Joakim Englund <joakim.englund@GMAIL.COM> wrote

>I'm currently learning about general strategies of how to select the "best" >covariance structure in repeated measurements trials using proc mixed. To >aid me I use "Applied Mixed Models in Medicine" by Helen Brown and Robin >Prescott. They propose to use a likelihood ratio test when choosing between >two different models. This seems reasonable to me and can be used whenever >the simpler model is nested within the more complex one. >

As an aside - is the book good?

>Question 1: Do you agree with this approach, or would you prefer to use >criterias such as AICC which penalises over-parameterised models? (I know >one mussed take other factors into consideration, such as trial design and >the meaning of the structure, but I would like your opinion on this >technical approach from an "all other things being equal" - viewpoint.) > >If one chooses to proceed with the likelihood ratio test approach, there can >be situations when the choice of structure stands between two models where >no one is nested within the other. In such an instance, there is no formal >test that can be applied. Brown and Prescott then suggests to pick the model >with the highest likelihood (even though that approach is not explicitly >stated). >

I await the response of Dale (and others) who really know this stuff cold. But I tend to favor AICC in general. In my experience, though, AICC, AIC and LR (where applicable) all give the same answer. I know this isn't mathematically necessary, but that's what I've nearly always found.

I'd be interested in other people's experience.

>Question 2: In the situation described above, would it not be more >appropriate to use something like AICC, as looking merely at the likelihood >would always favour more parameterised models? >

Clearly you can't use the LR test here.

There's debate among advocates of AIC, AICC, SBC and others, about which is best. I haven't kept up with this debate.

I look forward to discussion of this topic!

Peter

Peter L. Flom, PhD Statistical Consultant www DOT peterflomconsulting DOT com http://www.associatedcontent.com/user/582880/peter_flom.html