0
$\begingroup$

I was digging in the scipy code for this test Wilcoxon signed-rank test (stats.Wilcoxon) and I found that in scipy they compute the sum of the ranks for the differences that are positive and separately for the ones that are negative. Then they picked the smaller one and use that as W. That is substantially different from the test explanation in Wikipedia, or other sites where W = sum(all_differences).

Is this approach valid?

$\endgroup$
5
  • $\begingroup$ What is your question? $\endgroup$
    – whuber
    Jun 19 '13 at 14:30
  • $\begingroup$ I would like to know if the scipy implementation is wrong from the statistical point of view. Thanks for the help $\endgroup$ Jun 19 '13 at 14:47
  • $\begingroup$ Because the two statistics are interchangeable--when you know one, you can determine the other (from the sample sizes)--there are just two issues here: (1) what does the scipy documentation claim about $W$? and (2) does scipy produce correct p-values? Have you checked either of these? $\endgroup$
    – whuber
    Jun 19 '13 at 14:51
  • $\begingroup$ 1) scipy use this W to compute the p_value. Actualy the process is the same as decribed in Wikipedia, except for the computation of W. 2) How can I check this? $\endgroup$ Jun 19 '13 at 14:57
  • $\begingroup$ This is not an error. There are multiple equivalent ways to define most rank based statistics; if they yield the same p-value, they're the same test. If you want to check it, compare it with something that defines the statistic the other way and compare p-values for a case where they both do it exactly. (Alternatively, you can usually just work the algebra and see they are computing equivalent things and so must get the same p-value) $\endgroup$
    – Glen_b
    Jun 20 '13 at 0:07
2
$\begingroup$

First, wilcoxon test in scipy.stats does NOT use $W$ as the test statics, it instead uses $T$ as defined in Siegel's popular book: Non-parametric statistics for the behavioral sciences. And yes, as @whuber correctly pointed out, once you know $T$ and sample size, $W$ is also defined (@whuber, strictly speaking, not quite, one also need to know how 0 differences are handled).

Only can only know how the test is implemented by reading the source code. For scipy, Wilcoxon test can be found in your_python_package_folde/scipy/stats/morestats.py. Compare to R's wilcox.test, it is very simple. Go over the code, and you will see that it is equivalent to having correct=FALSE, exact=FALSE, paired=TRUE flags on in R.

Python:

>>> from scipy import stats
>>> x1=[48,  7, 12, 11, 62, 93, 79, 53, 28, 49, 74, 59, 57, 62, 22,  8, 30, 11,  2, 47]
>>> x2=[20, 13, 41, 61, 93, 11, 28, 61, 26, 91, 95,  5, 80, 45, 88, 99, 50, 96, 69, 93]
>>> stats.wilcoxon(x1, x2) # T and p value, two-sided
(60.0, 0.092963126712486244)

in R:

> x1<-c(48,  7, 12, 11, 62, 93, 79, 53, 28, 49, 74, 59, 57, 62, 22,  8, 30, 11,  2, 47)
> x2<-c(20, 13, 41, 61, 93, 11, 28, 61, 26, 91, 95,  5, 80, 45, 88, 99, 50, 96, 69, 93)
> wilcox.test(x1,x2,correct=FALSE,exact=FALSE,paired=TRUE)

    Wilcoxon signed rank test

data:  x1 and x2 
V = 60, p-value = 0.09296
alternative hypothesis: true location shift is not equal to 0 
$\endgroup$

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.