Elaine Walsh posted on Saturday, September 23, 2006 - 9:16 pm
Hello, I am working on a mixture model and attempting to identify trajectories related to a specific behavior measured at 6 time points. The N is 351 and there is no missing data. I am having difficulty with start values and receive the following message when I attempt to run a model with 4 classes:
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 0.120D-15.
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 17.
I would appreciate some assistance understanding the "ILL-CONDITIONED FISHER INFORMATION MATRIX" reference and would also like to know the best way to select start values for classes.
I would not use my own start values, but let the program generate those and do its random starts perturbation of them. If the default STARTS = 10 2 is not sufficient, I would increase it (say to 50 5 and 100 10, etc). The ILL-CONDITIONED message says that you have not found an acceptable solution - not a proper maximum of the likelihood.
Thank you for your response. I have tried a couple of different things and cannot get the syntax to run. I am using version 2.12. Is there a different way to state this for the version I am using? Thank you.
To eliminate the possibility that you have a non-identified model, send your input, output, data, and license number to email@example.com. But version 2.12 is far from the strength of the current version 4.1 (including automatic starting values and random perturbations of them since version 3) - you really should upgrade to 4.1.
Elaine Walsh posted on Thursday, October 12, 2006 - 10:04 pm
We purchased the new version and this now runs fine. Thank you for your help--I will check back if we run into other problems.
I am running a GMM model in which the 3-class model fits the data best. the largest class is the best-adjusted class, and i would like this class to be the reference group (class 3) when looking at the influence of the predictors. i attempted to include start values to accomplish this, but for some reason, the well-adjusted class is appearing as class 2 rather than class 3. here's the syntax i'm using:
%c#1% [i*2 s*-.03] ;
%c#2% [i*1.3 s*.4] ;
And here are the estimated i and s values for each group (i'm delineating what i WANT each class to represent - the parentheticals indicate the percent of the sample in each class):
class 1 (14%): i=2.52, s=-.028 class 2 (10%): i=1.37, s=.45 class 3 (75%): i=1.44, s=.03
for the post above, i implied but did not ask my question, which is whether mplus requires that the largest class NOT be the last class. if the largest class can be the last/final class, how is my syntax mis-specified?
Thanks bengt! i have one more question. In my GMM model, I include several predictors of class membership, but i also want to include 4 distal outcomes (all latent variables). Is it possible to do all of this in a single model? I saw example 8.6 in the user's manual, but i wasn't sure if this could be adapted for distal outcome variables that are latent.
Yes, that is possible. But if you apply this directly, you will end up with the assumption that your 4 latent distal vbles are uncorrelated given the latent class vble, which might not be what you want.
Dena Pastor posted on Thursday, October 14, 2010 - 1:21 pm
I some questions have about these starting values used during the initial stage optimizations: -Are randomly generated starting values produced for all parameters with the exception of variances and covariances? -Are all variances given starting values of 0.05, all covariances a value of 0? Iím assuming these starting values can be altered by providing user-supplied starting values and that there is no way to have Mplus generate random values for the starting values of variances and covariances, correct? -Are random starting values used for the class weights? -When the default settings are used and user-supplied starting values are absent are the randomly generated starting values are pulled from a uniform distribution centered at 0 and extending 5 units (so this distribution has a minimum value of -5 and a positive value of 5)? -Am I correct in thinking that when user-supplied starting values are provided that the randomly generated starting values are pulled from a uniform distribution centered at user-supplied starting values and extending 5 units? -Am I correct in thinking that this range of the uniform distribution can be altered using the STSCALE option (e.g., in the absence of user-supplied starting values and STSCALE=2, the starting values will be pulled from a uniform distribution ranging from -2 to 2)? -How can I go about obtaining the starting values that are being used by Mplus?
I am running an LCA and am attempting to compare a 3 class solution to the 2 class solution, but my 3 class model is not converging. I am getting the following warning messages even after increasing Starts to 500 50 (which took 3hrs 41 minutes to run) using MPlus 6.11. can you help explain these error messages? Thanks.
"Unperturbed starting value run did not converge.
1 perturbed starting value run(s) did not converge.
THE LOGLIKELIHOOD DECREASED IN THE LAST EM ITERATION. CHANGE YOUR MODEL AND/OR STARTING VALUES.
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.
THE MODEL ESTIMATION DID NOT TERMINATE NORMALLY DUE TO AN ERROR IN THE COMPUTATION. CHANGE YOUR MODEL AND/OR STARTING VALUES."
Stine Hoj posted on Tuesday, April 01, 2014 - 3:40 pm
I am running a series of GMMs in order to identify the optimal number of classes.
When attempting to fit a 5-class GMM, the model would not converge when using only random starts (STARTS=1000 250) or when using the growth factor means from an LCGA as starting values.
When I used the growth factor means from the 4-class GMM as starting values for 4 of the classes, the model estimation terminated normally. However, I am unsure of whether this is a suitable approach to selecting starting values?
I'd like to understand more about the random perturbation of start values. I've seen the technical document (https://www.statmodel.com/download/Starts.pdf) and the equation for the perturbed starting values. Three questions please: 1) What is the possible range of the scale variable? 2) What the base scale of the parameters? If I want to get random start values for observed categorical indicator variables in a latent class model (the rho parameters), do they range from 0-1 or are they on the logit scale? What about the class probabilities (the gamma parameters)? 3) What are the default starting values for the rho parameters?
1) Stscale can be any positive number. 2) It alternates. For even seed numbers it is on the probability scale and for odd seed numbers it is on logit scale. See the top of page 2. 3) It is data driven - for binary items and say 2 classes, in class one it is log(p0/p1)+1 and in class two it is log(p0/p1)-1, where p0 and p1 are the observed proportions for the binary variable. You can see the starting values in tech1.
benedetta posted on Monday, November 30, 2015 - 2:00 am
I am running a Diggle-Kenward selection model to deal with MNAR in longitudinal data in wide format and also assess the effect of exposure (x) on outcomes. The syntax is as follows:
When I specify the model as above I get the message 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 0.431D-10.
I have also tried changing starts value (i.e. to 50 5 and 100 10, etc).
If I don't regress the slope on the x, the model converges and standard errors are estimated. I can't figure out what is wrong with the model where both intercept and slope are regressed on x.
I tried to test a simple interaction/moderation model by using the example 5.13 in the Manuel. The codes are as follow:
VARIABLE: NAME ARE EP1-EP4 ER1-ER4 AG1-AG18 EN1-EN10 AC; USEVARIABLES ARE EP1-ER4 AG7-AG9 EN1-EN10 AC; ANALYSIS: TYPE = RANDOM; ALGORITHM = INTEGRATION; MODEL: F1 by EP1-EP4 ER1-ER4; F2 by AG7-AG9; F2 by AG7 AG8 (1); F3 by EN1-EN10; AC ON F3 F2; F3 ON F2 F1; F1xF2 | F1 XWITH F2; F3 ON F1xF2; OUTPUT: sampstat stdyx tech1 tech4;
However, I received the following message:
THE ESTIMATED COVARIANCE MATRIX COULD NOT BE INVERTED. COMPUTATION COULD NOT BE COMPLETED IN ITERATION 235. CHANGE YOUR MODEL AND/OR STARTING VALUES.
Dear Dr. Muthen, I am trying to rescore a cross-sectional database according to a classification already obtained in the past. I found a lot of information about latent transition analysis when rescoring the same individuals, but in my case I would like to rescore on a whole new sample. There are my questions :
-Is it ok to use start values from my base latent class model to rescore my new sample ? -Is there a more precise term than rescoring for this king of analysis? -Also, can I use the 3-step option to compare the distributions of my new sample vs the distribution of my base sample?