Re: Curve Fit Stats

[sample@lyon.edu (John D. Sample)]

At 9:55 AM on 9/12/96, <phys-l@mailer.uwf.edu> wrote:
> A question for both HS and college instructors. I've been looking at
> some graphing and curve fitting packages,  trying to figure out just
> what statistics should be provided by the software when performing a
> least-squares fit.
>
> > From my background, what's useful is the uncertainties on the fitted
> parameters, ie,  the standard error of the slope and intercept, or
> whatever the parameters may be. That tells you how well determined
> the slope is, for instance. The scatter standard deviation measures
> how far away from the fitted line, on average, the points are. That's
> useful too for comparison to the uncertainties of each y measurement.
>
> John E. Gastineau
> 304 296 1966
> Morgantown, WV
> http://www.imagixx.net/~jgastin
> email: jgastin@imagixx.net

I agree that the uncertainty of the fit parameters are more meaningful than
correlation coefficients.  You can get a very high R while clearly having a
poor fit.  So I tell intro students to ignore R, and look at whether the
data falls randomly above/below the curve and look at the relative size of
the parameters to the uncertainties of the parameters.

We have Kaleidagraph for the Macs in our lab.  Unfortunately, getting the
uncertainty is not a default and the students have to jump through some
minor hoops to get them.  There is a shareware program I've seen,
MacCurveFit which is very intuitive to use and gives the uncertainties, but
you can't edit the format of the graph much at all.

J. D. Sample                          (501) 698-4625
Math-Physics Dept                     sample@lyon.edu
Lyon College
2300 Highland Road
Batesville, Arkansas 72501