04-07-2016 01:39 AM
I recently run GEE model of Falls rates to compare the effect of 12 months intervention -
The trial was of cluster design where I randomised villages to intervention group=1 and control group=0.
I used the command GENMOD as follow
PROC GENMOD DATA=today;
CLASS group (ref='0') multiple_fallers (ref='0') village_id /param=ref ;
MODEL number_falls= group /LINK=LOG DIST=NEGBIN TYPE3
Now I would like to test this in zero-inflated negative binomial model as people with history of multiple falls at baseline fell more in the intervention group than their counterparts in the control and it could be that the effect of the intervention will be different if I model it differently
I find it hard to find the codes which I need to use taking into account the adjustment for cluster and the variability of follow-up time between the clusters (villages); the mean follow-up in weeks varied greatly between villages.
Any help on the command I need to use and where to place these 4 variables would be greatly appreciated
group (intervention indicator),
FUweeks (time remain in the study)
fallers (history of falls) and
05-03-2016 03:48 PM
This note shows how to fit zero-inflated Poisson and negative binomial models using PROC NLMIXED (see footnote for ZINB log likelihood). To handle the clustering, you could add a random effect by adding a RANDOM statement.
05-05-2016 09:04 AM
I would strongly support @StatDave_sas's suggestion, as a GEE model really does not capture the results as well. Both PROC GENMOD and PROC GEE take into account data ordering when constructing solutions for the repeated nature, so that a different sorting by individual subject within village could result in a different GEE solution. The ZINB method in NLMIXED is much better at modeling the data you have.