This dataset is one of several published by the National Institute of Standards and Technology, NIST, for the express purpose of testing statistical software. In the published archive, this example is labelled Roszman1 and is included in the Nonlinear Regression category.

Here, y is the number of quantum defects observed in iodine atoms and x is the excited energy state. As a software test, this example is considered of average difficulty in spite of the fact that there are a number of local optima very close to the global optimum shown in the figure.

Model

y = A - B x - atan (C/(x - D))/Pi

Parameters

The indicated optimization was carried out using unweighted least-squares [R-squared = 0.99841]. This assumes that the residuals (deviations from the curve along the y-axis) are Gaussian, with zero mean and the same variance at all points. However, residuals are often not Gaussian but Laplacian (double-exponential). If this is the case, then the least-squares criterion is invalid and the ML parameters must be found by minimizing the average magnitude of the deviations (residuals). Standard regression packages do not offer this option but Regress+ does. When this new criterion is employed, we get new ML parameters, as follows:

Alternate Parameters

With the minimum average deviation criterion, R-squared decreases to 0.99838 but the average deviation is 3.24272e-03 whereas it was 3.28547e-03 with the least-squares criterion.

Needless to say, Regress+ is ideally suited to assessing whether or not a set of residuals is Gaussian or Laplacian (or something else!) As it happens, the residuals from this least-squares fit are acceptably Gaussian and acceptably Laplacian, probably because the dataset is not big enough to show the difference.

This dataset is one of the examples included in the Regress+ software package.