Anonymous posted on Monday, November 07, 2005 - 10:31 am
I have two questions regarding using Mplus to do discrete time survival analysis.
Question 1 on Mplus fixing a parameter in discrete time survival analysis estimation
I am doing a discrete time survival analysis (DTSA) with only one class. The model runs fine using different covariates: the model converges and the parameters and standard errors all look reasonable. However, for a run using one particular covariate I got this warning message:
ONE OR MORE PARAMETERS WERE FIXED TO AVOID SINGULARITY OF THE
INFORMATION MATRIX. THE SINGULARITY IS MOST LIKELY BECAUSE THE
MODEL IS NOT IDENTIFIED, OR BECAUSE OF EMPTY CELLS IN THE JOINT
DISTRIBUTION OF THE CATEGORICAL VARIABLES IN THE MODEL.
THE FOLLOWING PARAMETERS WERE FIXED:
Despite this warning, Mplus reports that the model estimation terminated normally.
The parameter that is fixed is the threshold for the last period:
VG8$1 32.166 0.000 0.000
I also noticed that in the summary of categorical data proportion section I got:
VG8 Category 1 1.000 Category 2 0.000
It means there is no occurrence of the event at all in period 8.
All the other parameters have reasonable magnitudes and signs, and the standard errors are also reasonable.
I am wondering whether I can trust and report the results for the other parameters in the model.
I found the following exchange in Mplus Discussion. It is under Multinomial Logistic but I think it might apply to my discrete time survival analysis too.
Anonymous posted on Thursday, June 24, 2004 - 3:35 pm
I have a gmm with a dichotomous predictor of class membership (gender). There are no females in one of my classes and the regression coefficient associating this class with gender is 70.8. I also get an error message stating that this coefficient was fixed to avoid singularity. What should I do in this case?
bmuthen posted on Friday, June 25, 2004 - 9:09 am
You can report this solution. The large fixed reg coeff simply means that the probability is 1 for being in this class when the dichotomous predictor is 1 as opposed to 0. The value 70.8 is arbitrary - any large value, say greater than 15 (it depends on the other coefficients' sizes), suffices to give probability 1.
Question 2 on condition number
The DTSA I did was done using numerical integration because I had a latent variable as a covariate. As I said above, the model converges and the parameters and standard errors all look fine. And I used “start=20 2;” to check if the estimation has indeed converged to a maximum. The only problem is that in all my runs, the condition number ranges from 0.326E-07 to 0.660E-09. According to Mplus User’s Guide Version 3 p.296, when numerical integration is involved, a condition number less than 1.0E-06 may indicate non-identification. Should I be concerned with my low condition number? Will sharpening some of the convergence criteria or increase the number of integration points increase the condition number?
bmuthen posted on Sunday, November 13, 2005 - 10:33 am
Q1. Yes you can trust this solution. You have no new events in the last period so the threshold parameter for that last event history indicator is infinity (a very large threshold essentially giving a zero probability). Mplus automatically fixes the large threshold.
Q2. It is hard to say if you should be concerned, The condition number might be obtaining a small value due to large thresholds in which case this is harmless. You can also check if your SEs are uncharacteristically large which is an indication of non-identifiability. Yes, you can also try sharpening the convergence criterion (mconvergence) and increase the integration points.