It make perfect sense IF the scale of the X's is representative of how much
they vary in "real life". For example, suppose you have a chemical process
where the temperature normally varies by about a degree and the
concentration of something normally varies by about a percent. If your
experiment varies temperature by 10 degrees and concentration by 2%, you
will amplify the "importance" of temperature by about 5 times as much as the
importance of concentration.
By the way, the method you describe is built into most software. For
example, JMP calls them "standardized betas".
-------------------------------------------
Emil M Friedman, PhD
2304 Richmond Road
Beachwood, OH 44122
[log in to unmask]
216-591-1750 (voice)
775-249-6744 (fax)
----- Original Message -----
From: "Siobhan O'Keeffe" <[log in to unmask]>
To: <[log in to unmask]>
Sent: Thursday, May 08, 2003 5:28 AM
Subject: Input variables
> Allstaters
> A project I am working on is looking at the "relative importance" of input
> variables in helping to predict an output variable using linear
regression.
> I was wondering about the viability of the following method.
> Scale all the input variables by z-score, perform the regression, then
> the "relative importance" can be read off from the size of the beta co-
> efficients. Would anyone have any comments on this method or be able to
> recommend something better.
> Ideally the project requires output of the following
> Var X is the most important, which is 2.5 times more important than Vay Y.
> Regards
> PS I use the word important in the sense of best predictor
>
> Thanks Siobhan
>
|