3

It is postulated that one of the main issue is to find an appropriate covariance structure for repeated measures designs [Ref1]. SAS' PROC MIXED contains a number of covariance structures.

Despite "many choices among models to fit to a given data set in the mixed model setting... [and] we must always remember that all models are wrong (because they are idealized simplifications of Nature), but some are useful [citation]." there are different recommendations for choosing among the covariance models are known [Ref2], [Ref3].

I do experiments with simple one-way within-subjects RM ANOVA (balanced and unbalanced) design without a between-subject factor described here.

My question is
*how is it important if I have a "condition" (not time) as a within-subject factor and how to choose an appropriate covariance structure in this case*?

Perhaps this is a strange case as I have a native substance and its chemical derivatives with similar molecular structures. So probably I have correlation caused by the substance itself?

Community
  • 1
abc
  • 1,747
  • 3
  • 17
  • 32
  • 1
    Thanks for the references: I had read in Littell et al.'s "SAS for Mixed Models" [book](http://books.google.co.nz/books?id=z9qv32OyEu4C&printsec=frontcover#v=onepage&q&f=false) that using AIC might be a reasonable way to choose between two covariance structures **for otherwise identically specified models** – James Stanley Mar 21 '13 at 01:46

1 Answers1

2

If you have no basis for a particular correlation structure and the unstructured correlation option requires too many parameters, I would try two different ones to see if the results are sensitive to the choice . In my experience they usually aren't sensitive. I would probably compare AR(1) to compound symmetry.

Michael R. Chernick
  • 39,640
  • 28
  • 74
  • 143
  • thank you for you suggestion. But what characteristics do I have to look at? – abc Jun 15 '12 at 04:04
  • You just want to see the extent to which the results change depending on the choice of covariance structure. – Michael R. Chernick Jun 15 '12 at 04:18
  • OK. And the *minimal* Akaike Information Criterion is one of them. Right? (as noted by G.E.Dallal "[Choosing Among Covariance Structures](http://www.jerrydallal.com/LHSP/repeat2.htm)") – abc Jun 15 '12 at 10:15
  • Not exactly. AIC is a criterion to minimize when selecting between models. What I was referring to is really a subjective assessment of the change in fit or the change in the AIC if you want to look at it that way. – Michael R. Chernick Jun 15 '12 at 10:37