Full_Name: Marketa Kylouskova
Version: 1.2.1
OS: Windows
Submission from: (NULL) (147.231.7.2)
Hi,
I believe there is a bug in the extension of wilcox.test, namely in
determination of confidence intervals. I concentrated on the two-sample test
only. I examined and tested its performance on this working data set:
x<-rnorm(10,3,1) # So I have two samples, I will test whether
y<-rnorm(10,0,1) # x is greater than y, i.e. one-sided alternative “greater”
say
x<-c(3.770684,4.654342,3.496403,1.772743,1.624953,2.645835,3.099477,1.706758,3.507709,1.982924)
y<-c(-0.8161288,0.1632923,0.6421997,1.9270846,-0.4668112,0.3587806,0.3312529,-0.5393900,
0.1057892,1.7963575)
then
wilcox.test(x,y,alternative="greater",conf.int=T)
gives
Wilcoxon rank sum test
data: x and y
W = 94, p-value = 0.0001624
alternative hypothesis: true mu is greater than 0
95 percent confidence interval:
-3.344416 NA
which is apparently contradictory (p-value is much less than 0.05 while exact CI
contains 0). I examined the code and I found that while the exact wilcoxon test
is applied on ranks of x (thus testing that mu = x – y is greater than 0), the
exact CI is computed based on differences y - x, hence finding a CI for mu = y –
x). The same applies, in my opinion, to approximative CI, where he optimization
interval is set as (mumin,mumax) with
mumin <- min(y) - max(x)
mumax <- max(y) - min(x)
I didn’t go through other aspects, although I believe that the continuity
correction should be checked if some changes are made, and also that some closer
look onto definitions of uci and lci in the exact CI computations should be made
(here it depends on your definition of the CI for exact distributions, i.e.
whether the CI = {uci<=mu<lci} (then I think it’s OK) or CI =
{uci<=mu<=lci}
(then I think lci= diffs[ql] and not diffs[ql + 1]). Could this be specified in
the documentation?
Thank you in advance for checking my suspicions :) and for your reaction.
Marketa Kylouskova
kylouskova@euromise.cz
-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-
r-devel mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html
Send "info", "help", or "[un]subscribe"
(in the "body", not the subject !) To:
r-devel-request@stat.math.ethz.ch
_._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._
>>>>> kylouskova writes:> Full_Name: Marketa Kylouskova > Version: 1.2.1 > OS: Windows > Submission from: (NULL) (147.231.7.2)> Hi,> I believe there is a bug in the extension of wilcox.test, namely > in determination of confidence intervals. I concentrated on the > two-sample test only. I examined and tested its performance on this > working data set:> ...> which is apparently contradictory (p-value is much less than 0.05 > while exact CI contains 0). I examined the code and I found that while > the exact wilcoxon test is applied on ranks of x (thus testing that mu > = x – y is greater than 0), the exact CI is computed based on > differences y - x, hence finding a CI for mu = y – x). The same > applies, in my opinion, to approximative CI, where he optimization > interval is set as (mumin,mumax) with> ...Thanks for pointing this out. The confidence interval was for the difference in locations of y and x and should have been the other way round. Fixed now. -k -.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.- r-devel mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html Send "info", "help", or "[un]subscribe" (in the "body", not the subject !) To: r-devel-request@stat.math.ethz.ch _._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._
> > I didn’t go through other aspects, although I believe that the continuity > correction should be checked if some changes are made, and also that some closer > look onto definitions of uci and lci in the exact CI computations should be made > (here it depends on your definition of the CI for exact distributions, i.e. > whether the CI = {uci<=mu<lci} (then I think it’s OK) or CI = {uci<=mu<=lci} > (then I think lci= diffs[ql] and not diffs[ql + 1]). Could this be specified in > the documentation?the CI is [uci, lci) with lci = diffs[ql + 1]. You do not reject the hypothesis if W <= ql and the step function is equal to ql for all real numbers between diffs[ql] and diffs[ql + 1] and jumps in diffs[ql + 1] only. Therefore, all this points have to be included in the CI but not diffs[ql + 1] itself (one can however include it, making the CI 'larger'). Taking lci = diffs[ql] is wrong! See the example on page 688 of Bauer's article. Torsten> > Thank you in advance for checking my suspicions :) and for your reaction. > > Marketa Kylouskova > kylouskova@euromise.cz > > > -.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.- > r-devel mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html > Send "info", "help", or "[un]subscribe" > (in the "body", not the subject !) To: r-devel-request@stat.math.ethz.ch > _._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._ >-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.- r-devel mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html Send "info", "help", or "[un]subscribe" (in the "body", not the subject !) To: r-devel-request@stat.math.ethz.ch _._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._