Has the nonlinear regression in GraphPad Prism been tested against the NIST set of standard data sets?
The National Institute of Standards and Technology has created a set of challenging nonlinear regression problems to be used to test computer programs. We have tested GraphPad Prism with these problems. You can download these files, as one zip file (older version here).
For these tests, we set the options (in the weighting tab) to compute numerical derivatives using the slower, more accurate, method and to use the slower, stricter definition of convergence.
The NIST says that a good program should match its results to four digits of precision, and Prism 4 and meet this goal in all but three of the examples.
- Lanczos1. The best fit values of the parameters match to four signficant digits, but the standard errors do not. It is not possible to fully test Prism with the data set, because Prism only accepts data input with up to six or seven digits of precision, and the data in Lanczos1 has more digits than that. The fit is nearly perfect, and standard errors of all the parameters are tiny.
- Lanczos2. The data are fit almost perfectly by the model. Prism gives correct results for the first three digits, but is off for some parameters after that. The really tiny sum-of-squares leads to some round off errors.
- Lanczos3. One of the parameters matches to three digits, and the other parameters to four digits. Prism reaches a slightly smaller sum-of-squares than the NIST reports, so it seems that Prism's results are better than the ones reported by NIST.
April 2009 Update
Three of the files in the zip file were updated (MGH10, Ekerle4 and MGH17) for Prism 5. With Prism 5, the default setting is to give up if the fit hasn't converged in 1000 iterations. These are challenging fits (by design) so take more iterations. These files now allow 10,000 iterations. Also MGH10 was updated to use strict convergence criteria (as are the others; this one was mistakenly set to use less stringent criteria).