**Date:** Wed, 2 Nov 2005 13:55:40 -0500
**Reply-To:** Richard Ristow <wrristow@mindspring.com>
**Sender:** "SPSSX(r) Discussion" <SPSSX-L@LISTSERV.UGA.EDU>
**From:** Richard Ristow <wrristow@mindspring.com>
**Subject:** Re: Precision of the results
**In-Reply-To:** <S229461AbVKBLop/20051102114448Z+90574@avas-mr07.fibertel.c om.ar>
**Content-Type:** text/plain; charset="us-ascii"; format=flowed
A side issue, but it's something everybody should be aware of

At 06:44 AM 11/2/2005, Hector Maletta wrote:

>1. Precision of SPSS results is always the same (15 decimals). What
>varies is the visible number of decimals, which is established through
>the FORMAT command.

A small thing: actual precision is 52 bits; 15 decimal digits is quite
close. (SPSS uses the IEEE standard 754 64-bit floating-point
representation. So does most software: it's a good format, and it's
supported by floating-point hardware on just about all machines.)

However, that's the precision of SPSS *numbers*. Precision of
calculated *results* is never this high, and can be much lower. Never
assume you have 52 accurate bits, just because there are 52 bits in the
number.

Appendix - side issues
----------------------

The numerical precision now used is high (52 bits), much higher than
that of any data we'll see, and algorithms are well-designed now. You
can forget about precision problems, in almost all uses of SPSS.

(As late as the 1970s, though, I got bit by an imprecise calculation in
SPSS release 6, maybe 7, for the mainframe.)

You CAN have precision problems in unusual cases, such as a correlation
matrix that's almost singular. Those problems should be re-specified.
Software won't always catch them, but be careful of any message that
talks about loss of precision, or inability to converge. And, as
always, know your data.