Influence of epochs' lenght on psd plot


We use an egi headset made of 128 channels to collect eeg data (resting-state data) on patients with severe traumatic brain injury. We are actually pre-processing these data in order to do some connectivity analysis later.

During our pre-processing one step consist of epoching the data in order to perform an ICA to exclude artifacts such as eye blinks.

We were advised to epoch with a duration of 2 seconds, however it seems this is affecting a lot the psd plot. So we tried to extend the duration and it seems to decrease the effect on the psd.

Do you have any idea why reducing the duration of the epochs can influence the psd ?

This is the psd of the raw eeg data :

This is psd of the epoched data (2 seconds):

This is psd of the epoched data (10 seconds):

Thanks a lot in advance !!


Hello @victoriasus1, this is to be expected.

Firstly, we use a different approach for estimating PSD on epochs than we do on continuous raw data (multitaper on epochs, Welch FFT on raw), which can already yield different results.

So this can explain differences between raw data and epochs.

Secondly, the shorter the time periods to estimate PSD on, the “worse” the results will be in the lower frequency ranges, which is what you can see in your comparison between 2-sec vs 10-sec epochs: it’s especially below 10 Hz where the differences are really obvious. An algorithm just requires a longer observation to correctly identify “slow” oscillations, compared to “fast” ones.

There’s nothing you can really do about this, as we’re hitting a mathematical / physical limitation here. The only solution I can imagine is maybe to use longer epochs for doing the PSD estimation, and then cropping them to a shorter duration for your other processing steps.

1 Like