PROC GENMOD. Error with GEE

[Demographer]

Hi,

I want to model the probability of being obese with a set a risk factors. Data are from a longitudinal survey (so there are many observations for one single individuals). I thus use GEE with proc GENMOD, which I think is the appropriate model for this kind of situation. However, I have this error:

NOTE: Class levels for some variables were not printed due to excessive size.

NOTE: PROC GENMOD is modeling the probability that obe1='1'.

NOTE: Algorithm converged.

ERROR: Error in computing the variance function.

ERROR: Error in parameter estimate covariance computation.

ERROR: Error in estimation routine.

I identified the problematic variable, which is cntry (for country), but the error is still there. Here is the code for this simplified model:

proc genmod data=work.data1 descending ;
  class  smoking(ref='0') obesity(ref='0') ah(ref='0') edu3(ref='2') depression(ref='0') vig_pa(ref='0') cntry(ref='DE') noid sex;
   model  obe1=/*obesity sex age_num age_num*age_num edu3 smoking ah depression vig_pa*/ cntry /*cntry*sex dur*//  dist=bin link=logit;
   repeated subject=noid / type=exch;
   weight pond;
run;

Analysis Of GEE Parameter Estimates
Empirical Standard Error Estimates
Parameter		Estimate	Standard Error	95% Confidence Limits		Z	Pr > |Z|
Intercept		-1.6629	.	.	.	.	.
cntry	AT	0.5749	.	.	.	.	.
cntry	BE	0.4967	.	.	.	.	.
cntry	CZ	0.9894	.	.	.	.	.
cntry	DK	0.7170	.	.	.	.	.
cntry	EE	0.8038	.	.	.	.	.
cntry	ES	0.4774	.	.	.	.	.
cntry	FR	13.7220	.	.	.	.	.
cntry	GR	0.2397	.	.	.	.	.
cntry	IT	0.7010	.	.	.	.	.
cntry	NT	0.0756	.	.	.	.	.
cntry	SE	1790526	.	.	.	.	.
cntry	SI	0.6285	.	.	.	.	.
cntry	DE	0.0000	0.0000	0.0000	0.0000	.	.

And this is the crosstab for the obe1*cntry. I don’t see any category that may be problematic.

proc freq data=work.data1;
table obe1*cntry /nocol norow nopercent;
weight pond;
run;

Table of obe1 by cntry
obe1	cntry
	AT	BE	CZ	DE	DK	EE	ES	FR	GR	IT	NT	SE	SI	Total
0	719.568	836.852	508.02	2711.3	275.363	97.5724	3317.06	3159.86	609.726	3309.08	739.526	351.252	139.237	16774.4
1	180.909	197.059	220.64	744.693	41.8123	41.3259	913.276	736.899	147.012	717.024	146.01	62.5714	49.4657	4198.7
Total	900.477	1033.91	728.66	3456	317.176	138.898	4230.34	3896.76	756.738	4026.1	885.536	413.823	188.702	20973.1

Note 1: the model works when I remove the weight statement. The range of the weight is from 0.006735 to 46.43589.

Note 2: the model works when I use another dependent variable (and using the weight statement).

[StatDave]

In any modeling procedure, you should not specify variables in the CLASS statement that are not used in other statements in the procedure. Doing so can result in additional observations being ignored if some are missing on these variables. In binary response models, decreasing the number of observations used can easily cause the data to become too sparse which can cause results like what you show. 

Also, weights are typically not needed for logistic models like this. If the weights you are using are sampling weights, then you should be using PROC SURVEYLOGISTIC, not GENMOD. GENMOD does not have the variance estimators needed to do a proper survey data analysis.

[Demographer]

It's sampling weight. Is it possible with PROC SURVEYLOGISTIC to use the GEE method (some observations are for a same individuals,but different years, so there is a correlation that needs to be accounted)?

[StatDave · Accepted solution]

I believe you can use the CLUSTER statement to deal with the clustering in your data.