I want to estimate risk for a certain outcome. For example age on disease X.
I coded dummy variables from ordinal and nominal data (age, gender, etc) to create 2x2 tables. So, I turn my three age categories into three 2x2 tables. Lets say, I have an age category "young" vs "really old" and a "middle aged" vs "really old" and then I want to calculate the RR for each of these to test if there is an increased risk of being young vs being really old and contracting disease X as well as for being middle aged vs really old and contracting disease X.
So I ran a logistic regression (proc genmod) to get my RR's and I compared my results later with the proc freq method and my results loose statistical significance in the logistic regression because my CI are much wider. (Plus my actual RR is more extreme as well, meaning if the RR is high in log reg, then it is even higher in proc freq.)
I assume this is because each procedure uses a different calculation but I don't understand the statistics behind it well enough to understand the reaction. And which test statistic is more believable?
Thanks so much for any help!
here the two test statistics:
proc freq data=test;
proc genmod data=test;
model disease = age_young/ dist = poisson link = log;
estimate 'Beta' age_young -1 1/ exp;