Thank you both for your help. Matt, thank you for the long explanation, it is very informative ! Steve, in your code the Procedure_Type is a fixed effect, did you mean in the 3rd line of the code to write Procedure_Type, or should I add another random statement ? When I ran the code, the G matrix wasn't positive defined, could it be because I had 2 random statements with Subject_ID ? I also got weird OR estimated: OR DF L_CL U_CL Treatment1 Treatment3 <0.001 11 <0.001 <0.001 Treatment2 Treatment3 <0.001 11 <0.001 <0.001 Procedure1 Procedure2 >999.999 26 440.053 >999.999 Effect NUM DF DEN DF F P.V Treatment 2 11 39.32 <.0001 Procedure_Type 1 26 45.09 <.0001 Treatment*Procedure_ 2 26 Infty <.0001 When I ran Matt's code I have also received some wide confidence intervals for OR, however the F value did not converged to infinity (I did not have Procedure_Type as a fixed effect) Can these outputs be a result of what Matt has mentioned, the model balking due to the fact that some subjects had 2 surgeries with the same treatment, which are my repeated measures ? Some more outputs: Pearson Chi-Square / DF = 0.69 (which is good) covariance prarameter estimates Intercept Subject_ID 0.5162 1.4348 Treatment Subject_ID 1.14E-13 . Edit1: When I changed Steve's code and replaced the Subject_ID by Procedure_Type in the random statement, I have received the following error message: Estimation by quadrature is available only if the data can be processed by subjects. Make sure that all G-side RANDOM statements have SUBJECT= effects. If there are multiple SUBJECT= effects they need to form a containment hierarchy, e.g., SUBJECT=A, SUBJECT=A*B, SUBJECT=A(B), ... . Edit 2: When I removed the interaction term (which wasn't significant) I got slightly more reasonable results, the code was: proc glimmix data=data method=quad; class Subject_ID Treatment Procedure_Type; model Response = Treatment Procedure_Type / dist = binary link=logit oddsratio; random Treatment / subject=Subject type=vc; run; By the way, silly question maybe, how do I interpret an OR value of less than 1 ? I always find it hard (treatment 3 has many more 0's - non events, than treatments 1 and 2) For example if I get Treatment Treatment OR 1 3 0.009 Edit 3: I have "re-coded" the response, to get OR larger than 1, and I get an OR of 60 with CI of 6 to 600, due to small sample and perhaps other factors, how can I report such results ? I mean, it is significant (1 is not inside the CI), but I am not being very precise, am I ? One of the reasons for this is the sample size, in the control treatments, there are hardly any "successes". I have 15 samples in one of them, with 14 failures and 1 success, so the "1" is the problem.
... View more