st: numeric accuracy

[Stas Kolenikov <skolenik@gmail.com>]

I run a certain, rather extensive, microsimulation project on several
flavors of Stata (MP8 down to SE) and try to benchmark my results. On
some of the numeric outcomes I compare (percentiles of a
distribution), I get differences between different flavors as large as
300*c(epsdouble). The categorical outcomes always match, and some
other numeric outcomes (mean of a distribution), the difference is
~2*c(epsdouble). But the factor of 300 bothers me. I wonder what the
typical constants are in front of the c(epsdouble) that are used in
certification scripts. When I compared results of my -confa- with
-gllamm-, I was happy to find the difference to be of the order 1e-4.
But these are different estimation methods; here I am running exactly
the same code on exactly the same data, and I hoped to see a close
reproducibility from one run to another.

-- 
Stas Kolenikov, also found at http://stas.kolenikov.name
Small print: I use this email account for mailing lists only.
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/