Thirty years ago, Longley (1967) calculated by hand the solution to a linear regression problem and found that many mainframe programs produced incorrect results. Since that time, statistical and econometric software has progressed far beyond linear regression, but testing strategies to assess software reliability have not made similar progress. Recently, McCullough (1998) proposed a methodology for assessing the numerical reliability of statistical software in three areas: estimation, random number generation, and statistical distributions (e.g. for calculating pvalues). Estimation is assessed using the Statistical Reference Datasets (StRD, available at http:// www.nist.gov/itl/div898/strd), a collection of accuracy benchmarks compiled by the National Institute of Standards and Technology (NIST). Marsaglia's (1995) DIEHARD program applies tests of randomness to the output of a random number generator (RNG). Both KnuÈ sel's (1989) ELV program and Brown's (1998) DCDFLIB program can test the accuracy of statistical distributions.
The methodology will be applied to four econometric software packages: EViews 3.0, LIMDEP 7.0 for Windows 95, SHAZAM 8.0, and TSP 4.4, some details of which are provided in Table I. No attempt is made to assess or compare these packages on any basis other than computational; previous reviews can be referred to for evaluation of other aspects. Speci®c details of implementing this methodology will be addressed cursorily, since they are discussed elsewhere (McCullough, 1998(McCullough, , 1999)).
EViews 3.0, designed for Windows 95, is heavily oriented toward pull-down menus. A previous version was recently reviewed by Sparks (1997). LIMDEP 7.0 for Windows 95 supersedes LIMDEP 7.0 for DOS, which was reviewed by Wright (1996). While the DOS version was single precision storage/double precision calculation, the new version is double precision for both.