« The Upside of Data Cleaning | Main | Blog Forum - The New Legal Realism: It’s Not About Breakfast »

16 June 2006


Chris Zorn

Good point, Michael. And, as Jeff implies, version differences are often a less significant issue than differences between packages. For interested people, a nice resource that covers this issue and a host of others is Altman, Gill, and McDonald's (2003) "Numerical Issues in Statistical Computing for the Social Scientist" (John Wiley & Sons). A couple sample chapters and some other related materials are available here:


Among other things, they show how, thanks to issues with numerical precision and formulae, a number of software packages (including Excel!) can't even calculate a standard deviation correctly. (For a hint at possible implications of this for Sarbanes-Oxley enforcement, see http://hardware.slashdot.org/article.pl?sid=06/06/05/185201 ). Sobering stuff.

Jeffrey Segal

I recall years ago that SPSS used to add 5.0 to the constant of probit equations. This was before maximum likelihood procedures were perfected, and there was apparently some value in keeping the cumulative normal distribution above 0. But if that was the reason, why they couldn't subtract the 5.0 after the calculations was never clear. Needless to say, folks who didn't RTFM and didn't notice this undoubtedly made massive mistakes.

Oh yeah, for logit models, they added 5 to the constant and then divided other coefficients by 2 "in order to make them comparable to probit coefficients."

The comments to this entry are closed.


May 2018

Sun Mon Tue Wed Thu Fri Sat
    1 2 3 4 5
6 7 8 9 10 11 12
13 14 15 16 17 18 19
20 21 22 23 24 25 26
27 28 29 30 31    

Site Meter

Creative Commons License

  • Creative Commons License
Blog powered by Typepad