Date:         Wed, 29 Jul 2009 08:44:27 -0400
Reply-To:     Jerry Davis <jwd@UGA.EDU>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Jerry Davis <jwd@UGA.EDU>
Subject:      Re: proc glimmix - denominator degrees of freedom
In-Reply-To:  <aac7ab09-091b-4935-a754-0cc08dfdea81@2g2000prl.googlegroups.com>
Content-Type: text/plain; charset="US-ASCII"

RFisher wrote:

>  I'm using proc GLIMMIX as part of a study to examine habitat use of a
> small passerine. Each bird was radio-tagged and tracked once per day
> for several weeks, resulting in a set of habitat measurements of where
...snip...
> It seems like the few studies I've read either don't report what
> option they used or have reported using the Satterthwaite
> approximation (unfortunately with little justification). I'm not sure
> what I should be looking for in my data (or my objectives) to decide
> on what option to choose. Or do I just go with the default?

People close to the development of GLIMMIX generally recommend
KENWARDROGER as the best all around DDFM method.  It adjusts the
variance-covariance matrix then calculates Satterthwaite type degrees of
freedom.

It can have undesirable consequences if the covariance matrix has second
derivatives when using some types of covariance matrices.  The DDFM
option KR(firstorder) was added for these situations.  More details can
be found in the documentation for GLIMMIX and the DDFM option.

So, unless you have a specific reason not to use it, DDFM=KR might be a
good default choice.  It can be more computationally intensive than
other methods.

Jerry