Proc PLS Regression -- Getting Back To Original Predictors

[NKormanik]

As I'm understanding things, PLS regression comes up with a number of 'components.'  But how do these components tell us which of the original predictors is most important?

Thanks!

Nicholas Kormanik

[Ksharp · Accepted solution]

There is VIP plot you can refer to .

ods graphics on/width=1000px height=500px;
ods output VariableImportancePlot= VariableImportancePlot;
proc pls data=class missing=em nfac=2 plot=(ParmProfiles VIP) details; * cv=split cvtest(seed=12345);
class sex;
model age=weight height sex;
output out=x predicted=p;
run;
proc sort data=VariableImportancePlot;
by descending VIP;
run;

[PaigeMiller]

The variables with high loadings (either positive or negative) are the important variables in that component. There's no real cutoff, the meaning of "high" is relative to the other loadings in that component.

The loadings are available via the DETAILS option of the PROC PLS statement. You can save the loadings to a SAS data set if you want via

ods output xloadings=xloadings;

[Ksharp]

[NKormanik]

@Ksharp @PaigeMiller   You guys rock.  Thanks so much!

[NKormanik]

@Ksharp Noticed that your labels in the above graph are in Chinese.

How about that....

[Ksharp]

?? I am from China and running Chinese version SAS. If you are running English version SAS with same code , You would get English result . It is called Variables Important Plot . a.k.a VIP

[NKormanik]

@Ksharp   Yes, I know.  Mainland, would be my guess.  SAS with everything in Chinese just seemed even more overwhelming than SAS in English.  Like, OMG!!