Programming the statistical procedures from SAS

convergence in glimmix

Reply
N/A
Posts: 0

convergence in glimmix

In a several cases where I used GLIMMIX the iteration process seems to have converged, judged by monitoring the estimates using ITDETAILS, but in the end GLIMMIX returns saying “Did not converge”. See the example below:

Iteration Restarts Subiterations ObjectiveFunction Change MaxGradient Parm1 Gradient1

0 0 5 6184.6039075 1.02270787 0.00022 0.06671 0.000220
1 0 4 6413.5521306 0.17842137 0.001003 0.05578 -0.00100
2 0 2 6571.6354826 0.11029232 0.000097 0.05660 0.000097
3 0 2 6704.3698719 0.09277192 0.00002 0.05604 0.000020
4 0 2 6844.8990306 0.08137020 1.185E-7 0.05613 1.185E-7
5 0 1 6985.4373867 0.07196559 1.395E-6 0.05610 1.395E-6
6 0 1 7127.3675232 0.06453834 2.23E-6 0.05611 2.23E-6
7 0 1 7270.0181151 0.05841590 8.605E-7 0.05611 8.605E-7
8 0 1 7413.3989255 0.05332177 3.84E-7 0.05611 3.84E-7
9 0 1 7557.3827666 0.04901085 1.888E-6 0.05611 1.888E-6
10 0 1 7701.9044939 0.04532079 3.827E-6 0.05611 -3.83E-6
11 0 1 7846.8985916 0.04212767 0.000123 0.05611 0.000123
12 0 1 7992.3120998 0.03933864 1.603E-6 0.05611 1.603E-6
13 0 0 8138.0980503 0.03688319 9.126E-6 0.05611 -9.13E-6
14 0 0 8284.2161647 0.03470529 3.439E-6 0.05611 -3.44E-6
15 0 0 8430.6311405 0.03276120 4.144E-6 0.05611 -4.14E-6
16 0 0 8577.3119572 0.03101547 4.259E-6 0.05611 -4.26E-6
17 0 0 1.340781E154 2.00000000 . 0.05611 .

WARNING: Pseudo-likelihood update fails in outer iteration 18.
NOTE: Did not converge.

I have tried changing the convergence criterion, but need to relax it to PCONV=0.1 for GLIMMIX to finish the optimization ‘normally’ (i.e. without the “Did not converge” notice), but then the actually estimate it returns is 0.5604.

Two questions that I wonder anyone can help me with:
- How come that requiring PCONV=0.01 does not result in stopping the optimization process in a normal way even though the parameter estimates do not change any more?
- What do the columns “Change”, “Max Gradient” and “Gradient1” actually mean?
- What other options exist to tweak convergence?

Regards,
Freddy Fikse
Ask a Question
Discussion stats
  • 0 replies
  • 188 views
  • 0 likes
  • 1 in conversation