Proc Glimmix random intercept large estimates

[yuchinher]

Hi,

I am trying to model a three level logistic regression using PROC GLIMMIX. When I first added only my level 3 clustering variable into the model, the estimates and odds ratio I got look fine and reasonable. However, when I added both the level 3 and level 2 clustering variable to the model, the estimates and odds ratio I got became really big that are probably unreal (e.g., odds ratio of 40 to odds ratio of 300). I wonder if anyone knows the reasons behind and how I can fix it? Thanks for the help! 

The code I used for having both level 3 and level 2 clusters, which gave me really large parameters:

proc glimmix data=new noclprint method=laplace;

class clustervar_L3 clustervar_L2;

model event(event='1')=dur dur2 income incouple rent houseprice age sex

/solution cl link=logit dist=binary 

ddfm=bw oddsratio(DIFF=LAST LABEL);

random intercept / solution type = un subject= clustervar_L3;

random intercept / solution type = un subject= clustervar_L2(clustervar_L3);

covtest / wald;

run;

[PaigeMiller]

---

@yuchinher wrote:

I am trying to model a three level logistic regression using PROC GLIMMIX. When I first added only my level 3 clustering variable into the model, the estimates and odds ratio I got look fine and reasonable. However, when I added both the level 3 and level 2 clustering variable to the model, the estimates and odds ratio I got became really big that are probably unreal (e.g., odds ratio of 40 to odds ratio of 300).

---

My first thought is that adding additional class variables into the model causes quasi-complete separation, or maybe even complete separation.