Hi, my name is Alicia Merline. I attended the November Mplus training in VA. I am using Mplus to perform GGMM on substance use data across 5 timepoints.
Mplus produced the following error message:
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 -0.549D-16.
THE STANDARD ERRORS OF THE MODEL PARAMETER ESTIMATES COULD NOT BE COMPUTED. THIS IS OFTEN DUE TO THE STARTING VALUES BUT MAY ALSO BE AN INDICATION OF MODEL NONIDENTIFICATION. CHANGE YOUR MODEL AND/OR STARTING VALUES. PROBLEM INVOLVING PARAMETER 14.
I have been getting similar error messages consistently regardless of which data I use (cigarettes, alcohol or marijuana) and the number of classes I specify. Sometimes the parameter number is different. I have tried fixing the problemed parameter to zero and this results in output stating that next parameter has a problem.
Here is an example of my input (in this case for cigarettes with 5 latent classes.
Although Mplus produces several warnings, the classes identified by the program look quite reasonable.
As you may be able to tell from my file name and area of research, I am part of the Monitoring the Future project at the University of Michigan. John Schulenberg is my main collaborator for the current research project. I've also discussed this project with Patrick O'Malley. Dr. O'Malley is the only other member of our group with Mplus experience.
I am not sure if you have already done this. But I would start with two classes and only the %overall% part of the MODEL command allowing the Mplus defaults to be used. Only once I got the two class model with defaults working would I go to 3 classes and/or relax some of the defaults. Also, I would use STARTS = 100 10; instead of the default of STARTS = 10 1; if taking the stepwise approach does not work.
Also, it is not clear which version of Mplus you are using. If you cannot resolve this send your input and data along with your license number to email@example.com.
bmuthen posted on Monday, March 07, 2005 - 5:00 pm
I would second Linda's suggestions. To get insights into why your model is not identifed, you can look at tech1 to see what is parameter 14 that the error refers to. For example, if a class has only one or two individuals in it, parameters specific to this class are not identified. Also, I would recommend considering 2-part modeling as an alternative to censored modeling - it is both easier to work with (faster) and more flexible (refs on our web sie).
I have done the most work with the alcohol model, so I'm just going to talk about this model right now.
The two class model works fine with no error messages. When I move to a three class model the problems begin. I tried not specifying starting values and still get errors. In this case, the parameter 14 is listed as causing the problem. 14 is the Alpha for the intercept of the third latent class (-10.365). Size of groups is not a problem. The smallest group has 481 members.
When I run this same model but using start values, Parameter 5 is listed as having a problem. 5 is the Alpha for the quadratic term in latent class 1 (159.714).
I should note that latent class 3 in the unrestricted model and latent class 1 in the model with starting values are the same.
I tried a new model with starting values for the first two classes only. In this model I again get a problem with the intercept for the same group (now classified into group 2).
When I look back at my previous attempts at getting a 5 class solution I see that the problem class has a mean of 1 at each time. This is true of the problem clss in the 3 class solution. So, the real issue seems to be with the (large) class of respondents who abstain from heavy drinking at all 5 timepoints.
Thank you for your suggestions, they certainly helped me to define the problem more clearly.
I think that I have two options now--exclude abstainers from analyses (we have a history of doing this at MTF) or perhaps try the 2-part modeling. Do you think 2-part modeling would solve the problem?
Thank you again,
bmuthen posted on Tuesday, March 08, 2005 - 5:34 pm
Why don't you send your input, output, and data to firstname.lastname@example.org - we can probably see the problem with your 3-class GMM.
Sending my input and output would not be a problem. However, because of its sensitive nature, accesss to the MTF data is limited quite strictly. I am inquiring with the PIs about possibly sharing data with you.
I installed the update today and now the error messages have changed completely. Now I get a message stating that two parameters have been fixed to achieve identifiability. These two parameters are the alphas for the intercept and slope of the abstainer group. I am assuming that they have been fixed at 0, is that correct?
You can see what they are fixed at in the output. This happens when values go very large. The standard errors are fixed at zero.
Anonymous posted on Thursday, April 07, 2005 - 4:57 pm
Hi I am using GGMM for longitudinal data with missing values. here is the warning messages from Mplus when I am using mixture missing;
THE MODEL ESTIMATION DID NOT TERMINATE NORMALLY. THE COMPUTATION OF THE FISHER INFORMATION MATRIX COULD NOT BE COMPLETED.
THE STANDARD ERRORS OF THE MODEL PARAMETER ESTIMATES COULD NOT BE COMPUTED. THIS IS OFTEN DUE TO THE STARTING VALUES BUT MAY ALSO BE AN INDICATION OF MODEL NONIDENTIFICATION. CHANGE YOUR MODEL AND/OR STARTING VALUES. THE SAMPLE COVARIANCE MATRIX COULD NOT BE INVERTED FOR CLASS 2. PROBLEM INVOLVING VARIABLE X4.
X4 is a dichotomous covariate. the output doesn't have goodness of fit test due to the warnings. I am just wondering if the estimates for the parameters are still reliable. Thanks, Lulu
It is likely that x4 has no variability in class 2. In this case, the regression coefficient for x4 cannot be determined. Fit statistics are not available for mixture models. Also, be sure to use Version 3.12.
Socrates posted on Monday, February 20, 2006 - 3:13 pm
Dear Dr. Muthén
When running different GMMs on different datasets, I often get the following error warning:
THE MODEL ESTIMATION TERMINATED NORMALLY
WARNING: THE RESIDUAL COVARIANCE MATRIX (PSI) IN CLASS 1 IS NOT POSITIVE DEFINITE. PROBLEM INVOLVING VARIABLE s.
In this discussion, I read that this error warning is serious. However, what can I do to avoid it / to deal with it? Since, as I mentioned before, this message appears very often, I have the feeling that perhaps it is a technical problem?
Thanks and best regards!
bmuthen posted on Monday, February 20, 2006 - 3:22 pm
A typical cause is that your estimated variance for s is negative. This suggests fixing this variance at zero so that s is a fixed rather than a random effect. This happens more often than in single-class growth models because the classes absorb much of the variation in the growth factors.
Socrates posted on Tuesday, February 21, 2006 - 6:20 am
Dear Dr. Muthén
Many thanks for the fast reply! While the estimated variance for s is not negative, I found in the TECH4 output a correlation of the variable s with another variable s2 that is greater than one. s is the slope of a process A and s2 the slope of a parallel process B. Does this circumstance also suggest fixing the variance of s at zero?
bmuthen posted on Tuesday, February 21, 2006 - 9:27 pm
No it does not. A common mistake with parallel processes is to omit contemporaneous correlations between the outcome residuals for the different processes. This then channels too much of the correlation between the observed outcomes through the growth factors, causing unit correlation.
socrates posted on Thursday, November 02, 2006 - 7:01 pm
dear dr. muthen
running a GGMM with predictors of class membership probabilities and growth parameters results in the following error message:
WARNING: THE RESIDUAL COVARIANCE MATRIX (PSI) IN CLASS 1 IS NOT POSITIVE DEFINITE. PROBLEM INVOLVING VARIABLE S.
WARNING: THE RESIDUAL COVARIANCE MATRIX (PSI) IN CLASS 2 IS NOT POSITIVE DEFINITE. PROBLEM INVOLVING VARIABLE S.
WARNING: THE RESIDUAL COVARIANCE MATRIX (PSI) IN CLASS 3 IS NOT POSITIVE DEFINITE. PROBLEM INVOLVING VARIABLE S.
WARNING: THE RESIDUAL COVARIANCE MATRIX (PSI) IN CLASS 4 IS NOT POSITIVE DEFINITE. PROBLEM INVOLVING VARIABLE S.
WARNING: THE RESIDUAL COVARIANCE MATRIX (PSI) IN CLASS 5 IS NOT POSITIVE DEFINITE. PROBLEM INVOLVING VARIABLE S.
s ist the residual variance of the slope. it is negative and does not differ significantly from zero. when fixing s@0 the model converges without error message. is this prcedure sound? and if not: what can I do to avoid the problem of negativ residual variances?
Matt Moehr posted on Tuesday, January 16, 2007 - 7:03 pm
I have a couple of follow-up questions regarding the error message for non-positive definite psi matrices, and the subsequent fixing of S.
1) What does the negative variance for S indicate? There's no variation among the individual slopes? I'm running a single class model (although with multiple cohorts), so the class differences aren't sucking up all the variation. Also, the individual plots would seem to contradict this.
2) By assuming that the variance of S is zero, does the interpretation of any other parameter change? Can I and S still be used in the linear equation, y = I + S*x, to come up with predicted values?
3) I went ahead and used S@0, but then got this warning: "All continuous latent variable covariances involving S have been fixed to 0 because the variance of S is fixed at 0." So I fixed S@1, and I got a model with good looking estimates but much worse model fit. Can I pick arbitrary, non-zero values for the variance of S?
1. It can indicate the the slope growth factor is a fixed effect instead of a random effect or it can indicate that the model is not appropriate for the data.
2. No. Yes.
3. It makes sense to fix the variance to zero if it is is very small and non-significant. I don't think fixing the variance to one or another arbitrary value is justified.
mihyun park posted on Wednesday, April 06, 2011 - 8:07 am
dear dr. Muthen
Mplus produced the following error message:
***WARNING Data set contains cases with missing on all variables. These cases were not iincluded in the analysis. Number of cases with missing on all variables: 1 1 WARNING(S) FOUND IN THE INPUT INSTRUCTIONS
After this warning, i repeat again..and then, number of observations on my result output decreased 2843 to 1422. I don't know why this result happened and what tells me..