I am extremely new to this, so bare with me:
I am trying to run a generalized estimating equation analysis to assess the association between television viewing in childhood (coded categorically as "tv": <=0.5 hours, 1 hour and =>2 hours/day) and change in body mass index (bmisds) over time. I have repeated measurements on individuals at ages 4, 11 and 13 years (time= 1, 2, 3) and would like to adjust for some confounders associated with obesity (the rest of the covariates).
So far, I have been using the following code in SAS:
proc genmod data=all;
class tv(ref="2.00") M_ID time(ref="1");
model bmisds=time tv time*tv mom_gc brfed preec gender_merged
firstborn sectioyes mat_ed activityhs4 GA_weeks MBMIsvkon1
maternal_age_birth weightsds_1;
repeated subject=M_ID/type=exch corrw;
run;
My questions are:
1) How do I determine the best correlation matrix to use? I have been playing around with them and some of them completely alter statistical significance of the model. Based on what I have been reading online, an unstructured or exchangeable correlation matrix may fit my data best. Do you agree?
2) How would I interpret a coefficient that is significant for the interaction between television viewing (=>2 hours/day) and time (13 years) but is not significant for television viewing alone? Can I say that watching higher amounts of television is associated with a higher body mass index at 13 years only?
