Date:         Tue, 6 May 2008 16:38:14 -0400
Reply-To:     Wensui Liu <liuwensui@GMAIL.COM>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Wensui Liu <liuwensui@GMAIL.COM>
Subject:      Re: Scoring nlmixed
Comments: To: Dale McLerran <stringplayer_2@yahoo.com>
In-Reply-To:  <237232.66091.qm@web32207.mail.mud.yahoo.com>
Content-Type: text/plain; charset=ISO-8859-1

beautiful!
it is so simple but I just couldn't get it.
thank you so much, Dale!

On Tue, May 6, 2008 at 3:48 PM, Dale McLerran <stringplayer_2@yahoo.com> wrote:
>
> --- "Dr. Eli Y. Kling" <Eli.Kling@GMAIL.COM> wrote:
>
>  > Hi Guys,
>  >
>  > For reasons I will not boar you here, I prefer scoring my predictions
>  > by adding the data to be socred to the training data with the target
>  > field 'missing'. This works like a charm for the good old procedures
>  > like Reg, GLM, and Logistic. (I know, I Know, you have the score
>  > procedure and the score command).
>  >
>  > I have been converting some code using proc logistic to use proc
>  > NLMIXED; which by the way is a cool bit of statistics and a powerfull
>  > sas tool. my problem is that (to my knoledge) it does not support
>  > scoring in any mannor apart for the training data. I realize that
>  > scoring here is much more tricky as random effects are involved.
>  >
>  > For now I export the coeficiants for the fixed effects and do the
>  > scoring in a data step. But I am not sure what to do with the random
>  > effect. Any Ideas? Any code out there to do a generic scoring for
>  > NLMIXED on a fresh dataset?
>  >
>  >
>  > P.S. I used to be very active on the old SAS-L when I lived in Vienna
>  > and Denmark. Now I am in England and I am back on the sas game. A
>  > special Hi to anybody who still remembers me.
>  >
>  > With regard,
>  >
>  > Eli Y. Kling
>  >
>
>  The following approach should work as long as the data to be scored
>  are from the same subjects as the data used to fit the model.
>  Combine both fitting and scoring data into a single data set and
>  construct an indicator variable of the data to be scored.  Then
>  write NLMIXED code in which you construct your likelihood function.
>  If the data are from the set to be scored, then set the log-
>  likelihood contribution to zero.  That way, you will only use the
>  fitting data in the estimation process.  You can then write a
>  PREDICT statement which outputs estimated response values.  If you
>  put a WHERE clause on the output data set that restricts the data
>  to those which need to be scored, then you will get results only
>  for the data to be scored which were not used to fit the model.
>
>  To be a little more concrete, you would have something like the
>  following:
>
>     data combined;
>       set fit_data(in=a)
>           score_data(in=b);
>       by id;
>       score = b;
>     run;
>
>     proc nlmixed data=combined;
>       eta = ...;
>       mu = <f(eta)>;
>       if score=1 then
>          ll = 0;
>       else
>          ll = ...;
>       model y ~ general(ll) subject=id;
>       random ...
>       predict mu out=predicted(where=(score=1));
>     run;
>
>
>  HTH,
>
>  Dale
>
>  ---------------------------------------
>  Dale McLerran
>  Fred Hutchinson Cancer Research Center
>  mailto: dmclerra@NO_SPAMfhcrc.org
>  Ph:  (206) 667-2926
>  Fax: (206) 667-5977
>  ---------------------------------------
>
>
>
>       ____________________________________________________________________________________
>  Be a better friend, newshound, and
>  know-it-all with Yahoo! Mobile.  Try it now.  http://mobile.yahoo.com/;_ylt=Ahu06i62sR8HDtDypao8Wcj9tAcJ
>



--
===============================
WenSui Liu
ChoicePoint Precision Marketing
Phone: 678-893-9457
Email : wensui.liu@choicepoint.com
Blog : statcompute.spaces.live.com
===============================