| CC: | bug-Statistics-ttest [...] rt.cpan.org |
| Subject: | Divide by zero in fdist module |
| Date: | Wed, 25 Jun 2008 16:27:10 -0400 |
| To: | bug-Statistics-Distributions [...] rt.cpan.org |
| From: | b.e.buck [...] clearwire.net |
line 325/326 sub _subf2 has data vulnerabilities when called from t-test. It
can wind up trying to take log's of negative numbers or dividing by zero if $n
(the sample size) is sufficiently large that ($n - 2) * log($x) winds up being
very negative, resulting in an exponent of a negative number. $x is frequently
negative such that this can happen easily for large $n. Module 1.02 of
Statistics::Distributions PERL 5.84
A trivial fix is to return $x = 1 for $z=>0 and $x = 0 for $x=>negative, but
that may not be the desired behavior. You'd have to be much better at
statistics and understanding what this part of the f-test calculation is
actually *doing* to know if that's the right behavior.
Thanks for any help on this, -- Ben