trouble in signrank function for R2011b ?

Hello,

When calling signrank(x) in the R2011b version, 0s values in x do not seem to be omitted anymore as in previous releases. These values should indeed be ignored when computing the signed rank statistics ( http://faculty.vassar.edu/lowry/ch12a.html ).

Let me give two examples, with and without 0s, which should theoretically give the same result if I understood well:

signrank([-5 -6 -7 -8 -4 2 -10])

ans =

    0.0312

whereas

signrank([-5 -6 -7 -8 -4 2 -10 0 0 0])

ans =

    0.0820

... And the result becomes different depending on whether the bias is for positive or negative values, while it should not be, because the test should be two sided: indeed:

signrank(-[-5 -6 -7 -8 -4 2 -10 0 0 0])

ans =

    0.0156

Using a previous version of matlab (R2007b) as well as the statistical package R (<http://www.r-project.org/>) with the same vectors, I saw that:

1) all the p-values stay close to the p-value given by R2011b for the vector where 0s are manually removed (first example I gave)

2) the p-values they give are independent on the sign of the bias, as it should be.

Could this be a bug, or am I doing something wrong ?

Answer 1

[Tom Lane]

This does look like a bug. If you are comfortable editing code in your MATLAB installation, I recommend editing signrank.m to change this

t = (abs(diffxy) < epsdiff);

to this

t = (abs(diffxy) <= epsdiff);

Alternatively you can do an explicit comparison with zero:

signrank([-5 -6 -7 -8 -4 2 -10 0 0 0],0)

Thanks for reporting this. I'll see that it gets fixed as soon as possible.