Systems estimation: A comparison of SAS, SHAZAM and TSP

The Accuracy of Econometric Software

“…First, Calzolari and Panattoni (1988) had already proven that they could write benchmark-quality code when they provided a benchmark for FIML estimation. These FIML results were used successfully by Silk (1996) to find flaws in packages and were also easily reproduced by software developers. Having mentioned one method of system estimation, for completeness we note that Zellner and Thornber (1966) long ago remarked on different packages giving different answers to two-and three-stage least squares estimates of Klein's model I, and there is apparently no published resolution of the problem.…”

Econometric Software

“…In fact, once a software package is chosen, even in the best case the user is almost completely dependent upon the specific provisions made by its developer. If only simple linear techniques are considered, the difference between the textbook world and the less-than-ideal real world may not always be significant, but in the case of nonlinear estimation, in particular, the gap can be wide indeed [172,173,219].…”

“…Indeed, economic journals normally reject as not germane articles that focus too obviously on specific computational issues, even when critical to the evaluation of the published findings. A possible exception to this is numerical accuracy [175,253], but only recently and still in token form [34,38,154,167,169,170,172,174,219,229,231,232]. To date, economists deal with computational issues at a full arm's length and depend upon others, often statisticians and the statistics literature, to probe the details [10,122,155,220,239], even if some related issues have been considered within the bounds of econometrics [16,17,21,22].…”

The Numerical Reliability of Econometric Software

“…XXXVII (June 1999) provided a benchmark for FIML, including the asymptotic covariance matrix thereof computed by various methods. See Julian Silk (1996) for a discussion.…”