All this discussion about failure of binomial confidence intervals to give
"exact" coverage also applies to the Fisher "exact" test, whose actual level
(probability of rejecting the null hypothesis of equal proportions, when in
fact the proportions are equal) is usually less than the nominal level,
depending on the true proprtions. In the frequentist setting, it's the same
problem - there are only a finite number of possible outcomes.
Al Feiveson
I was wondering about that. So, is there also a raging controversy over
whether some alternative to Fisher is superior, e.g. Yates correction for
continuity? Like Nick Cox said in an earlier post, it sounds like "exact"
is more of a propaganda term than an accurate description of the
test. (Kind of like saying you've got the "best" product on the market.)
