Validation

Introduction

CurveExpert Basic is validated against the Statistical Reference Datasets Project of the National Institute of Standards and Technology. These datasets can be downloaded directly at http://www.itl.nist.gov/div898/strd/general/dataarchive.html, but are also included verbatim in the CurveExpert Basic distribution for you to use yourself. The datasets validated against are all of the sets in the “Linear Regression” and “Nonlinear Regression” categories.

Running the Suite

You can validate CurveExpert Basic against these datasets simply by choosing Tools->Run NIST Validation. All validation cases, with all “starts”, will be run. The NIST test suite provides different starts (i.e., initial guesses) for each nonlinear regression validation. CurveExpert Basic covers all of these starts in the test suite, and the results are output to the message window.

Some competing packages do not include all starts in their validation sets, but CurveExpert Basic shows the results of all of them. At the time of this writing, there is one failure (dataset BoxBOD, Start #1), where, for the given start, the initial correction to parameter causes underflow in the model, and from this there is no recovery.

Performing validations yourself

If you would like to perform validations against the test data yourself, all data files are contained in the “example_data/NIST” folder in the initial data directory (see Where files are located). Or, you can always download the files from NIST at the link above.

Note

The NIST datasets all list the dependent variable (usually y) before the independent variable (usually x). So, when you read the files, select Rotate Columns in the file import dialog, or alternatively, rotate the columns after you read in the dataset.

All nonlinear regression validation runs are performed in the automated suite with rtol=ptol=1.0e-14, and maxiter=1000.