GEE logor=exch vs. type=exch

[mcmag]

I am trying to determine why I received the following errors when I ran GEE on my data using logor=exch and why I did not receive those errors when I ran my data with type=exch:

Error in log odds ratio regression computation.

Error in parameter estimate covariance computation.

Error in estimation routine.

The respective code is listed below:

proc genmod data=both;

class inhalyn (ref="0") grades (ref="3") wave uid / param=ref;

model inhalyn=grades / dist=bin link=logit;

repeated subject=uid / withinsubject=wave logor=exch;

estimate "1 vs 3" grades 1 0 / exp;

estimate "2 vs 3" grades 0 1 / exp;

run;

proc genmod data=both;

class inhalyn (ref="0") grades (ref="3") wave uid / param=ref;

model inhalyn=grades / dist=bin link=logit;

repeated subject=uid / withinsubject=wave type=exch;

estimate "1 vs 3" grades 1 0 / exp;

estimate "2 vs 3" grades 0 1 / exp;

run;

SAS was able to give me estimates and ORs when using type=exch. Inhalyn is a binary variable and grades is a 3-level categorical variable. I used the same code for other 3-level categorical variables and was able to get estimates and ORs when using logor=exch, so I'm trying to figure out what's different about this variable/scenario that would require type=exch instead of logor=exch.

Your help is much appreciated!

[SteveDenham · Accepted solution]

This sounds like a job for Tech Support.  I have a guess, but it is probably wrong.  Any possibility that an OR is pathological for one of the waves, such that the log odds ratios are not exchangeable, but the correlations are?  By pathological, I mean a zero or some value greater than the machine arithmetic recognizes? This would result in errors when the alternating method is applied.

Steve Denham

[mcmag]

I think you're right!

Thanks so much for your help.