Dear Allin Thanks, It works both on my XP and Ubuntu machines. I found that there is still a minor problem when sample size <9 (maybe due to no appropriate approximation of z distribution with small sample size), but it is excellent enough. Yi-Nung Yang Associate Professor and Chair International Business Chung Yuang Christian University, Taiwan. 2011/4/5 Allin Cottrell <cottrell(a)wfu.edu&gt;:

On Wed, 29 Jun 2011, yinung at Gmail wrote: No, but it is easy to calculate. <hansl> nulldata 30 series x = normal() series y = normal() # for comparison difftest x y --rank-sum # calculate rank-sum manually matrix Z = ({x} ~ 1) | ({y} ~ 0) Z = msortby(Z, 1) ~ seq(1,rows(Z))' w1 = sumc(selifr(Z[,3], Z[,2])) </hansl> Allin

On Sat, 2 Jul 2011, yinung at Gmail wrote: The problem is actually the ties. (If the series are of different lengths you just put the shorter one first.) This fixes the ties issue: <hansl> nulldata 9 matrix mx = {1,2,3,4}' matrix my = {1,3,5,7,9}' matrix Z = (mx ~ 1) | (my ~ 0) Z = msortby(Z, 1) series s = Z[,1] matrix r = ranking(s) Z = Z ~ r Z w1 = sumc(selifr(Z[,3], Z[,2])) </hansl> However, the need to go back and forth between matrices and series is not very elegant. In CVS I have now enabled the ranking() function for vectors. So this will now work: <hansl> matrix mx = {1,2,3,4}' matrix my = {1,3,5,7,9}' matrix Z = msortby((mx ~ 1) | (my ~ 0), 1) w1 = sumc(selifr(ranking(Z[,1]), Z[,2])) </hansl> Allin