The collection of functions in the NaN Suite enhances the functionality of the original 'nan'-functions in the Statistics Toolbox for N-D matrices. In addition, the new function nansem.m calculates the standard error of the mean ignoring NaNs.
The NaN Suite functions are used similar to the original functions (e.g. mean, std, var), that is they take a matrix as the first input, and the dimension along which the operation should be computed as the second input. For example, nanmean(A,4) calculates the mean of A along the 4th dimension of A ignoring NaNs.
Sorry for the scarce fantasy but could not find better words than those expressed by David Redish below:
"Thank you for fixing this. We have been bitten regularly by Mathworks' strange decision to define nanxyz(all-nans) as 0, which propogates nans incorrectly. This suite has helped us tremendously."
woops someone already pointed this out. Ignore comment below
Hi I think there is a bug in nanmedian - i have a dataset (single row), DS, which contains no NaN values, but nanmedian(DS) = NaN. Matlab's normal median(DS)=0.
I think this is the problem - when the median actually equals zero, nanmedian thinks it =nan.
Just what I was looking for :-)
Haven't looked at why, but seems to run a little slowly.
The issue Binu found is also present in the nanvar function. where:
NaNsuite y = nanvar(x,dim,flag)
MATLAB y = nanvar(X,flag,dim)
Thanks for the contribution!
There is a bug in the order by which the Nan suite calls the 'dim' argument. This matters if you have replaced the matlab 'NaN' functions with the Nan suite.
NaNsuite y = nanstd(x,dim,flag)
MATLAB y = nanstd(X,flag,dim)
y = std(X,flag,dim)
I'd give five stars if Andrei's bugfix were included. Will upgrade to 5* if the bug is fixed :)
Great package, thanks!
Also agreeing on the bug as well as fix found by Andrei Terebilo on 09 Oct 2009. Would be nice if the downloadable package could be fixed.
Thank you so much, this is useful!
Little bug here (otherwise great, thanks):
>> nanmedian([0 0 0 0 0 1])
Just modify line 71:
i = find(y==0);
i = find(s==0);
is there any function to do quintiles ignoring NaN? Thanks!
This is a great suite of functions. I did notice the max/min toolbox functions in M7 filtered NaN's from the calculations. However, the other functions did not. So, that's where the NaN suite came in handy.
Thank you for fixing this. We have been bitten regularly by Mathworks' strange decision to define nanxyz(all-nans) as 0, which propogates nans incorrectly. This suite has helped us tremendously.
Isn't life as simple as this:
nanmean(a) = mean(a(find(~isnan(a))))
Great minds think alike, I see. I was just about to write out this code myself when I realized someone out there must also be using nan as a non-measurement, and want it ignored while computing average and standard deviation.
thanks a lot
This happened to be just what I needed.
so useful -- thank you!
included an additional function nanvar.m which computes the variance for N-D matrices. It is very similar to nanstd.m without taking the square root.
bug fix in nanmedian. The function now returns NaN for an entire vector of NaNs.
bug fix in nansum.m
Create scripts with code, output, and formatted text in a single executable document.