"Vitaly Kupisk" <[EMAIL PROTECTED]> wrote in message
[EMAIL PROTECTED]">news:[EMAIL PROTECTED]...
> Hi,
>
> I am looking for a good orthogonal linear regression (L1) algorithm
> and have been pointed to Bargiela and Hartley
> www.doc.ntu.ac.uk/RTTS/Papers/rttg-publ02.ps. I don't quite get what
> they mean by "solving equation 29" -- obviuosly a 0 vector is a
> solution (often the only one); they must mean something else, so I
> can't reconstruct the algorithm.
>
> Can anyone explain their algorithm, or does anyone know of an
> available implementation of it, or for that matter any good orthogonal
> linear L1 regression algorithm/implementation?
>
> I saw some references in the group (e.g Jackson, J. E. (1991), "A
> Users Guide To Principal Components", John Wiley and Sons, New York,
> chapter 15.) but haven't looked at them yet.  Are these older
> algorithms much inferior to Bargiela-Hartley's?
>
> Please copy me on your replies to the group.
>
> Thank you
>
> Vitaly Kupisk
> [EMAIL PROTECTED]
--------------------------------------------------------------
The basic problem with orthogonal regression is that it minimizes the sum of
the square of the error as measured by a vector perpendicular to the
regression line through each point. This is an assumption about the variance
of the X and Y values, which cannot be verified from the data. The better
method is to make an assumption on the ratio of the variances, or to assume
that the X values have no errors. In many cases, where the X values are
instrument measurements, the X variance can be estimated prior to collecting
the data, and the regression based on minimizing the Y variance.

With orthogonal regression, one depends on the validity of the linearity
assumption. Slight non-linearity can give problems.

As Woodhouser points out, scaling (or normalizing) the X and Y values can
have a very profound effect on the values of the regression coefficients.

David Heiser


.
.
=================================================================
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