[SciPy-user] Fit statistics for sum of squared relative error
Tue Feb 13 02:09:16 CST 2007
I've been doing fits to lowest sum of squared relative
error for a while now. These are useful when a data
set exhibits increasing heteroscedasticity as the
values of the independent variable increase, i.e.,
data scatter proportional to distance along the x axis.
You can test this at http://zunzun.com by selecting a
fitting target of "Lowest sum of squared relative errors"
when fitting, and this is also in the Python Equations
package at http://sf.net/ptojects/pythonequations.
Having investigated fit statistics for some time now,
it seems *everything* is geared toward absolute error,
with not the slightest drop that I can find regarding
relative error. These statistics are needed when
performing SSQREL rather than SSQABS fitting.
Can I safely use existing routines for covariance
matrices and parameter standard errors simply by
substituting dy(relative)/dx and relative error
whenever dy(absolute)/dx and absolute error are
used? I apologize, but this is over my head and
I would like to report the fit statistics properly.
P.S. I don't yet have much in the way of fit statistics
on the web site, this is what I'm currently working on.
I found much of what I need in the BSD-style licensed MPFIT.py
written by Mark Rivers.
More information about the SciPy-user