# How to configure PSD and TFR methods applied on the delta and alpha band in EEG data

**URL:** <https://mne.discourse.group/t/how-to-configure-psd-and-tfr-methods-applied-on-the-delta-and-alpha-band-in-eeg-data/5838>\
**Category:** Support & Discussions\
**Tags:** eeg\
**Created:** [October 31, 2022, 4:01pm UTC](https://mne.discourse.group/t/how-to-configure-psd-and-tfr-methods-applied-on-the-delta-and-alpha-band-in-eeg-data/5838 "2022-10-31T16:01:14Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![mscheltienne](https://yyz2.discourse-cdn.com/free1/user_avatar/mne.discourse.group/mscheltienne/32/827_2.png) [@mscheltienne](https://mne.discourse.group/u/mscheltienne)\
**Post date:** [October 31, 2022, 4:01pm UTC](https://mne.discourse.group/t/how-to-configure-psd-and-tfr-methods-applied-on-the-delta-and-alpha-band-in-eeg-data/5838/1 "2022-10-31T16:01:14Z")

</div>

Hello,

I’m still looking into details about the (temporal-)spectral analysis I can apply to EEG data, and for now, I am focusing on the delta-band (1, 4) Hz and alpha-band (8, 13) Hz on long epochs (8s, 16s or 24s). The dataset is sampled at 512 Hz and filtered between 1. and 40. Hz.

### PSD

#### multitaper

From my limited knowledge, `multitaper` method provides a more robust spectral estimation than welch’s periodograms, but is more computationally expensive. But the `bandwidth` smooths together frequencies in a large range. The default, 4 Hz, will smooth together frequencies at ± 4 Hz around each frequency. Is this potentially an issue when looking at low frequencies, namely delta and alpha bands? Would you recommend a different `half-bandwidth` value?

#### welch

If I choose to go this way instead of multitapers, as the lowest frequency I am interested in is 1 Hz, I guess the welch’s segment length should be set to a 1-second minimum. Is this assumption correct? Would you set it to more, e.g. 2 seconds?

With a 1-second minimum, on 8 seconds epochs, I could get 8 segments without overlap, or way more with overlap. Besides the added computational load, increasing the number of segments via a large overlap should improve the spectral estimate, correct?

### TFR

This time I am interested in 24 seconds long epochs. This dataset is from a neurofeedback study, thus the temporal resolution isn’t crucial as there isn’t a stimuli onset. I have the temporal resolution set to 2 seconds, i.e. `n_cycles = 2 * freqs`.

For the methods, I experimented mostly with multitapers for now, but would you recommend using Morlet wavelets for low-frequencies?

For the frequencies of interest, I have it set at `np.arange(1., 15., 1)`:

- 1 Hz because this is the high-pass cut-off frequency.
- 15 Hz, because I was aiming for 13 Hz (upper limit of my defined alpha-band) + half-bandwidth (with multitapers). Is this a good idea, or would you provide a different range of frequencies of interest?  
Also, what should impact the decision about the granularity of the frequencies of interest? 1 Hz? half-bandwidth in the case of multitapers? Something else?

This is already a lot of questions on the computation part, I haven’t found a consensus on settings to use for those common bands (other than the defaults), and I am a bit struggling in figuring out what should weigh my decisions.

Looking forward to some input!