Date:         Thu, 1 Nov 2007 13:49:18 -0400
Reply-To:     "Elmaache, Hamani" <Hamani.Elmaache1@CRA-ARC.GC.CA>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         "Elmaache, Hamani" <Hamani.Elmaache1@CRA-ARC.GC.CA>
Subject:      Re: Goodness of fit for ZIP random effects model using proc
              nlmixed
Comments: To: Wensui Liu <liuwensui@GMAIL.COM>
In-Reply-To:  A<1115a2b00711011018m2ed8468v1a63e9fc1bdee051@mail.gmail.com>
Content-Type: text/plain; charset="iso-8859-1"

Hi Wensui Liu.
It's very interesting to compare the estimated density to the observed
density. Can you please indicate how to do it when your are using Proc NLMIXED? How to display these information?

Thanks a lot.
Hamani.

-----Original Message-----
From: SAS(r) Discussion [mailto:SAS-L@LISTSERV.UGA.EDU]
Sent: November 1, 2007 1:18 PM
To: SAS-L@LISTSERV.UGA.EDU
Subject: Re: Goodness of fit for ZIP random effects model using proc
nlmixed


Hi, Anne,

to access the goodness-of-fit of a count data, using deviance or
deviance residuals, which is based upon expected value, is not
adequate.

The reason you are using ZIP model might be because of the
over-disperson / excess zeroes. So variance has to be considered. If I
were you, I would compare the estimated density to the observed
density, which takes the variance into consideration.

I have addressed this issue in my submitted sugi paper and hope it
will be accepted.

Just my $0.02.

On 11/1/07, Anne Mann <conceptualness@hotmail.com> wrote:
> Hello,
>
> I am trying to figure out how to assess the adequacy of a zero-inflated
> poisson regression with a random effect model that I've fit using SAS proc
> nlmixed (I've included the code below, taken from a SAS book).
>
> I was reading a paper by Yau and Lee in Stats and Medicine 2001 (bibtex
> below) which describes how one might perform a scaled deviance test or plot
> the deviance residuals.  I would like to do this in SAS and was hoping that
> I could somehow get the output to conduct these tests from proc nlmixed, but
> am not sure where to start.  To do either of these tasks, I would need to
> get the log-likelihood for the vector of responses given the random effects,
> and need to get the log-likelihood for the vector or expected values of the
> responses given the random effects.  Is there an easy way (or any way) to
> get these quantities? Or more generally, does anyone have any suggestions
> for assessing the adequacy of a ZIP random effects model using SAS?
>
> Thanks very much for the help.
> Ann
>
>
> proc nlmixed data=zip;
>  parameters b0=0 b1=0 a0=0 s2u=1;
>  linpinfl=a0;
>  infprob=1/(1+exp(-linpinfl));
>  lambda=exp(b0 + b1*x + u);
>  if count=0 then
>    ll=log(infprob + (1-infprob)*exp(-lambda));
>  else ll=log((1-infprob)) + count*log(lambda)-lgamma(count+1)-lambda;
>  model count~general(ll);
>  random u~normal(0,s2u) subject=s;
>  estimate"inflation probability" infprob;
> run;
>
> @article{yau2001zip,
>  title={{Zero-inflated Poisson regression with random effects to evaluate
> an occupational injury prevention programme}},
>  author={Yau, K.K.W. and Lee, A.H.},
>  journal={Statistics in Medicine},
>  volume={20},
>  number={19},
>  pages={2907--2920},
>  year={2001}
> }
>


--
===============================
WenSui Liu
Statistical Project Manager
ChoicePoint Precision Marketing
(http://spaces.msn.com/statcompute/blog)
===============================