|Date: ||Wed, 20 Jun 2007 12:52:36 -0700|
|Reply-To: ||Andrew Hill <hill.andrewd@GMAIL.COM>|
|Sender: ||"SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>|
|From: ||Andrew Hill <hill.andrewd@GMAIL.COM>|
|Subject: ||Re: AIC in proc nlin|
|Content-Type: ||text/plain; charset=ISO-8859-1; format=flowed|
I'll hunt around and see what I can find.
Nice weather after the rain the past few days, yes?
You are right, I am trying to minimize SSE. The response variable is
pretty wonky, so I was using bootstrapping to get an estimate of model
fit and the parameter estimates.
I might be able to use a gamma, but a 3 parameter weibull might also
work, if nlmixed supports it. I'll have to see.
Thanks for your help.
On 6/20/07, Dale McLerran <firstname.lastname@example.org> wrote:
> --- Andrew Hill <hill.andrewd@GMAIL.COM> wrote:
> > Thanks Matthew.
> > I'll have to look at these more closely. I am hoping that I can find
> > one that I can just move the model statement to and just use. It
> > would
> > make my live simpler.
> > Does proc nlin mixed require one to use a mixed model?
> > The one I'm currently using is an exponential model.
> > change = exp(b0 + b1x1 + b2x2 + b3x3 + b4x4 + b5x5).
> > I don't want to use proc reg because the data are definitely
> > non-normal so the assumption that go into that are wrong for this
> > data.
> > Thanks,
> > Andrew
> From what you have written here, I would recommend the NLMIXED
> procedure all the more. The equation that you specified for AIC
> in your previous post assumes that the residuals are normally
> distributed. But you indicate above that the response is not
> normally distributed. Therefore, you should not be using the
> form you specified for computing AIC.
> In response to your specific question as to whether the NLMIXED
> procedure requires that you fit a mixed model, the answer is a
> resounding NO!!! Just search the SAS-L archives for some examples
> (most contributed by yours truly) where NLMIXED is employed to model
> all manner of likelihood functions. What the NLMIXED procedure
> does require is that one specify a distribution. Based on the
> specified distribution, NLMIXED computes maximum likelihood (or
> approximate ML) estimates of the parameters of the model.
> I can't tell from way over here, but it sounds as though you
> minimized an error sum of squares function in NLIN. Specifying
> a distributional model can be more difficult. There may not be
> a distribution which is exactly appropriate for your data. But
> you can probably specify a distribution which models your
> response reasonably well.
> Dale McLerran
> Fred Hutchinson Cancer Research Center
> mailto: dmclerra@NO_SPAMfhcrc.org
> Ph: (206) 667-2926
> Fax: (206) 667-5977
> Choose the right car based on your needs. Check out Yahoo! Autos new Car Finder tool.