The question is about using contrasts in regression analysis (here, Poisson regression with robust error variance). We have devided the participants in our sample into 6 groups, according to their body mass index at baseline (3-category variable: normal, overweight, obese) and the metabolic status at baseline. data d.mho; set d.mho; /*2737*/ if bmi_cat2 = 1 and Met_abnorm = 0 then gp=1; /*metabolically healthy normal weight*/ if bmi_cat2 = 1 and Met_abnorm = 1 then gp=2; /*metabolically unhealthy normal weight*/ if bmi_cat2 = 2 and Met_abnorm = 0 then gp=3; /*metabolically healthy overweight*/ if bmi_cat2 = 2 and Met_abnorm = 1 then gp=4; /*metabolically unhealthy overweight*/ if bmi_cat2 = 3 and Met_abnorm = 0 then gp=5; /*metabolically healthy obese*/ if bmi_cat2 = 3 and Met_abnorm = 1 then gp=6; /*metabolically unhealthy obese*/ run; We choose the "metabolically healthy normal weight" people as the reference group. In this case, is it possible to calculate the contrasts only between group 5 and 6 (metabolically healthy obese vs metabolically unhealthy obese), while keeping 'group 1' as the reference group: proc genmod data=&data descending; class id_vol &adjust_class &class(ref=&ref) ; model &outcome = &class &adjust_class &adjust_contin / type3 dist = poisson link = log; repeated subject= id_vol / type=unstr modelse; ods output GEEEmpPEst=modcat ; estimate 'MHO vs MUO' &class 0 0 0 0 1 -1/ exp; run; I've also seen examples where the contrast coefficients are not necessarily integers (but couldn't understand): estimate 'P contrast' &class -1.5 -0.5 0.5 1.5/ exp; Thanks for your help!