3
$\begingroup$

I have a question about the joint distribution of the mid p-value and p-value.

We know that, for right tailed test with discrete test statistic $X$ with distribution $F$, the p-value is defined as $P=Pr(X \geq observed~X)$ and the mid p-value is defined as $mid~P=Pr(X \gt observed~X)+\frac{1}{2}*P(X=Observed)$.

I would like to find out $Pr(P_{mid} \leq t~and~P \gt t)$ for some $t \in (0,1)$. Here $P$ is the p-value and $P_{mid}$ is the mid p-value.

$\endgroup$
6
  • $\begingroup$ What test do you have in mind? $\endgroup$
    – whuber
    Aug 10 '12 at 23:13
  • $\begingroup$ Thanks so much for the reply. I am hoping to get a general solution for any discrete distribution F. $\endgroup$
    – user13154
    Aug 11 '12 at 1:06
  • $\begingroup$ The answer depends strongly on the test statistic, not just the distribution: that's why you need to tell us what test you're using. $\endgroup$
    – whuber
    Aug 11 '12 at 21:05
  • 1
    $\begingroup$ Ok. We can focus on the Fisher's exact test. Thanks. $\endgroup$
    – user13154
    Aug 12 '12 at 2:31
  • 1
    $\begingroup$ Simple translation of the question: "I would like to find out values of the hypergeometric distribution." $\endgroup$
    – whuber
    Aug 14 '12 at 14:05
1
$\begingroup$

It is well known that the distribution of $P_{\rm mid}$ more closely approximates a uniform distribution than that of $P$.

$P$ is obviously always slightly larger than $P_{\rm mid}$. If $X=x$, then the difference between the two p-values is $$P-P_{\rm mid}=P(X=x)/2$$ On average this is $$E(P-P_{\rm mid})=\frac12 \sum_x P(X=x)^2.$$ In general, if $X$ can take on many possible distinct values, so that $\max_x P(X=x)$ becomes small, then $E(P-P_{\rm mid})$ will also be small.

If $P(X=x)$ is a decreasing function of $x$, as it usually is for $x$ sufficiently large, then $P-P_{\rm mid}$ will also decrease with $x$.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.