PROC GLMMIX did not converge even with only one binary covariate

[lichee]

Hi SAS Gurus,

I am trying to run HGLM with PROC GLIMMIX using the data of around six million of individuals across states with state effect. Even though I started with the simplest model with only one covariate included only, the model did not converge. Any of you have any insight or suggestion?

Below is the log:

Any states are excluded when they had extremely low counts of individuals with the outcome variable flag_success_discharge=1.

Thank you,

Lichee

[PaigeMiller]

Plenty of discussion about GLIMMIX not converging here in the forums. One example: https://communities.sas.com/t5/Statistical-Procedures/Glimmix-did-not-converge/td-p/387883 You should search for other discussions.

[jiltao]

What happens if you add METHOD=LAPLACE option in the PROC GLIMMIX statement?

Also, you might want to spell out: dist=binary in the MODEL statement.

Thanks,

Jill

[lichee]

Thank you! 

I added METHOD=LAPLACE, and now the initial estimates did not yield a valid objective function.

[Ksharp]

If I was right, your response variable "flag_success_discharge" is continuous variable,
NOT a category variable. Use PROC FREQ to check its levels firstly.

[lichee]

Thank you! The dependent variable is dichotomous. Below is the crosstab:

[Ksharp]

Try specify the distribution :
/d=b
====>
/dist=binary

[lichee]

Yes, I re-ran the code with dist=binary specified and got the same message "Did not converge"

[jiltao]

Try using PARMS statement to enter you own starting values. This might take several trial and error.

If that does not help, can you send in your data?

Thanks,

Jill

[SteveDenham · Accepted solution]

Try increasing the number of iterations in an NLOPTIONS statement. Binary data often require more than the default 20 iterations.

SteveDenham

[Ksharp]

@Rick_SAS 's blog could give you a clue.

https://blogs.sas.com/content/iml/2019/04/03/g-matrix-is-not-positive-definite.html