PROC MIXED: provide Box-Cox Transformation for general linear mixed model
If you are going to add Box-Cox (a good idea), I think you should also add the "folded exponential" transformation for proportions, since Box-Cox is not well suited for proportions. Good reference, with sas macro code for the folded exponential with proc mixed: Piepho, H.P. 2003. The folded exponential transformation for proportions. The Statistician 52(4): 575-589.
A good reference for a macro for Box-Cox that utilizes proc mixed is: Piepho, H.P. 2009. Agronomy Journal 101(4):865-869.
I could understand adding these to PROC GLIMMIX, but as MIXED currently does not support any transformations, I would tend to keep them out of the MIX (so to speak).
I think we are considering different things. GLIMMIX has link functions (where the function of the expected value is modeled in relation to X and Z), but MIXED and GLIMMIX could also have transformations of y (where the expected value of the function of y is modeled in relation to X and Z). I think SAS is considering a generalization of the boxcox(y) transformation in TRANSREG (see example 6 of the procedure), which is for fixed effects, to an approach for a mixed-effects model. One can do the transformation outside of MIXED, of course, but I am guessing SAS is thinking about embedding the functionality within the procedure (where profile ML is used to find the exponent parameter). The above references generalize the approach in TRANSREG for mixed models using macros, where MIXED is called multiple times. This would not be a substitute for using link functions.
That I could understand. Why not expand TRANSREG to a mixed model basis? If there were no RANDOM statements, it would default back to the current (and very useful) tool. Of course, that would mean four different mixed model procedures, each with their own idiosyncracies. Maybe I was hasty in my approach...
You must be a registered user to add a comment. If you've already registered, sign in. Otherwise, register and sign in.