When I run a growth mixture model I get figures of 999.000 in the Std and StdYX columns. Everything else looks OK. And if I'd run the program without the command Output: Standardized; I'd never have come across these figures. Does this indicate a problem or am I worrying without cause??
bmuthen posted on Friday, November 29, 2002 - 6:51 am
999 means that it was not possible to compute the quantity. For example, if a variance is zero or negative, dividing by the square root of the variance in a standardization is not possble.
jbond posted on Tuesday, September 30, 2003 - 1:30 pm
Hello. I am following example 25.1 in the mplus users guide, except it is a 4 class model and there are two continuous outcome variables, each measured at 4 time points. The problem seems to be with identification (the iterations terminate due to an ill-conditioned fisher information matrix). I've tried several combinations even with using the exact analog of example 25.1, with only volume of consumption as the time varying variable (that is, removing the aavariable) but all seemed to have the same problem. The distributions of volume and # aa meetings (applying the USEOBSERVATIONS selection criteria) although skewed, still have quite a bit of variation. Is there anything obviously wrong in the specification or is it likely a numerical identification problem? Thanks much for any input,
The syntax used was (sorry if this is not the best way to provide the info):
TITLE: LCA For Number of AA Meetings;
DATA: FILE IS "G:\Trajectories\AA Careers\aacareer.dat";
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 0.513D-13.
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 21.
FINAL CLASS COUNTS AND PROPORTIONS OF TOTAL SAMPLE SIZE BASED ON ESTIMATED POSTERIOR PROBABILITIES
Class 1 242.57932 0.49006 Class 2 188.33055 0.38047 Class 3 37.29576 0.07534 Class 4 26.79438 0.05413
CLASSIFICATION OF INDIVIDUALS BASED ON THEIR MOST LIKELY CLASS MEMBERSHIP
Class Counts and Proportions
Class 1 245 0.49495 Class 2 192 0.38788 Class 3 32 0.06465 Class 4 26 0.05253
Average Class Probabilities by Class
1 2 3 4
Class 1 0.911 0.073 0.013 0.003 Class 2 0.091 0.882 0.024 0.003 Class 3 0.045 0.035 0.921 0.000 Class 4 0.014 0.001 0.000 0.985
AACAPDT1 WITH AACAPDT2 490.326 AACAPDT3 115.399 AACAPDT4 73.934
AACAPDT2 WITH AACAPDT3 1719.602 AACAPDT4 845.029
AACAPDT3 WITH AACAPDT4 438.563
VOLCAPT1 WITH VOLCAPT2 ********* VOLCAPT3 ********* VOLCAPT4 *********
VOLCAPT2 WITH VOLCAPT3 ********* VOLCAPT4 *********
bmuthen posted on Tuesday, September 30, 2003 - 2:35 pm
You are trying to do a mixture model with class-invariant covariance matrix and class-varying means. In line with Everitt-Hand's book referred to under the Mplus section with classic mixture examples, this is a difficult model to work with (unequal covariance matrices would be even harder). The LPA model in contrast, fixes the off-diagonal covariance matrix elements to zero. I notice that several of your variables are on a very high scale - your analysis might be simpler if you scale your variables down to variances in the 1-10 range.
I have been attempting to estimate a model in line with Example 7.26 (CFA with a non-parametric representation of a non-normal factor distribution) in the Mplus manual. I have tried estimating the model (which has 3 latent classes) exactly as with example 7.26, and I have also attempted to alter the factor means/intercepts for each of the latent classes using contrasts (e.g., -1, 0, 1) or by setting the first mean/intercept to "0" and freely estimating the rest. The models are estimating without error, but all of the models are returning standardized estimates of 999 for all loadings, means, variances, etc. I am aware that 999 = indeterminate, but I am not so sure that 999 = a problem with my models given your post on 11/02 in this thread:
"999 means that it was not possible to compute the quantity. For example, if a variance is zero or negative, dividing by the square root of the variance in a standardization is not possble."
Because Example 7.26 specifies the variance of the latent factor to be 0 in the overall model, does this mean that all models based on this example will not have estimable standardized coefficients? If, alternatively, all of the 999 values are indicative of a problem with my model, are my fit estimates still reliable/interpretable (e.g., BIC) given that standardized estimates are not estimable?