How do I use nancov on a a matrix that has multiple consecutive rows of NaN in some of the columns?

I am trying to use nancov on a matrix that has multiple consecutive rows of NaN in some of the columns, and nancov simply returns a matrix full of zeros. (The NaN values are due to gaps in the data.) Any suggestions for alternatives or ways to get around this?

[the cyclist]

Are you able to show us the data (or a small subset) that exhibit the issue? Also, why do you think the result is incorrect? Could it be that the covariance is truly zero?

[Jenny]

Below is an example matrix, when I have situations like column 2 or 3, it doesn't work. When I have situations like column 1, it works fine. I think the result is incorrect because if I manually change the NaN values and run it again, nancov works, by which I mean I don't get a matrix full of zeros. Also, I have other data sets like this that only have situations like column 1 and they work fine.

58.28 0.5798 7.833 -60.29 0.768 2.14

-4.235 NaN 68.08 NaN 0.9708 -6.435

-0.5155 NaN -4.611 0.4154 NaN 0.32

NaN NaN NaN 30.5 NaN 96.01

0.4329 -0.2091 NaN 8.744 4.387 NaN

-22.59 37.98 0.0592 150 -49.83 NaN

Answer 1

[the cyclist]

After removing the observations (i.e. rows) that have any NaNs in them, you are left with only one row. You can't take the covariance of a single observation.

I don't fully understand why MATLAB creates output of all zeros. (It seems like an output of NaN would be more appropriate.) But that's the underlying cause.