Bookmark and Share

Notice: On March 31, it was announced that Statalist is moving from an email list to a forum. The old list will shut down at the end of May, and its replacement, is already up and running.

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: RE: RE: median equality test for non normal variables

From   Roger Newson <>
To   "" <>
Subject   Re: st: RE: RE: median equality test for non normal variables
Date   Tue, 25 May 2010 17:34:05 +0100

Yes, the Wilcoxon ranksum test calculates its P-value using a permutation method. Therefore, if we calculate a confidence interval for the Hodges-Lehmann median difference by inverting the ranksum test, then we are assuming that the 2 sub-population distributions differ only in location. This is in contrast to the -cendif- confidence interval for the Hodges-Lehmann median difference, which inverts a confidence interval for Somers' D, and which therefore contrasts to the usual Lehmann confidence interval formula as the unequal-variance t-test contrasts to the equal-variance t-test.

For a comparison of the performance of formulas for estimating the Hodges-Lehmann median difference, you might like to look at my presentation on this subject at the 2007 UK Stata User Meeting, downloadable from SSC at

Best wishes


Roger B Newson BSc MSc DPhil
Lecturer in Medical Statistics
Respiratory Epidemiology and Public Health Group
National Heart and Lung Institute
Imperial College London
Royal Brompton Campus
Room 33, Emmanuel Kaye Building
1B Manresa Road
London SW3 6LR
Tel: +44 (0)20 7352 8121 ext 3381
Fax: +44 (0)20 7351 8322
Web page:
Departmental Web page:

Opinions expressed are those of the author, not of the institution.

On 25/05/2010 17:04, Feiveson, Alan H. (JSC-SK311) wrote:
Isn't it true that the Wilcoxon rank sum test is designed only for possibilities of one distribution being a translation of the other? So the null would be identical distributions; the alternatives would be that the distributions differ only by a translation.

So if distributions have different shapes but the same medians one might naively assume the "null" is true, but as this example shows, such a condition will likely be rejected by -ranksum-.

Here's another example with continuous data:

One distribution is gamma(1,1), while the other is a reflection of the first plus a translation so that both have the same median.

drop _all
  set obs 100
  gen y=rgamma(1,1)
  summ y,det
  local med = r(p50)
  set obs 200
  gen group = 1 in 1/100
  replace group=2 in 101/200
  gen negy = -y[_n-100] if group==2
  replace y = 2*`med'+negy if group==2
  noi sum y if group==1,det
  noi sum y if group==2,det
  noi ranksum y,by(group)
  noi qreg y group

Note -ranksum- rejects its null (that the two distributions are identical, not that the medians are the same), whereas -qreg- accepts its null of equal medians.

Al Feiveson

-----Original Message-----
From: [] On Behalf Of Ronan Conroy
Sent: Tuesday, May 25, 2010 5:06 AM
Subject: Re: st: RE: RE: median equality test for non normal variables


There is an interesting question concerning the difference between
what people think they are doing when applying a 'nonparametric' test
and what is actually happening.

Consider the following data:

input var group
1 0
2 0
3 0
4 0
4 0
4 0
4 0
4 1
4 1
4 1
4 1
5 1
6 1
7 1

Note that the median coincides with the highest value in group zero
and the lowest value in group 1.

What we get now depends critically on what we ask for:

Test for equality of medians using -qreg- : P=1.000 (the medians are
the same)
Wilcoxon rank sum test : Prob>  |z| =   0.0196
Median test (which does not test for equality of medians, NB) :
Pearson chi2(1) =   3.8182   Pr = 0.051
Median test, continuity corrected : Pearson chi2(1) =   1.6970   Pr =
Ordered logit regression with group as a predictor : P =  0.997
'Harrell's C' (as calculated by -somersd-) : .76, P<  0.001

I have put quotes around Harrell's C, as this quantity is simply a
rescaling of Mann Whitney's U, dividing it by its maximum possible
value, and was first proposed by Richard Herrnstein in 1976
(Herrnstein, R. J., Loveland, D. H.,&  Cable, C. (1976). Natural
concepts in pigeons. Journal of Experimental Psychology: Animal
Behavior Processes, 2, 285-302), who termed it rho. Fans of
terminological chaos will also recognise the entity as the area under
the ROC curve. Harrell's C is identical with rho only when the data
are uncensored (James A. Koziol, Zhenyu Jia.T he Concordance Index C
and the Mann-Whitney Parameter Pr(X>Y) with Randomly Censored Data
Biometrical Journal 2009:51(3);467 - 474.)

I fancy that there is an amusing paper on this, clarifying the
hypotheses being tested in each case, if anyone has time to write one...

I am looking again at the t-test, which, after a couple of Kolmogorov-
Smirnovs, is beginning to look more and more attractive.

Ronan Conroy
Royal College of Surgeons in Ireland
Epidemiology Department,
Beaux Lane House, Dublin 2, Ireland
+353 (0)1 402 2431
+353 (0)87 799 97 95
+353 (0)1 402 2764 (Fax - remember them?)

P    Before printing, think about the environment

*   For searches and help try:

*   For searches and help try:
*   For searches and help try:

© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   Site index