Assessing significance of coherence values

,
1

Dear all,

I have a question related to the coherence computed with the bst_cohn function.
I am trying to assess coherence across two experimental task (i.e. coherence between the signal in Cz during Task 1 and during Task 2). My problem is to define if this inter-task coherence is significantly different from zero.

As far as I understand, the output parameter pValues gives in output a list of pvalues of size equal to the number of frequency bins.
I was wondering if there is a way to assess the significance of the coherence values by discrete frequency band (i.e. Alpha 8-13 Hz) and not by frequency bin.

Does someone has a suggestion about which is the correct way to achieve this?

I also considered running a bootstrap analysis by shuffling the original time series n times, and recomputing the inter-task coherence for each iteration to get a null distribution of the coherence values. Unfortunately, this become computationally too time-consuming (~100 h) on an ordinary computer.

Many thanks for your help,
Antonio

2

Your job of shuffling data, computing coherence on that to find a null-hypothesis case and comparing it with actual coherence values is pretty correct.

There is no reason that it takes 100h. Just select shorter time series (few seconds) and run it like 500 hundred times and then find the distribution by those values (either pull all to one or have a per element distribution).

3

Hi Hossein,

thanks for your answer. The reason it takes so much is that my recordings are 30 sec long and sampled at 500 Hz.

There is one thing that I don't understand from your suggestion. If I compute the coherence on shorter time lengths during the shuffling loop, will these values still be comparable with the original computed on the full time series?

The length of the recordings does not influence the accuracy of the coherence estimation?

Thanks,
Antonio