If you have a basis for assigning weights to each data point, doing so lets
you make use of additional knowledge about your measurements, e.g., take into
account that some points may be more reliable than others. That may affect
the final values of the parameters.
Weighting the data provides a basis for interpreting the additional `fit`
(Note:fit ) output after the last iteration. Even if you weight each point
equally, estimating an average standard deviation rather than using a weight
of 1 makes WSSR a dimensionless variable, as chisquare is by definition.
Each fit iteration will display information which can be used to evaluate
the progress of the fit. (An '*' indicates that it did not find a smaller
WSSR and is trying again.) The 'sum of squares of residuals', also called
'chisquare', is the WSSR between the data and your fitted function; `fit`
has minimized that. At this stage, with weighted data, chisquare is expected
to approach the number of degrees of freedom (data points minus parameters).
The WSSR can be used to calculate the reduced chisquare (WSSR/ndf) or stdfit,
the standard deviation of the fit, sqrt(WSSR/ndf). Both of these are
reported for the final WSSR.
If the data are unweighted, stdfit is the rms value of the deviation of the
data from the fitted function, in user units.
If you supplied valid data errors, the number of data points is large enough,
and the model is correct, the reduced chisquare should be about unity. (For
details, look up the 'chi-squared distribution' in your favourite statistics
reference.) If so, there are additional tests, beyond the scope of this
overview, for determining how well the model fits the data.
A reduced chisquare much larger than 1.0 may be due to incorrect data error
estimates, data errors not normally distributed, systematic measurement
errors, 'outliers', or an incorrect model function. A plot of the residuals,
e.g., `plot 'datafile' using 1:($2-f($1))`, may help to show any systematic
trends. Plotting both the data points and the function may help to suggest
another model.
Similarly, a reduced chisquare less than 1.0 indicates WSSR is less than that
expected for a random sample from the function with normally distributed
errors. The data error estimates may be too large, the statistical
assumptions may not be justified, or the model function may be too general,
fitting fluctuations in a particular sample in addition to the underlying
trends. In the latter case, a simpler function may be more appropriate.
You'll have to get used to both `fit` and the kind of problems you apply it
to before you can relate the standard errors to some more practical estimates
of parameter uncertainties or evaluate the significance of the correlation
matrix.
Note that `fit`, in common with most NLLS implementations, minimizes the
weighted sum of squared distances (y-f(x))**2. It does not provide any means
to account for "errors" in the values of x, only in y. Also, any "outliers"
(data points outside the normal distribution of the model) will have an
exaggerated effect on the solution.