Non-positive definite matrix PreviousNext
Mplus Discussion > Structural Equation Modeling >
Message/Author
 Laney Sims posted on Thursday, August 31, 2006 - 9:02 am
I have a simple model for a longitudinal study with two variables over three points in time. My input statement is:

Analysis:
type = missing H1;
estimator=ML;
Model:
pbrt2 on pbrt1 aainv1;
aainv2 on pbrt1 aainv1;
pbrt3 on pbrt2 aainv2;
aainv3 on pbrt2 aainv2;
pbrt1 with aainv1;
pbrt2 with aainv2;
pbrt3 with aainv3;

Output:
sampstat;
standardized;

I am getting the following message in the output:

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 0.176D-13. PROBLEM INVOLVING PARAMETER 23.

Could you explain how my results might be affected by this?

Thank you.
 Linda K. Muthen posted on Thursday, August 31, 2006 - 6:50 pm
The results are not valid if a model is not identified. If you ask for TECH1 in the OUTPUT command, you can see what parameter 23 is. If this does not help, please send your input, data, output, and license number to support@statmodel.com.
 Laney Sims posted on Friday, September 01, 2006 - 7:37 am
Parameter 23 is the value in the PSI matrix corresponding to aainv1 vs. aainv1, which appears to be the variance of that variable. I don't see anything unusual about the variance of aainv1 (0.234) compared to the others in the SAMPSTAT output or the MODEL output. Do you have any suggestions, or should I email my files for additional help?

Thank you.
 Linda K. Muthen posted on Friday, September 01, 2006 - 8:09 am
Please send your input, data, output, and license number to support@statmodel.com.
 Kelvin Choi posted on Thursday, September 04, 2008 - 6:52 pm
I have a similar problem as described above. I asked for TECH1 and found out the parameter is edu vs. edu (covariance coverage=0.935). Are the results still trustworthy? If not, how can I resolve this issue?

Thanks.
 Linda K. Muthen posted on Friday, September 05, 2008 - 6:22 am
Please send your input, data, output, and license number to support@statmodel.com.
 Vanessa Wight posted on Tuesday, August 25, 2009 - 7:21 am
I am having a similar problem when I weight my data. That is, the model converges fine with unweighted data, but when I add the weight, I get the same 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 0.209D-10. PROBLEM INVOLVING PARAMETER 114.

If the results were fine with unweighted data, can I trust the weighted results?
 Linda K. Muthen posted on Tuesday, August 25, 2009 - 7:39 am
I could not say without more information. Please send your input, data, output, and license number to support@statmodel.com.
 Melinda Gonzales-Backen posted on Thursday, June 17, 2010 - 9:32 am
I'm having a similar problem. I am running a multigroup analysis. The warning also points to the variance of one of my variables. I looked at the descriptives for this variable and everything looks ok, but there is a lot of missing data on this variable. Could the amount of missing data cause this problem? Is there a way that I can still include this variable in my model? It is a very important control. Thanks!
 Linda K. Muthen posted on Thursday, June 17, 2010 - 10:27 am
I could not say without more information. Please send your input, data, output, and license number to support@statmodel.com.
 Kimberley Breevaart posted on Thursday, March 31, 2011 - 4:22 am
Dear Linda,

I am testing a sequential mediation model with Mplus. However, when I reverse my model (to say something about causality), I get the following error message:

MAXIMUM LOG-LIKELIHOOD VALUE FOR THE UNRESTRICTED (H1) MODEL IS -5603.305

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 0.236D-12. PROBLEM INVOLVING PARAMETER 45.

It does run my model, but my reversed model indicates better fit, which makes no sense theoretically. Is it possible this error message influences my results?

Kind regards, Kim
 Linda K. Muthen posted on Thursday, March 31, 2011 - 5:17 am
I would need to see the full output to answer this. Please send it and your license number to support@statmodel.com.
 Elise Pas posted on Friday, June 24, 2011 - 11:40 am
I have gotten the same error as discussed above for a CFA conducted with a clustering variable. 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 -0.260D-17. PROBLEM INVOLVING PARAMETER 23

In addition, my output said: THIS IS MOST LIKELY DUE TO HAVING MORE PARAMETERS THAN THE NUMBER OF CLUSTERS MINUS THE NUMBER OF STRATA WITH MORE THAN ONE CLUSTER.

I entered the cluster variable to account for nesting in my data but will not be including any between-level specifications. To test if the issue is in fact the cluster size, I ran the same CFA without the cluster ID (or the type=COMPLEX command) and the entire error message disappeared. It seems I received this message because of my cluster size (n = 23) being smaller than the number of freed parameters (i.e., 31) but what I do not understand is why this error message would also indicate an issue with a parameter (which in this case was the psi of an observed variable). Are the number of freed parameters actually the problem? And if so, how do I solve this problem?

Thanks,
Elise
 Linda K. Muthen posted on Friday, June 24, 2011 - 4:50 pm
The number of clusters is the number of independent observations in your data set. The warning is telling you that you have more parameters than you have independent observations. The impact of this on the results has not been studied. This is simply a warning.
 Hallie Bregman posted on Thursday, May 03, 2012 - 12:12 pm
Hi,

I am running a path model with three latent "predictors", a single outcome, and 4 covariates (age, gender, two dummy codes for ethnicity). When I run the path model without any of the covariates, my model converges and achieves adequate fit. However, when I include paths from each of the 4 covariates to the outcome variable, I recieve the following error messages:

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 0.316D-18. PROBLEM INVOLVING PARAMETER 52.


WARNING: THE LATENT VARIABLE COVARIANCE MATRIX (PSI) IS NOT POSITIVE
DEFINITE. THIS COULD INDICATE A NEGATIVE VARIANCE/RESIDUAL VARIANCE FOR A
LATENT VARIABLE, A CORRELATION GREATER OR EQUAL TO ONE BETWEEN TWO LATENT
VARIABLES, OR A LINEAR DEPENDENCY AMONG MORE THAN TWO LATENT VARIABLES.
CHECK THE TECH4 OUTPUT FOR MORE INFORMATION. PROBLEM INVOLVING VARIABLE PGENDER.

There are no negative residual variances, correlations greater than 1, etc. Also, with parameter 52, this is a covariance between PGENDER and PAGE, and I cannot find a problem in the data. Can you make any recommendations about how to proceed? Thanks!
 Linda K. Muthen posted on Thursday, May 03, 2012 - 2:41 pm
Please send your output and license number to support@statmodel.com.
 EFried posted on Friday, May 11, 2012 - 4:47 am
I am facing the same problem, in a GMM with 5 measurement points, trying to run a model with 3 classes.


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 0.937D-11. PROBLEM INVOLVING PARAMETER 19.

TECH1 output looks normal, as does TECH4 (no PSI problems). The solution that MPLUS finds is implausible for this model, so I would need to explore what causes these issues.

Thank you
 Linda K. Muthen posted on Friday, May 11, 2012 - 6:19 am
Please send the output and your license number to support@statmodel.com.
 Susan Pe posted on Tuesday, August 28, 2012 - 4:58 pm
Hi I am doing 2level, different time points nested within individuals. I have 3 items. My Mplus commands include
cluster = firmid;
within = time;
Analysis:
Type = twolevel;
Model:
%within%
RD by exp newcar oldcar;
If I run CFA with 2 items, I get:
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
If I run CFA with 3 items, I get:
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 -0.562D-10.
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 5.
What is the problem? Thank you.
 Linda K. Muthen posted on Tuesday, August 28, 2012 - 5:19 pm
A factor with two indicators is not identified. Please send the output and your license number to support@statmodel.com.
 Danyel Arlyssa Vargas posted on Wednesday, October 31, 2012 - 6:49 pm
I have a similar problem. I have received the error: THE MODEL ESTIMATION TERMINATED NORMALLY
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 -0.148D-09. PROBLEM INVOLVING PARAMETER 149.

Iím running a multigroup model and this is the free model. This error did not occur for the constrained model. I believe this parameter is the disturbance covariance for two of the outcome variables. Would you please tell me how this affects my results?

Thank you very much!
 Linda K. Muthen posted on Thursday, November 01, 2012 - 7:27 am
Please send the output and your license number to support@statmodel.com.
 Danielle Roubinov posted on Friday, February 01, 2013 - 10:11 am
I'm running a basic regression model with an interaction. I'm also using auxiliary variables. I received the following error message when I tried to probe the interaction.

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 0.227D-10. PROBLEM INVOLVING PARAMETER 13. THIS IS MOST LIKELY DUE TO VARIABLE M11WORKR BEING DICHOTOMOUS BUT DECLARED AS CONTINUOUS.

Parameter 13 is the variance of my interaction term. M11WORKR is one of my auxiliary variables. I have it listed using the command auxiliary = (m). It is dichotomous, but I'm not sure where in my syntax it is declared as continuous.
 Linda K. Muthen posted on Friday, February 01, 2013 - 1:38 pm
Variables on the AUXILIARY list are treated as continuous variables. If a variable on this list is binary, it can prompt the message above because the mean and variance of a binary variable are not orthogonal. If this is the cause of the message, you can ignore it.

All variables are assumed to be continous unless they are put on a list like CATEGORICAL, CENSORED, COUNT etc.
 Danyel A.Vargas posted on Tuesday, October 08, 2013 - 2:07 pm
Hello,
I'm trying to conduct an LPA with four indicators.

I have received the error:
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 -0.102D-15. PROBLEM INVOLVING PARAMETER 1.

I have checked parameter 1 - it's on the NU matrix, but I'm not sure what this means. Can you please help?

Thanks,

Danyel
 Linda K. Muthen posted on Tuesday, October 08, 2013 - 5:20 pm
Please send the output and your license number to support@statmodel.com.
 Jami Gauthier posted on Friday, November 08, 2013 - 2:37 pm
Hello,

I'm running into a similar issue. When running a fairly simple path analytic model, I get this 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 IS0.202D-19.PROBLEM INVOLVING PARAMETER 99.

Parameter 99 is the variance of one of the variables in the model,and is rather large (as it should be). The model still runs, and standard errors are still estimated. Is this message ignorable? If not do you have suggestions for how to handle this issue?

Thanks!
Jami
 Linda K. Muthen posted on Friday, November 08, 2013 - 5:13 pm
Please send the output and your license number to support@statmodel.com.
 Bradford Barnhardt posted on Wednesday, November 20, 2013 - 7:59 pm
Hello,

I encountered a NPD matrix recently while fitting a structural model with composite variables. The offending factor had a small negative residual variance. I addressed the problem, on the advice of a colleague, with this syntax: factor@.00001; which eliminated the NDP, and appeared to have little or no effect on the fit of the model.

Looking through the user's manual, I can't figure out what exactly this syntax means, why it seems to serve this function, and whether it is appropriate. If it is not appropriate, can you suggest another means of addressing small negative residual variances?

Many thanks!
 Linda K. Muthen posted on Thursday, November 21, 2013 - 6:07 am
You are fixing the residual variance of factor to .00001. If a residual variance is estimated at a small negative non-significant value, many fix it to zero.
 Danyel A.Vargas posted on Thursday, February 06, 2014 - 11:34 am
Hello, I'm running a multigroup model with dummy variables as predictors. One of my groups is very small (n= 14). I received this error. Does this mean by results are not valid and I cannot perform a multigroup tests?

Thank you,

Danyel

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 -0.128D-17. PROBLEM INVOLVING PARAMETER 24.

THIS IS MOST LIKELY DUE TO HAVING MORE PARAMETERS THAN THE SAMPLE SIZE
IN ONE OF THE GROUPS.
 Linda K. Muthen posted on Thursday, February 06, 2014 - 1:27 pm
A sample size of 14 is too small for multiple group analysis.
 Danyel A.Vargas posted on Friday, February 07, 2014 - 7:26 am
Thank you very much.

Danyel
 Tunde Ogunfowora posted on Saturday, March 15, 2014 - 2:33 pm
Hi, I am having a similar problem as many of the previous posters in terms of the following error messages:

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 -0.149D-17. PROBLEM INVOLVING PARAMETER 54.
THE NONIDENTIFICATION IS MOST LIKELY DUE TO HAVING MORE PARAMETERS THAN THE NUMBER OF CLUSTERS. REDUCE THE NUMBER OF PARAMETERS."

I have checked the parameter and still can't make sense of the error message.

Thanks.
 Linda K. Muthen posted on Sunday, March 16, 2014 - 10:49 am
You most likely have more parameters than clusters. Check that. This warning is reminding you that independence of observations is at the cluster level. The effect of having more parameters than clusters may have an effect on the results.
 Tunde Ogunfowora posted on Sunday, March 16, 2014 - 11:31 am
Thank you for the reply. I do have about 63 parameters and 52 clusters. Does this affect only the parameter estimates (due to standard errors) or does this influence the chi squared tests and fit indices as well?

I recall my mplus instructor suggesting that one may use Bayes to get more accurate parameter estimates. In the case that I cannot get more clusters, is it acceptable to report the chi-squared results and fit indices from the ML solution and the parameter estimates from the Bayes solution?

Thanks!
 Linda K. Muthen posted on Monday, March 17, 2014 - 2:48 pm
This would effect that standard errors and fit statistics. The only way to know how much would be to do a Monte Carlo simulation study based on your analysis.

I don't think Bayes would make any difference.

You should report only results from one estimator.
 Tunde Ogunfowora posted on Monday, March 17, 2014 - 3:19 pm
Thank you. I ran the Bayes estimator but I did not get any error messages. The parameter estimates are similar to those obtained from the ML solution.
 Bengt O. Muthen posted on Tuesday, March 18, 2014 - 2:04 pm
Bayes estimation does not do this particular kind of check; it is ML-oriented. This doesn't mean that the Bayes solution is free of the potential problem listed in the ML run.
 Tunde Ogunfowora posted on Tuesday, March 18, 2014 - 2:36 pm
Thank you. It seems like my options are to increase the number of clusters in my data or reduce the number of parameters estimated (difficult in this case because of the theoretical model of interest).
 Bengt O. Muthen posted on Tuesday, March 18, 2014 - 2:53 pm
Or, you can do a little simulation in Mplus to check if you don't need to worry (that may very well be true).

This would be a suitable topic for a methods grad student to work on.
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: