On Thu, 16 Oct 2003 18:38:20 GMT, DZ <[EMAIL PROTECTED]> wrote:

>Enda Kelly <[EMAIL PROTECTED]> wrote:
>
>> Hi. I have a query regarding whether it is logical to place a
>> confidence interval about a P-value. The computation involved uses a
>> permutation method to produce a P-value for a hypothesis test.
>
>Variance associated with the estimate of such P-value will be
>proportional to p(1-p) where p is the "true" p - which can be defined
>as p obtained with the infinite number of permutations.

That doesn't make sense.  Why would an infinite number of permutations
give a "true" p-value?  The p-value is the probability that data
following a hypothesized distribution would give a value of a
discrepancy measure that meets or exceeds the one you observed.  If
you have a point null hypothesis and it's true, then the p-value
should have a Uniform (0,1) distribution, no matter what the sample
size.  If the alternative is true and you consider it fixed while you
look at larger and larger sample sizes, then you'd expect the p-value
to tend to zero.

So in answer to the original query, no, confidence intervals for
p-values don't make sense.  Confidence intervals make sense when there
is some true fixed value and you're constructing a random interval
which has certain coverage properties in repeated sampling.  They
don't make sense conditional on one particular dataset.  


Duncan Murdoch
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to