Date:         Tue, 10 Jul 2007 13:31:51 -0700
Reply-To:     Andrew Hill <hill.andrewd@GMAIL.COM>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Andrew Hill <hill.andrewd@GMAIL.COM>
Subject:      Re: Can I do this multiplr regression?
In-Reply-To:  <1184069229.522546.12810@57g2000hsv.googlegroups.com>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed

This can be done, and is done all the time in modeling growth from
time t0 to time t1 in many different contexts.

The things your colleague will want to check are correlation between
the predictor variables and have a way to account for co-linearity, if
it exists.

Out of curiosity, could he also substitute x2 in for x1 have a a valid model?

Hope this helps.

Andrew

On 7/10/07, Paige Miller <paige.miller@kodak.com> wrote:
> On Jul 10, 2:28 am, davidlcass...@MSN.COM (David L Cassell) wrote:
> > x...@LSU.EDU wrote:
> >
> > >A colleague asked me if he can do this multiple regression. He first
> > >calculated differences from X1 and X2 and used it as the dependent variable
> > >(Y). Then he wanted to run regression Y = X0 X1 X3 X4 where Y is the
> > >difference between X1 and X2, and he wanted to put X1 as one predictor. X0,
> > >X3, X4 are independent from Y.
> >
> > >Can this multiple regression be done? Which assumption is violated?
> >
> > This, in itself, does not violate any regression assumptions.  Of course:
> >
> > [1] regression assumptions may be violated that have nothing to do with
> > using Y=X1-X2.
> >
> > and
> >
> > [2] there may be serious logical violations in doing this.
> >
> > #2 depends on the data, the meaning of the variables, the underlying
> > theory involved, the scope of the data, etc.  It may be entirely feasible
> > to perform the regression, even though scientifically it could be a hideous
> > nightmare.  I think this is a subject-matter problem more than a
> > statistical one.
>
> I would add that since Y is a function of X1 and other variables, then
> putting X1 in the right hand side as a predictor is very problematic.
>
> Specifically, Y may be highly correlated with X1. Now that in itself
> isn't a problem. The problem is that if the large regression doesn't
> improve the fit over the regression  involving Y and only X1 and an
> intercept, then the large regression is useless. It simply is
> explaining the a priori correlation between Y and X1.
>
> --
> Paige Miller
> paige\dot\miller \at\ kodak\dot\com
>