"H. D. Vinod" <vinod@murray.fordham.edu> writes:
> For those who may have missed The American Statistician article ,
> you may be interested to know that B. D. McCullough has assessed the
> numerical reliability of S-PLUS, SAS, and SPSS. The articles can be
> downloaded from the TAS website
http://www.amstat.org/publications/tas/
> It seems that the random number generator(RNG) for s-plus fails four
> of the DIEHARD tests and ANOVA fails SmnLsg9 problem Nonlinear
> regression fails to get good R-square for the Lancsos1 problem.
I think you mean the Residual Sum of Squares (RSS) for the nonlinear
regression problem, not R-square. If you check the Lanczos1 problem,
you will find that it is generated data with an "exact" solution. Of
course, in computer arithmetic the "exact" solution gives a small
residual sum of squares instead of 0 because of the round-off error.
The certified solution from NIST has a residual sum of squares of
1.4307867721E-25 from responses that range from 2.5 down to 0.06239.
Do you really care that the S-PLUS solution only had 3 significant
digits in the Residual Sum of Squares when the certified value is
10^(-25)? It's an artificial problem of the sort that is never
encountered in practice.
I have been corresponding a bit with Bruce McCullough about the
results for nonlinear least squares in S-PLUS. I have the impression
that he thought that the nonlinear least squares solutions from S-PLUS
were among the most reliable of the three packages compared. In all
cases the nls() function either produced a solution or correctly
indicated that it was unable to converge to a solution.
There is a slight inaccuracy in the description of the nls function in
the article. McCullough states that the convergence criterion is RSS
(residual sum of squares). In fact the criterion is based on the
orthogonality of the residual vector to the derivative vectors as
described in
@article{bate:watt:1981,
Author = {Bates, Douglas M. and Watts, Donald G.},
Title = {A Relative Offset Orthogonality Convergence Criterion for
Nonlinear Least Squares},
Year = 1981,
Journal = {Technometrics},
Volume = 23,
Pages = {179--183},
Keywords = {Regression}
}
-----------------------------------------------------------------------
This message was distributed by s-news@wubios.wustl.edu. To unsubscribe
send e-mail to s-news-request@wubios.wustl.edu with the BODY of the
message: unsubscribe s-news
|