Daivd:  i was also under the impression that X and Y had to be in the same 
units for orthogonal regression to make sense.  is this true?  it's not 
intuitive to me why this would be the case.


Quoting David Heiser <[EMAIL PROTECTED]>:

> 
> 
> -----Original Message-----
> Wrom: PBARHDMNNSKVFVWRKJVZCMHVIBG
> [mailto:[EMAIL PROTECTED]]On
> Behalf Of JJ Diamond
> Sent: Saturday, April 06, 2002 9:12 AM
> To: [EMAIL PROTECTED]
> Subject: Re: linear regession
> 
> 
> [EMAIL PROTECTED] (Shi) wrote in message
> news:<[EMAIL PROTECTED]>...
> > > taking you at your word, i assume you really mean "perpendicular"
> and
> > > not vertical deviations.  here's something that might help:
> > >
> > >   www.mathpages.com/home/kmath110.htm
> >
> > Thanks : )
> >
> > Yes I do mean "perpendicular" deviations.
> ----------------------------------------------------------------------
> What you (Shi) is doing is a regression in which errors of measurement
> occur
> with both Y and X variables.
> 
> If the data set has only one X variable, then the technique is called
> Orthogonal Regression. Since there is insufficient data, you have to
> assume
> a numerical value for the ratio of the Y error variance to the X error
> variance. Carroll and Ruppert (American Statistician, Feb 1996, page
> 1)
> discuss this and how to obtain solutions. Although Shi states
> "perpendicular" the angle depends on the variance ratio. Their
> references
> are a good source for further information, including algorithms.
> 
> If there are more than one X variable, the method is called Total
> Least
> Squares. It is all matrix stuff, and computer algorithms. It is
> dependent on
> how the errors are distributed among the X variables. There are a
> whole
> bunch of analytic approaches.
> 
> DAHeiser
> 
> 
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to