2
$\begingroup$

I am performing a Wilcoxon signed rank test on my dataset. Because the sample size is quite small I use exact p values (as described here.)

Somewhere else I once read that the p values associated to W can be doubled for two-tailed tests. (e.g. n = 6; W = 4 -> one-tailed: p = 0.11 or two-tailed: p = 0.22). Reading the article mentioned above I noticed that it is possible that the one-tailed p value is 0.5 (e.g. n = 6; W = 10). Following my assumption this would result in p = 1.0 for a two-tailed test.

Isn't it impossible for p to become exactly 1 since there is always a remaining probability for the null hypothesis to be wrong? Or maybe I am just doing a simple logic mistake.

$\endgroup$
2
$\begingroup$

It's easily possible for a p-value to be exactly 1 if you have a discrete test statistic (as here).

Isn't it impossible for p to become exactly 1 since there is always a remaining probability for the null hypothesis to be wrong? Or maybe I am just doing a simple logic mistake.

Your question here implies that you think that 1-p is the probability the null is wrong (and presumably then that the p-value is the probability the null is true); this is not a correct understanding of a p-value.

A p-value is the probability of getting a test statistic at least as extreme as the one you got when the null is true. In the case of the signed rank test it's easy to have the signed pair differences centered* at zero, so that the are no values for the test statistic that are less extreme than the one you observe.

plot of signed rank statistic showing "at least as extreme" for w=4 and w=10 with 6 pairs

There's no value for the test statistic less extreme than 10 -- it's as consistent with the null as it's possible to get. Since every possible value is at least as extreme as the observed statistic, the p-value is 1.

Somewhere else I once read that the p values associated to W can be doubled for two-tailed tests. (e.g. n = 6; W = 4 -> one-tailed: p = 0.11 or two-tailed: p = 0.22).

If you have a symmetric distribution of the test statistic and the one-tailed p-value is no more than 0.5 -- but a one-tailed p-value could be 0.6, or even as much as 1.0 ... you don't double those. (You subtract them from 1 and then double.)

* (in the way that the signed rank statistic measures center)

| cite | improve this answer | |
$\endgroup$
  • 1
    $\begingroup$ An explanation for the downvote would be helpful -- in what way is this answer wrong or inadequate? I'd like to improve it if there's some issue with it. $\endgroup$ – Glen_b Dec 14 '17 at 23:54

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.