Date: Fri, 30 Mar 2007 11:41:55 -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: logistic regression
In-Reply-To: <1175267695.041248.245320@y66g2000hsf.googlegroups.com>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
SL,
very smart macro by transforming a logit fit to a ols fit problem.
But after the running the program, the specific transformation of
predictor is still up in the air and a judgement call by statistician.
In this case, I will suggest using GAM procedure to give more insight
of it.
On 3/30/07, Shiling Zhang <shiling99@yahoo.com> wrote:
> On Mar 30, 9:49 am, "Savvy" <snamud...@gmail.com> wrote:
> > I am doing a binary logistic regression. All of the variables in my
> > model are untransformed variables. Should I be concerned that none of
> > the transformations came in?
>
> It really depends on the relationship between log odds and covariates.
> The relationship between log odds and covariates may be very
> complicatied when there are interactions and correlations among your
> covariates. Usually the first step is to investigate the relationships
> by plotting log odds and each and every covariate candidate. If it is
> not linear, you may consider tranforming your covariates and using any
> theritical work to guide if there is any interaction. This is JUST the
> first step and a of dirt work.
>
> Here is an example to plot the logodds with a numeric variable. The
> plots will give you the relationship between log odds and covariate x1
> is quadratic, and covariate x3 is linear and covariate x4 is none.
>
> HTH.
>
>
> data t1;
> do i = 1 to 5000;
> x1=ceil(ranuni(1234567890)*70)+5; x2=x1*x1;
> x3=rannor(1234567890);x4=rannor(1234567890);
> y=0.5-0.05*x1+0.001*x2+0.5*x3 > rannor(1234567890);
> output;
> end;
> run;
>
> %macro logit_plot(idsn=t1,odsn=all,y=y,x=x1 x3 x4, group=100,
> combined=0,standard=n);
> title ;
> %let n=%eval((%sysfunc(countc(%cmpres(&x),%str( )))
> +1)*(%length(&x)>0));
> %let keepit=&y &x;
> %let rankvar=;
>
> %do i = 1 %to &n;
> %let rankvar=&rankvar %scan(&x, &i)_r;
> %end;
>
> proc rank data=&idsn(keep=&keepit) out=_&idsn group=&group
> TIES=MEAN ;
> var &x;
> ranks &rankvar;
> run;
>
> %if %upcase(&standard)=Y %then %do;
> proc standard data=_&idsn mean=0 std=1 replace print out=_&idsn;
> var &x;
> run;
> %end;
>
> data _v&idsn/view=_v&idsn;
> set
> %do i = 1 %to &n;
> %let var=%scan(&x, &i);
> _&idsn(rename=(&var=_X_ &var._r=_X_R) in=_&var )
> %end;
> ;
> %do i = 1 %to &n;
> %let var=%scan(&x, &i);
> if _&var then class_var="%upcase(&var)";
> %end;
> run;
>
> proc means data=_v&idsn nway noprint;
> class class_var _X_R ;
> var &y _x_;
> output out=&odsn(keep=class_var _X_R &y _x_) mean=&y _x_;
> run;
>
> data &odsn;
> set &odsn;
> &y._logit=log(&y/(1-&y));
> run;
>
> title 'logit values VS NUM covariate values';
> proc print data=&odsn; run;
>
> %if "&combined"="1" %then %do;
> title ">logodd plot of &y._logit VS. %upcase(&x)<";
> %let byit=;
> %let plotit=%str(plot &y._logit*_x_ =class_var/;);
> symbol i=j value=circle cv=r;
> symbol2 i=j value=dot cv=b;
> symbol3 i=j value='=' cv=g;
> symbol4 i=j value='$' cv=pk;
> symbol5 i=j value='&' cv=o;
> symbol6 i=j value='%' cv=p;
>
>
> %end;
> %else %do;
> %let byit=%str(by class_var;);
> %let plotit=%str(plot &y._logit*_x_ /regeqn;);
> symbol i=rlclm95
> value=diamond
> cv=r
> ci=b
> co=g
> ;
> %end;
>
> proc gplot data=&odsn;
> &byit
> &plotit
> run;
> quit;
>
> proc datasets lib=work;
> delete _&idsn _v&idsn(memtype=view)
> ;
> quit;
> %mend logit_plot;
>
> %logit_plot(idsn=t1,odsn=all,y=y,x=x1 x3 x4, group=100,
> combined=0,standard=n)
>
--
WenSui Liu
A lousy statistician who happens to know a little programming
(http://spaces.msn.com/statcompute/blog)
| 
