20, however, the BIC is better than 3-class model.
I'm now wondering which model we should choose, could we say 4-class model is not stable due to extracting too many classes, so we go with 3-class model?

Thank you.

Keng-Han
 Linda K. Muthen posted on Monday, March 01, 2010 - 9:24 am
This is a question for which the answer is based on too much information for us to be able to say in this forum. You might need statistical consulting help.
 Ingrid Carlson posted on Tuesday, March 08, 2011 - 10:54 am
Dear Professors Muthen, I'm running an LCA MonteCarlo simulation, and I've been getting these errors even when I know the model is intrinsically identified:

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 0.982D-11. PROBLEM INVOLVING PARAMETER 4.

Can you tell me more about what this error message means?
 Linda K. Muthen posted on Tuesday, March 08, 2011 - 11:15 am
It says your model may not be identified and suggests it is due to parameter 4. You can look at TECH1 to see which parameter this is. If you cannot figure it out, please send your full output and license number to support@statmodel.com.
 Janet Smith posted on Tuesday, April 19, 2011 - 1:05 pm
Dear Professors Muthen

My model is the calculation of three classes for two known groups.

The syntax is as follows:

NAMES id group e1 e2 e3;
USEVARIABLES group e1 e2 e3;
IDVARIABLE ARE id;
MISSING ARE ALL (-999);
CLASSES = cg (2) c (3);
KNOWNCLASS = cg (group = 1 group = 2);

ANALYSIS:
TYPE = MIXTURE;
STARTS = 500 50;

I am getting the error message:

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 -0.240D-20. PROBLEM INVOLVING PARAMETER 9.

The problem appears to be with variable e3 for the latent class pattern 1 2.

PARAMETER SPECIFICATION FOR LATENT CLASS PATTERN 1 2
NU
E1 E2 E3
7 8 9

I have tried to fix the problem by increasing the number of starts but this does not work. Therefore is the next option to fix the parameter e3 for the second latent class of my first group?

If so, please can you tell me how to do this in the model?

Many thanks.
 Linda K. Muthen posted on Tuesday, April 19, 2011 - 2:12 pm
Please send the full output and your license number to support@statmodel.com.
 Dennis Schneider posted on Monday, February 04, 2013 - 9:54 am
Dear Professors Muthen,

i'm running an school development research analysis of over 250 schools to find out which of them are similar to each other. Therefor i'm using the LCA Method with MPlus 5.21. While increasing the number of tested classes I compare H0/AIC/BIC to figure out which solution might be the best.
To reduce the number of Warnings I increased the random starts (now 10000 100) and the iterations (now 5000). Even I tried bootstrapping but could not get rid of the Warnings. I also recoded my variables to make the distribution less skewness.

My Warings are:
WARNING: WHEN ESTIMATING A MODEL WITH MORE THAN TWO CLASSES, IT MAY BE
NECESSARY TO INCREASE THE NUMBER OF RANDOM STARTS USING THE STARTS OPTION
TO AVOID LOCAL MAXIMA.


WARNING: THE BEST LOGLIKELIHOOD VALUE WAS NOT REPLICATED. THE
SOLUTION MAY NOT BE TRUSTWORTHY DUE TO LOCAL MAXIMA. INCREASE THE
NUMBER OF RANDOM STARTS.

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 0.429D-11. PROBLEM INVOLVING PARAMETER 35.

regards Dennis Schneider
 Linda K. Muthen posted on Monday, February 04, 2013 - 11:10 am
You would need to use Version 7 and send the output and your license number to support@statmodel.com for us to be able to help you. This is not enough information.
Back to top
Add Your Message Here
Post:
Username: Posting Information:
This is a private posting area. Only registered users and moderators may post messages here.
Password:
Options: Enable HTML code in message
Automatically activate URLs in message
Action: