Granger causality on oscillation power

Hi Mike,

I have a few questions regarding granger causality.
If I conduct granger causality on the power (or the phase) extracted from time-frequency decomposed EEG data, is the result interpretable? How different this method is compared with the frequency domain granger causality described in your book Analyzing neural time series? Is it possible that the results analyzed using these two different ways are not consistent, for example, the directions are opposite?

Thank you for you time!

Hi Lovenix. The right way to do this is to apply GC on the broadband time series data and then look for rhythmicity in the autoregression coefficients. I write about this in my book, and it’s also implemented in GC toolboxes, for example the MVGC toolbox from Anil Seth.

If you want to measure this on the phase angle time series, you can use phase transfer entropy, which is basically the same thing as GC under certain assumptions (near-Gaussian distribution of the time series data, if I remember correctly).

Thank you very much!

Hi again Mike,

I’ve read about the GC chapter in your book. I know that I can use the MVGC toolbox to do the frequency domain GC and applying GC on the filtered signal is not appropriate. But I still would like to know why using the power time series of a specific frequency band from time-frequency decomposition to compute GC (using the normal GC instead of the frequency domain GC) is not right.


The power time series is a highly smooth version of a signal. So the effect will be to increase the number of parameters required to fit the autoregression model to the data. Power time series also remove the phase information, which is the crucial part of the signal that contains the exact timing of the fluctuations. That’s exactly what the autoregression model needs to latch onto :wink:

Thank you very much for your explanations!