Date:         Sun, 17 Jan 2010 19:01:57 -0800
Reply-To:     Gary Klein <kleingary@YAHOO.COM>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Gary Klein <kleingary@YAHOO.COM>
Subject:      Re: model comparison
In-Reply-To:  <12477987.1263776885019.JavaMail.root@elwamui-milano.atl.sa.earthlink.net>
Content-Type: text/plain; charset=iso-8859-1

I like Peter's answer, but always remember that comparing models using -2LL, they should be nested models (and always have the same number of observations - sometimes a missing value of a predictor can throw things off).

Ching, do you have a reference regarding halving the p-value.  I've never heard of it either.





--- On Sun, 1/17/10, Peter Flom <peterflomconsulting@MINDSPRING.COM> wrote:

> From: Peter Flom <peterflomconsulting@MINDSPRING.COM>
> Subject: Re: model comparison
> To: SAS-L@LISTSERV.UGA.EDU
> Date: Sunday, January 17, 2010, 5:08 PM
> Ching <kcwong5@GMAIL.COM>
> wrote
> >Hi everyone.
> >
> >I've done some reading about halving the p-value
> (difference in -2LL)
> >when comparing 2 different models produced by proc
> mixed. Is it more
> >appropriate this way? but why?
> >
> >Your advice is greatly appreciated. Have a great day.
>
> I'm not sure what you are asking.  The difference
> between two log likelihoods is not a p-value, but it has an
> associated p-value.  AFAIK, there is no reason to halve
> it.
>
> But if you are talking about the difference in the LL,
> then, well, yes, you should halve the difference, because
> both are doubled.
>
> HTH
>
> Peter
>
> Peter L. Flom, PhD
> Statistical Consultant
> Website: http://www
> DOT statisticalanalysisconsulting DOT com/
> Writing; http://www.associatedcontent.com/user/582880/peter_flom.html
> Twitter:   @peterflom
>