Garry Anderson <g.anderson@unimelb.edu.au> found the following bug:
>
> I am enquiring why the value of Fisher's exact P for a 2 x 2 table should
> decrease when there is a smaller sample size for each of the two samples
> and the two proportions remain at 100% and 0%?
>
> For example
> -csi 6 0 0 6,e-
> 1-sided Fisher's exact P = 0.0011
> 2-sided Fisher's exact P = 0.0022
>
> -csi 5 0 0 5,e-
> 1-sided Fisher's exact P = 0.0000
> 2-sided Fisher's exact P = 0.0000
<stuff deleted>
>
> Any suggestions would be appreciated.
>
> (SPSS gives P=0.1 for the 3 0 0 3 combination)
>
> I am using Stata 8.2, 30 Jan 2004, ado 11 Mar 2004.
Al Feiveson also noted this was also a problem with -tabi-.
I have fixed this bug, and the fix will be out in the next executable
update. It was restricted to the Windows version of Stata.
The problem (if anyone cares) was a subtle one and restricted to
these extreme cases -- numbers can be held in a computer at higher
precision than they are stored at. Some compilers treated two numbers
differently, holding one at a higher precision than the other, when
the two numbers should have been exactly the same. Disaster
occurred when these were subtracted and a value of zero was not
obtained.
Some results:
. csi 5 0 0 5, e
<stuff deleted>
1-sided Fisher's exact P = 0.0040
2-sided Fisher's exact P = 0.0079
. csi 3 0 0 3, e
<stuff deleted>
1-sided Fisher's exact P = 0.0500
2-sided Fisher's exact P = 0.1000
. tabi 5 0 \ 0 5, exact
<stuff deleted>
Fisher's exact = 0.008
1-sided Fisher's exact = 0.004
Hope that helps!
--Jean Marie
jlinhart@stata.com
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/