I have trouble with the magnitude of the error of parameters when using the
non-linear curve fitter. I'm using a self-defined "modified saturating
exponential curve" of the form y=a*(1-exp(-(x+b)^d/c)), with d fixed, and a,b,c
free parameters. For my work (dosimetric dating in geology) the parameter b
and its error is the most important.
I've fitted the same data-set with the above function in Origin 3.5 and in
Origin 4.1. The result were somewhat confusing. The attained parameters were
exactly the same, also Chi^2, which should be the case, but the errors of all
three parameters were only roughly a fourth as large in the new version than in
the old one (the exact ratios new/old are 0.258, 0.245, 0.262 for a,b,c
respectively)!!!
From my experience and from the comparison with other fitting programs
(sigma-plot), the "old" errors seem to be okay, whereas the errors in
Origin 4.1 seem to be too small. From the user's manual, it's clear that
Origin 4.1 also returns a 1-sigma error, so why the difference?
Has anyone else had similar problems?
Until I've solved this discrepancy, it renders the new version useless for me.

Other users might like to know that the problem pointed out by Clemens Woda
seems to have arisen between versions 3.5 and 3.7. V.3.7 and all subsequent
releases give the same (incorrect) error estimates when the Curve Fitter is
used with the weighting methods listed by Scott Bertrand.

We fully confirm this difference in error estimates and in fact I wrote to
Microcal about the problem in Sept 1996. Steve Roszko replied immediately: he
was able to reproduce the problem using a data set that I had sent him. In his
letter he stated: "It seems that Version 4.1 does not take into account the
error bar weighting when calculating errors......... This problem should be
fixed for the next available release or patch." However, patch 1 didn't provide
a fix: I haven't yet downloaded patch 2 so that I don't know if the fix is now
in place. Like you, I think that Versions 3.0 and 3.5 are providing valid error
estimates whereas Version 4.1 isn't: the estimates are much too low, as you
note. The discrepancy is striking when the data are weighted ("instrumental")
which is our usual choice. In my letter to Steve Roszko, my colleague David
Colquhoun and I asked for details of how the estimates of error are calculated
in Version 3.5 and 4.1, and we suggested that what might have happened was a
change (between versions)from the use of the observed to the expected
information matrix. As yet we have not received this information. The manuals
are sketchy on the issue (as is also true for all the rival products that I
have looked at).