Dear Statalist,
Hi, this is my first time Statalist forum. I am not familiar to GEE or not too much strong in statistics. I appreciate if you could consider my question.
I would like to ask about GEE for count models. I have repeated measure data. The purpose of analysis is to see whether there is difference in outcome (number of conditions) experienced between intervention (group=2) and control (group=1) groups. The sample size is n=103, with around equal numbers between groups. 'visit' is the number of measurement occasions (3 times). I consider 'ar1' as the suitable correlation structure (I comapred workng correlation and actual correlation). The outcome is the count of condition, but around 50-70% of respondent did not experience any condition and therefore there are many 0s at each occasion. Also, variance exceeds mean for 2-7 times at each occasion.
I checked the fit of count model using 'countfit' command (using cross-sectional analysis using baseline, not as repeated measure), and either zero-inflated negative binomial or negative binomial regression appeared to be most suitable.
Reading this earlier statalist forum http://www.stata.com/statalist/archi.../msg01187.html and other UCLA resource, I fitted below family and link combinations. I am not sure whether this makes sense, and also I am not sure whether the interpretation differs by the different combination?
xtgee outcome i.group i.visit i.group#i.visit, fam(nbinomial) link(nbinomial) i(ID) t(visit) corr(ar1) eform
xtgee outcome i.group i.visit i.group#i.visit, fam(nbinomial) link(log) i(ID) t(visit) corr(ar1) eform
xtgee outcome i.group i.visit i.group#i.visit, fam(poisson) link(log) i(ID) t(visit) corr(ar1) eform
The results were somewhat different:
IRR SE z p CI_lower CI_upper
First model: 2.group | .5280172 .1916958 -1.76 0.079 .2591915 1.075661
Second model: 2.group | .3958621 .1946298 -1.88 0.059 .1510225 1.037639
Tird model: 2.group | .3958621 .1667044 -2.20 0.028 .1734156 .9036489
I used a user written command 'qic' to see which model is best. While I do not think Poisson is suitable given overdispersion, the QIC value is the smallest for Poisson+log combination (QIC=271), indicating the better model fit. Other two specifications were similar (QIC = 280 for the First model and QIC=281 for the Second model).
Thank you so much for your time.
Best wishes,
Ayako
Hi, this is my first time Statalist forum. I am not familiar to GEE or not too much strong in statistics. I appreciate if you could consider my question.
I would like to ask about GEE for count models. I have repeated measure data. The purpose of analysis is to see whether there is difference in outcome (number of conditions) experienced between intervention (group=2) and control (group=1) groups. The sample size is n=103, with around equal numbers between groups. 'visit' is the number of measurement occasions (3 times). I consider 'ar1' as the suitable correlation structure (I comapred workng correlation and actual correlation). The outcome is the count of condition, but around 50-70% of respondent did not experience any condition and therefore there are many 0s at each occasion. Also, variance exceeds mean for 2-7 times at each occasion.
I checked the fit of count model using 'countfit' command (using cross-sectional analysis using baseline, not as repeated measure), and either zero-inflated negative binomial or negative binomial regression appeared to be most suitable.
Reading this earlier statalist forum http://www.stata.com/statalist/archi.../msg01187.html and other UCLA resource, I fitted below family and link combinations. I am not sure whether this makes sense, and also I am not sure whether the interpretation differs by the different combination?
xtgee outcome i.group i.visit i.group#i.visit, fam(nbinomial) link(nbinomial) i(ID) t(visit) corr(ar1) eform
xtgee outcome i.group i.visit i.group#i.visit, fam(nbinomial) link(log) i(ID) t(visit) corr(ar1) eform
xtgee outcome i.group i.visit i.group#i.visit, fam(poisson) link(log) i(ID) t(visit) corr(ar1) eform
The results were somewhat different:
IRR SE z p CI_lower CI_upper
First model: 2.group | .5280172 .1916958 -1.76 0.079 .2591915 1.075661
Second model: 2.group | .3958621 .1946298 -1.88 0.059 .1510225 1.037639
Tird model: 2.group | .3958621 .1667044 -2.20 0.028 .1734156 .9036489
I used a user written command 'qic' to see which model is best. While I do not think Poisson is suitable given overdispersion, the QIC value is the smallest for Poisson+log combination (QIC=271), indicating the better model fit. Other two specifications were similar (QIC = 280 for the First model and QIC=281 for the Second model).
Thank you so much for your time.
Best wishes,
Ayako
Comment