NaN in ranksum test

Hi,

I know there is a nanmedian function, but it seems there is no nanranksum function. I have tried the ranksum function with and without NaN values filling in empty data and I get different p values. Am I wrong? Is there a Wilcoxon ranksum / Mann-Whitney U test that ignores NaN values?

Thanks for your help!

Answer 1

[Daniel Golden]

I have the same problem in Matlab R2012a. My version of the Matlab ranksum() documentation doesn't mention anything about NaNs, but the latest (R2013a) documentation at http://www.mathworks.com/help/stats/ranksum.html states, "ranksum treats NaNs in x and y as missing values and ignores them." This is obviously not true in R2012a, where there is probably a bug in the treatment of NaN values. For example try the following:

>> ranksum([1 2 3 nan], [4 5 6])

ans =

      0.0571

>> ranksum([1 2 3], [4 5 6])

ans =

      0.1000

Obviously, the NaNs are not being ignored.

Other times, NaN inputs will result in NaN outputs:

K>> ranksum([-14.44 NaN 5.97 -117.55 -77.56 -45.00], [-78.59 -101.04 -26.15 -79.51 -48.10 -23.45 -42.18 -76.75 -55.42 -135.18 70.02 -57.44 -31.69 -146.01])

ans =

     NaN

But this isn't consistent. For example, removing any one of the vector values in the above example, even the non-NaN values, will result in a non-NaN output.

Here's a simple workaround if your inputs might have NaNs. If your input vectors are x and y, and you're running ranksum like:

p = ranksum(x, y)

Then just run ranksum like this:

p = ranksum(x(~isnan(x)), y(~isnan(y)))

Answer 2

[Açmae]

@ Daniel and Eric:

In R2012b and beyond, the test:

p = ranksum(x(~isnan(x)), y(~isnan(y)))

is performed in the function RANKSUM to remove any missing data, and thus takes care of NaN's. If you are using versions older than R2012b, then this is the workaround.