Review of GAUSS for Windows, including its numerical accuracy
✍ Scribed by H. D. Vinod
- Publisher
- John Wiley and Sons
- Year
- 2000
- Tongue
- English
- Weight
- 157 KB
- Volume
- 15
- Category
- Article
- ISSN
- 0883-7252
No coin nor oath required. For personal study only.
✦ Synopsis
The importance of Aptech's (1999) GAUSS programming language for applied econometrics is widely accepted. A fresh look is needed due to the rapidly changing computing environment involving higher speeds, lower costs, Windows and the Internet. Accordingly, this review focuses on (1) GAUSS for Windows version 3.2.37, and (2) the numerical reliability of GAUSS.
GAUSS has ¯exibility, power, and intuitive syntax, ease-of-sharing, and ease-of-use. An impressive array of free GAUSS software is available on the American University (1999) and The Econometrics Journal (online) (1999) web sites. It is so rich because many researchers who develop new econometric methods choose GAUSS for testing and implementation. While much of this software is written for the DOS version of GAUSS, it need not be rewritten to run on Windows. Rust (1993) complained about the slow (7-minute) startup time for Windows due to virtual memory manager initialization problems. The version reviewed here loads very fast.
Our review focuses on the numerical accuracy of GAUSS. We ®nd that the algorithms used by GAUSS sometimes fail rather badly to provide accurate results. We blame both the algorithms and the language itself. See McCullough and Vinod (1999, Section 2) for discussion of how a language can cause inaccuracies through poor handling of errors, under¯ows, etc., and how in¯uential the choice of estimation algorithms can be. We use the methodology proposed by McCullough (1998) to assess statistical software on three fronts: estimation, random number generation, and statistical distributions. McCullough (1999a, b) and McCullough and Wilson (1999) uncover numerous ¯aws by applying this methodology. We use Internet-based NIST (1999) benchmarks for Statistical reference data sets (StRD) to evaluate the numerical accuracy of GAUSS algorithms.
After discussing some basics of GAUSS syntax in Section 2, we describe our important results on the numerical accuracy of GAUSS in Section 3. Our review of the numerical accuracy of nonlinear methods in GAUSS uses Constrained Optimization (CO) module sold for extra cost. It