Re: st: RE: Can I have different results from Stata13 depending on the OS?

[Daniel Feenberg <feenberg@nber.org>]

On Tue, 21 Jan 2014, David Kantor wrote:

> Hello,
>
> I just want to add that I once had a situation in which I got different 
> results in Stata MP vs. a single-processor version (of the same generation, I 
> believe). I don't recall the details, but in the end, we decided to trust the 
> Stata-MP result. Stata Technical Support explained that it had to do with MP 
> making assumptions about what order some of the intermediary results could be 
> computed -- or what could be concurrently. I could dig up the details of this 
> matter, but at this point, the significant point is that, sometimes, MP can 
> yield different results as compared to the single-processor version. 
> Potentially, your two Statas could differ in that regard.
> --David

My guess is that 99.44% of the time a large difference in estimated 
coefficients is due to an accidental change in the data that could be 
detected by running summarize before the estimation command. Or perhaps 
the sort order is significant and has changed. If that isn't the case, 
then consider if the difference is statistically significant. I think 
Albert Beaton is the author of a JASA article almost half a century ago 
pointing out that very small inaccuracies in arithmetic can substantially 
affect the results of a regression only if the estimation is already very 
imprecise for statistical reasons, but small errors can't make imprecise 
results seem precise. Both PC and Mac versions of Stata use ieee 
arithmetic, so there is very little chance of a major arithmetic error in 
the hardware. I believe even the decimal to binary conversion is 
controlled by the standard (I think) so that traditional source of 
difference is likely gone.

The Beaton article was a response to the then famous Longley critique of 
statistical packages. Longley provided a dataset that gave different 
results in different packages. This was long before Stata existed. Many 
readers of the Longley piece were worried that all of quantitative social 
science was based on arithmetic errors, but that wasn't the case, as 
Beaton showed. Any author of a package now must use double precision ieee 
arithmetic or be subject to (possibly unjustified) ridicule.

Daniel Feenberg
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/