Query on Granger Causality Dominant Frequency vs. PSD

Support & Discussions
1

Hello everyone,

I’ve been determining the net Granger causality (GC) between two signals using the spectral_connectivity_epochs function, following a process similar to the one in this example. My GC results show a dominant frequency of 5.0 Hz, but this doesn’t align with the power spectral density (PSD) of the signals, where the dominant frequency is around 1.0 Hz. I’m wondering if the GC results can still be considered valid, and how to interpret this apparent discrepancy.

Here are some additional details:

  • The two signals were normalized (mean = 0, std = 1) before GC calculation.
  • Highpass filter of the raw data: 0.0 Hz
  • Lowpass filter of the raw data: 58.0 Hz
  • Sampling frequency: 116.0 Hz
  • Parameters for GC calculation:
    • fmin=0.5
    • fmax=20
    • rank=(np.array([1]), np.array([1]))
    • gc_n_lags=20

Any insights on whether the GC results are reliable, and how I should infer the relationship between the GC and PSD, would be greatly appreciated!

Thanks in advance for your help!

2

Hello,

With Granger causality, we’re quantifying “directed connectivity” based on how well a signal helps to predict the activity of another signal.

Now you may have two signals with very strong activity at 1 Hz, but that alone doesn’t mean there is some interaction which would allow us to predict the activity of a given signal.

Alternatively, you may not have a super strong PSD peak at 5 Hz, but if there is some interaction between the signals then what (relatively little) activity is there could have a very valuable predictive ability.

All in all it’s often tricky to try and impose some 1:1 relationship between power spectra and connectivity spectra, so I wouldn’t be too quick to discount the validity of what you’re seeing.

Cheers,
Thomas