06-10-2012 06:03 PM
I'm running Proc Mixed, using a Random statement for repeated measures. Have a bunch of variables, so run each one separately using the same code. Most have run just fine, but 3 variables all from a second database are giving me "non-est" for the means. I've tried a bunch of random changes to no avail. Rather than this hit-or-miss process, can someone tell me what things cause non-estimable means? Thank you!
06-11-2012 09:01 AM
There are a number of possible causes. A couple that I would look for:
is the new variable a linear combination of others in the data. Sometimes that may be linear in just this dataset (akin to autocorrelation) or a design issue (like having pre, post, and delta variables in the model together).
one of your repeats may not have any variability in the predictor.
06-12-2012 01:19 AM
Thank you both! I think I figured it out from your comments. My model was y = treatment trial day, but day essentially equals treatment in this database, whereas it didn't in the other one. I removed it and voila! Means!
06-11-2012 09:16 AM
Can you share the code you are using? Besides the issues that Doc mentioned, I would be concerned about missing cells, such that the interaction term could be estimated, but main effects could not. If this is the case, you may have to address everything using a means model.