09-15-2016 11:13 AM
We are utilizing a complex survey (NHATS) with stratification, clustering, and weighting. Our final models are Poisson with GEE. We understand that SURVEY procedures (SURVEYMEANS, SURVEYLOGISTIC, etc.) are necessary to use in order to appropriately account for this sample design and obtain accurate standard errors. We also know that PROC GENMOD does not have a direct corrollary with a SAS SURVEY statement.
We have had the suggestion to use SUDAAN, STATA, or some other program, but have used SAS for the rest of the analysis and would like to continue to do so. We found what appeared to be a very helpful article outlining with sample code how to use a non-survey procedure (PROC GENMOD) to obtain valid point estimates and extended programming to obtain valid standard errors for our model's parameter estimates (Poisson Regressions for Complex Surveys--attached here) and have followed the steps outlined therein carefully. Unfortunately Step 2 will not run due to the GEE portion of our model (if we take out the GEE coding lines starting with "repeated subjects" then the code runs, but with them included we get errors).
Here is one of our original models as an example (we did change to proc genmod data=jkweights appropriately when attempting to run the suggested code):
proc genmod data = analysis descending;
class SPID frailty time/param=ref ref=first;
model drive = frailty time frailty*time/type3 dist = poisson link = log;
repeated subject = SPID/ type=cs;
estimate "frailty t1" frailty 1 / exp;
estimate "frailty t2" frailty 1 frailty*time 1 0 0/ exp;
estimate "frailty t3" frailty 1 frailty*time 0 1 0/ exp;
estimate "frailty t4" frailty 1 frailty*time 0 0 1/ exp;
In comparison, here is the Poisson model (without GEE) that is presented in the paper we are attaching:
proc genmod data=jkweights;
class sex race marital private education;
model visits = sex race marital private education / dist=poisson;
Is anybody aware of how to take the information presented in this article and apply it to a Poisson model including GEE? Any help is greatly appreciated.
09-15-2016 11:36 AM
If you have access to SAS/STAT14.1, try porting the code to PROC GEE. There is an example (43.4 GEE for Binary Data with Logit Link Function) here:
that with a change in distribution and link would be a start, I think.
If that works, you may then want to port again to PROC GLIMMIX, using this exampe, with appropriate changes, for survey data:
09-15-2016 03:07 PM
Thank you, Steve, for your quick and insightful reply.
I have checked and I do not currently have access to SAS/STAT14.1 I have put in a request with my IT department to see if installing it will be possible.
In the meantime, if there are any other suggestions in the SAS community, I am running SAS version 9.4 at the moment.