Question

I am getting NaNs when I use perform correlation between two matrices, but I want to ignore them, and I read on some MATLAB forum that replacing NaNs with the means obtained by nanmean is a pretty good replacement. So I executed the following code, but Im still getting NaNs in my result. Why is that happening?

cor1nan=nanmean(correlations1);
mod1cor=isnan(correlations1);

for z=1:264
for a=1:264

 if(mod1cor(a,z)==1)
     correlations1(a,z)=cor1nan(z);
 end
 end
 end

Just noticed that some of the values in cor1nan are NaNs themselves.. how I go about solving this?

Answer 1

If you have columns that are all NaN, then nanmean will keep it as NaN if I read it correctly. Then your loop below will substitute in a NaN, thus keeping them.

A suggestion for how to solve it is as follows. Replace your code with:

  cor1nan=nanmean(correlations1);
  corAllnan = nanmean(cor1nan);
  cor1nan(isnan(cor1nan)) = corAllmean;
  mod1cor = isnan(correlations1);

  for z = 1:size(correlations1,2),
         correlations1(isnan(correlations1(:,z)),z) = cor1nan(z);
  end 

This will replace the NaNs in cor1nan with the mean of the non-NaN values in cor1nan