Peter, Thanks for sharing your paper again. The paper in the original post is about the variable selection in a multinomial logit. Even I knew that glmselect procedure couldn't be applicable to such a logit model. an expert like you should certainly realize it without reading too much into the paper.

On Fri, Sep 25, 2009 at 4:53 PM, Peter Flom <peterflomconsulting@mindspring.com> wrote: > I will try to look at this in detail, maybe this weekend, but a much simpler method is available PROC GLMSELECT offers both > LASSO and LAR > > See the paper David Cassell and I wrote together - > Stopping Stepwise - Why Stepwise Variable Selection Methods are Bad, and What you Should Use > > it's at various places - one SGF, two NESUGs, a NYASUG and either a WUSS or a PNWSUG. > > Here is the one from NESUG 2007 > > www.nesug.org/Proceedings/nesug07/sa/sa07.pdf > > > HTH > > Peter > > -----Original Message----- >>From: liudapaul <liudapaul@GMAIL.COM> >>Sent: Sep 25, 2009 4:30 PM >>To: SAS-L@LISTSERV.UGA.EDU >>Subject: Variable reduction method >> >>I've been reading a SUGI paper regarding to variable reduction but >>couldn't figure out one part of the code on page 6. I attach the code >>below and for anyone who is interested in the paper I also attached >>the link. >> >>*--------------------------------------------------------------* >>| Loop through all Numeric Variables in Data Set | >>*--------------------------------------------------------------*; >>%do range=1 %to &numobs %by 200; >>%let upper=%eval(&range+199); >> >>data varlist; >>set &inputds._keeplist; >>if &range<=_n_<=&upper ; >>run; >> >>data _null_; >>set varlist NOBS=numobs end=last; >>call symput(cats("var",_n_),variable); >>if last then call symput("numobs",numobs); >>run; >> >>proc corr data=&inputds. OUTP=%cmpres(corr_&range._&upper) NOPRINT; >>var %do j=1 %to &numobs; &&var&j %end; &dumlist; >>%if %bquote(&weight) ne %then %str(weight &weight;); >>RUN; >> >>%put Output Correlation Data Set Will be Called: %cmpres >>(corr_&range._&upper); >> >>ods listing close; >>data _null_; >>set varlist end=last; >>call execute("ods output Test ANOVA=reg_temp(KEEP=MODEL SOURCE DF MS >>FVALUE PROBF >>WHERE=(Source='Numerator'));"); >>call execute("proc REG data=%cmpres(corr_&range._&upper) ;"); >>call execute(variable||": model "||variable||"=&dumlist ;"); >>call execute("test &dumlist_comma;"); >>call execute("RUN;"); >>call execute("quit;"); >>call execute("ods output close;"); >> >>IF EXIST("WORK.REG_TEMP") THEN DO; >>call execute("Proc APPEND base=%cmpres(work.reg_total_&out_suffix) >>data=reg_temp(drop=source) force;"); >>call execute("run;"); >>END; >>RUN; >>%end; >> >>my question is what's the idea behind using a correlation matrix as >>input of a regression? >> >>Link to the paper: >>http://www2.sas.com/proceedings/forum2007/081-2007.pdf > > > Peter L. Flom, PhD > Statistical Consultant > Website: www DOT peterflomconsulting DOT com > Writing; http://www.associatedcontent.com/user/582880/peter_flom.html > Twitter: @peterflom >

-- ============================== WenSui Liu Blog : statcompute.spaces.live.com Tough Times Never Last. But Tough People Do. - Robert Schuller ==============================