What is the difference between mean(A,'omitnan') and nanmean(A) ?

Hi, I have a vector A (or could be a matrix) which contains finite values as well as NaNs. I want to compute the mean value of its non-NaN elements. What is the difference between mean(A,'omitnan') and nanmean(A) ? Thank you!

Answer 1

[Stephen23]

There is no practical difference for your vector.

Personally I would use mean, because nanmean is part of the Statistics Toolbox, so using it makes your code less portable.

[LeChat]

ok. It is the feeling I had. Thanks a lot!

[Nathan Williamson]

On a related note, in my code I found median(A,'omitnan') to run significantly faster than nanmedian(A). Not sure why. In the same code, mean(A,'omitnan') and nanmean(A) have a similar runtime.

[Andrew Frane]

Looking at the code for the most recent (2020) version of nanmean, it's just a wrapper-function that executes the mean function with 'omitnan'. So it should perform nearly as efficiently.

But looking at the code for the most recent (2020) version of nanmedian, for some reason it doesn't just execute the median function with 'omitnan'. Instead, it just uses the prctile function to calculate the 50th percentile, which isn't a very computationally efficient way to calculate the median.

Incidentally, if A is a vector, (or if you want a single grand mean of a matrix or higher-dimensional array), you could also do something like the following, which avoids both the back-compatibility problem with 'omitnan' and the toolbox-dependence problem with nanmean and has similar efficiency:

mean( A(~isnan(A)) )

Answer 2

[Walter Roberson]

nanmean() existed first, as part of the Statistics Toolbox.

Later 'omitnan' was added as an option to the regular mean() as part of basic MATLAB. nanmean() should now be used mostly for backwards compatibility.