[R] small bug in binom.test?

[ripley@stats.ox.ac.uk]

On Wed, 22 Jan 2003, Jerome Goudet wrote:

> At 22.01.2003  08:45 +0000, you wrote:
> >Why do you think that?
> >
> >The problems binom.test(49,50,0.5) and binom.test(51,100,0.5) are
> >symmetrical, so one would expect the same results for a two-sided test.
> >
> >The problem I guess is how a two-sided test is defined for a discrete
> >distribution.  For one-sided tests one would use the probability of X >=11
> >or X <= 9, and those are not equal.  For a two-sided test the code
> >attempts to find a point in the opposite tail with at least as large a
> >tail probability, and adds on that tail probability.  Thus for
> >binom.test(11,100,p=0.1) it used P(X < 9 || X >= 11), and for
> >binom.test(9,100,p=0.1) it used P(X <= 9 || X > 10), if I followed the
> >code right.

(I typed it wrong.  The first omits (9,10), the second (10,11).)

> I would have defined the two sided test as P(X<=9 || X>=11) (checking of 
> course that if the two values are equal, the probability is not counted 
> twice). Is this wrong?

Yes.  Let me repeat: `why do you think that?'.   (R implements standard 
statistical definitions, not Goudet's ones, so I expected you to give me a 
reference.)  Did you look up the references on the help page?

For continuous distributions, two-tailed tests are defined as equi-tailed,
and so your recipe is nothing like what happens for e.g. a gamma
distribution.  For discrete distributions there are complications but
there is a standard definition of an alpha-level test, which can be
converted to a p value via searching for possible alpha-level tests.

-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595