topic Proc Mixed - very large AIC, BIC, variance estimate in Statistical Procedures

Proc Mixed - very large AIC, BIC, variance estimate

SAS_User — Tue, 05 May 2009 09:27:31 GMT

My model converges and there are no other errors. However, the AIC & BIC are around 1700. The covariance parameter estimates are 99559, 160 and 28567 for the 3 random components. I have one fixed effect that is not significant. Very large AIC always bothers me, but I have not seen such large variance estimates before. Can someone throw light on this, please?

Re: Proc Mixed - very large AIC, BIC, variance estimate

reneeharper — Tue, 05 May 2009 14:54:20 GMT

SAS_User: I moved this post to the Statistical Procedures forum where there may be more folks to help with this question.

Re: Proc Mixed - very large AIC, BIC, variance estimate

statsplank — Tue, 05 May 2009 18:50:19 GMT

Hi SAS_User,

Recall that AIC and BIC from a single model do not mean much. If there are at least two models then say a difference AIC(model1)-AIC(model2) can be useful in model selection.

Re: Proc Mixed - very large AIC, BIC, variance estimate

SAS_User — Tue, 05 May 2009 19:44:34 GMT

Thanks for your thoughts on the AIC-that does make sense.
Do you have any words of wisdom for the variance estimates? i have started to convince myself that the variance estimates are in the order of 100,000 because my response variable is in the order of 3000. Would you say that is a reasonable way to look at this?
Thanks

Re: Proc Mixed - very large AIC, BIC, variance estimate

plf515 — Fri, 08 May 2009 10:54:38 GMT

Variance estimates are scale dependent.

You don't say what your topic is, or what your variables are, but suppose you are (say) looking at the effect of salary on life expectancy. Measure salary in dollars per year, variance will be huge. Measure it in tens of thousands of dollars per year, variance will be small.

Does this apply in your case?