I am currently analyzing some DV which represents aproportion and a fractional logit model seems most appropriate. In addition to this, my data has a multilevel structure, implying that I include a random intercept (hhic, legal and age are measures at the group-level agency).
This is the syntax:
proc glimmix data=consultations empirical;
class agency consultationtype(ref='1') complexitytext(ref="1") com_dummy(ref="0") format targetgroup;
model p_nonregulated=legal age hhic hhic*legal legal*age consultationtype complexitytext format duration l_mob_total
/dist=binomial link=logit solution;
random _residual_/subject=agency type=vc;
covtest/wald;
output out=predp pred(ilink)=p;
run;
My question concerns the interpretation of the coefficients, which I find difficult to interpret (in terms of change in the independent variable corresponds with variation in the dependent variable). More specifically, I have some coefficients which are pretty large (<3.00), although standard errors seems normal. My hunch is that this related to the scaling of the dependent and independent variable. Namely, the estimates where I have large coefficients concern covariates that range between 0 and 1 (HHIC is a proportion), while the esitmate (fixed effect) concerns how one unit change in the dependent variable shapes the outcome variable.
See the output attached.
Jan