06-10-2016 09:48 AM
I'm a business user focused on interpretation of results (so I can muck around with somebody else's code and have enough of a statistical background to be dangerous). I am looking for help in interpretting the outputs of PROC MIXED Influence Diagnostics.
1. My approach initially is to look at Cook's D and use a cutoff of 4/n, DFFITS with a cutoff of 2SQRT(p/n) and examine if RMSE changes much for levels with or without this level. is that a valid approach?
2. I read breifly somewhere that I could use PRESS to examine overfitting - does anybody have any direction they coud provide on how I should thinkg aobut this?
3. How should I interpret the other measures provided?
4. I have a model that provides no diagnostics, providing notes only of Loss of Rank or Change in Singularity. Need help in interpetting this (but I fear this means that my model is fragile and cannot live without all the levels that I'm using.
5. Anything that I'm missing / not thinking of?
Help greatly appreciated
06-13-2016 10:29 AM
In addition to the examples in the MIXED chapter of the User's Guide (available on-line), I recommend that you read the following: