On Thu, 02 Dec 1999 09:00:30 -0500, Paige Miller <[EMAIL PROTECTED]> wrote: > > Can someone please point me to examples where Excel 95 or Excel 97 > produce incorrect results? Or perhaps post a few examples here? Thanks! Here is from a note that was posted Dec 1, 1999 on sci.stat.consult, by someone touting a particular set of add-ons to Excel -- ===========start extract < reformatted > from James Huntington However, we have recently been comparing Analyse-it (see http://www.analyse-it.com/) our statistics software for Microsoft Excel against competitor products. Our testing is based on the publically available NIST datasets, and our testing includes most of the popular packages like XlStat, WinStat, Minitab, SAS, SPSS and so on. We were shocked by the results! We'll be publishing details on our web-site soon, along with the datasets & results for all to see, but briefly: * Excel's built-in statistical functions like STDEV(), VAR() and CORREL() yeild inaccurate results in many cases. Most of the Excel add-ins use the above functions so propagate these errors. Analyse-it does not use **ANY** of Excels functions so is not affected. * Even well established packages that are taken for granted like SPSS & SAS, which were independently tested by the American Statistical Association, did not perform well in many cases. In a few cases even Excel's results are more accurate! You can read the reports online at http://www.amstat.org/publications/tas/mccull.pdf and http://www.amstat.org/publications/tas/mccull-1.pdf ==================end of extract. Those URLs are: American Statistician, May, 1999, for one reference that you can read in full, if you can read post-script files from the site above (I found it downloaded rather slowly), which details or cites references for various tests. The Longley data, which illustrates some problems for non-centered algorithms, uses data like 1001,1002, ... ,1007 to examine variances, and correlations/regressions with something practically similar (maybe, add 1.0 to a 1007, to make the correlation not-perfect). The SD, correlations, etc.,-- except for the means -- should be exactly the same as if you looked at 1,2,3,4,5,6,7 vs 1,2,3,4,5,6,8. Now, instead of 2 zeroes in a number like 1001, use 4 or 5 or 6 zeroes -- 1000001, etc. Six digit numbers will break a non-centering algorithm, if the computer doesn't compute with 11+ digits of accuracy. -- Rich Ulrich, [EMAIL PROTECTED] http://www.pitt.edu/~wpilib/index.html
