Hi Hima, I think what you are looking for is a contrast of (all active treatments) vs. placebo. Is that correct? I think you would want to use an Estimate statement that would allow you to build a contrast of (1 1 1 -3) / divisor = 3 for treatment, but also include code for each covariate level. It would be a very large statement, and you'd want to keep it a single statement so that you can adjust your confidence intervals appropriately. I think something like this might work. I am only doing one covariate at a time, as that looks like what you intended. If you actually want to put all the covariates in at the same time with all their interactions, this is probably going to get too complicated. class subj trt age_grp visit; model chg = baselinevalue trt|visit|age_grp / ddfm = kr cl; and then either estimate 'Any Treatment vs. Placebo for Age < 80' trt 1 1 1 -3 trt*age_grp [1, 1 1] [1, 2 1] [1, 3 1] [-3, 4 1], 'Any Treatment vs. Placebo for Age >= 80' trt 1 1 1 -3 trt*age_grp [1, 1 2] [1, 2 2] [1, 3 2] [-3, 4 2] / divisor = 3 e adjust = scheffe; or estimate 'Any Treatment vs. Placebo for Age < 80' trt 1 1 1 -3 trt*age_grp 1 0 1 0 1 0 -3 0, 'Any Treatment vs. Placebo for Age >= 80' trt 1 1 1 -3 trt*age_grp 0 1 0 1 0 1 0 -3 / divisor = 3 e adjust = scheffe; Unless I made an error, I think those two estimate statements are equivalent. One uses positional syntax, and the other uses non-positional. For age_grp, positional (the second) looks cleaner. But realize if you have a covariate with 5 levels, that string of digits will have 20 more zeros sandwiched in there in various places. (personally I like seeing that). You can look up all this in the SAS help where they have a simpler example. The option e in the statement is very important as it will output the design matrix for you. This will let you check that you put all the 1s and 0s in the correct places for what you want. Note, they'll be divided by the divisor. The syntax of both are dependent on a few things: 1. the order of your class statement will change the order of your interaction terms, regardless of how you order them in the estimate statement; 2. the order of your variables sort will impact the order you need in the estimate statement. I assumed your treatments will naturally sort 1, 2, 3, placebo. If they don't, you'll have to mix it up. Note: your procedure for Eff_1 does not adjust for multiple comparisons. Some researchers don't adjust, but I think they are in the majority. I'd double check with your head statistician. If you need to adjust, you want to use diff = control adjust = dunnett since you are only comparing each level to a control. Steve, does this seem right? I don't need to add in the interaction term of all three, right, because we're summing over all the visit levels anyway? Best, Michael Message was edited by: Michael Cooney (added note about Eff_1)
... View more
