Hello- I noted the excellent exchange of info. back in 2012 on how to obtain accurate estimates of LSMEANS and standard errors using PROC GLIMMIX when the response variable has been log-transformed. Steve- thanks for the suggestion. However, I tried this approach but found no improvement in my residuals (non-constant variance problem). In a nutshell, here is my conceptual linear model: ln (x...a continuous resp.variable) = f (y, z)...where y is continuous and z is a 3-level class variable. So the GLIMMIX program statement with the "fix" for back-transforming the LSMEANS would be:   PROC GLIMMIX DATA=example plots=all ; CLASS z; MODEL x= y  z/SOLUTION LINK=LOG; LSMEANS z/CL DIFF ILINK; RUN;   The resulting residuals display no improvement and are essentially the same as the output from a non-transformed model without the link=log statement. I am doubtless making a stupid mistake! I would appreciate any help you can offer. Thanks. ... View more