Date:         Sat, 10 Nov 2007 06:17:45 -0800
Reply-To:     "cat.." <cat.b41@GMAIL.COM>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         "cat.." <cat.b41@GMAIL.COM>
Organization: http://groups.google.com
Subject:      Re: Univariate tests before multivariate modeling in logistic
Comments: To: sas-l@uga.edu
In-Reply-To:  <1194639366.199684.52590@k79g2000hse.googlegroups.com>
Content-Type: text/plain; charset="us-ascii"

On Nov 9, 9:16 pm, Paige Miller <paige.mil...@kodak.com> wrote:
> Received via e-mail:
> ---------------------------------------------------------------------------­-----
> "Sometime back, don't remember now how, you wrote that the covariates
> must appropriately be transformed before they can be used with PLS.
>
> "Can you please direct me to a paper that discusses this issue? I
> don'y know whta kinds of transformations I must perfom on the
> predictor varuiables before they can be used with PLS.
>
> "Also, once done with PLS, do we have to un-transform the inputs back
> to their original form. (I hope not!)"
> ---------------------------------------------------------------------------­-----
> Reply:
>
> The only transformation I can ever remember recommending is to center
> and scale your predictor (covariate) variables so that they have mean
> zero and variance 1. Even this is optional in the proper setting. Of
> course, in specific instances, you might want to take the logarithm or
> square root or other transform of your predictors, but this is done on
> an individual variable and individual dataset basis.
>
> You shouldn't have to un-transform your predictor variables. Good
> software should make this transparent. Of course, bad software
> exists...
>
> Reference: Rasmus Bro, Age K. Smilde (2003), "Centering and scaling in
> component analysis", J. of Chemometrics, Vol 17, No. 1, pp 16-33
>
> --
> Paige Miller
> paige\dot\miller \at\ kodak\dot\com

Hi Paige,

It makes no sense, in my opinion, to scale continuous covariates in a
logistic model.
Because it makes the interpretation of the coefficient complicated.
Eg: Age is a covariate and has coeff Beta in the model.
If you just center it before modeling, you can infer than an increase
of 1 year generates an increase in OR of exp(beta), which you cannot
do if you have also scaled the covariate.

Regards,

Catherine.