Pre-processing for FFT (or other ways of extracting the power spectrum of the signal)

Hi All,

I am a PhD student, learning how to analyse the EEG data collected from infants (so lots of noise, and a lot less usable trials). We use the EGI 128 channel hydro cap to collect the data, and I use the NetStation offered by the EGI and the EEGlab (and ERPlab when needed) to look at the data.

I have epoched data, with each epoch lasting 5 seconds, and I’m trying to run a power analysis (or a frequency analysis). I believe I should perform the FFT and see the power spectrum of the data, but I have a question regarding this.

So the question is;
Why do we need no baseline (and hence baseline correction) for the power decomposition of the data?

The background is;

I was told that the power analysis requires no baseline (and hence baseline correction), and I don’t think I understand why - even though i was given explanations multiple times. Could anyone advise me on this?

More specific to what I’m doing - so, like I said, I have multiple segments of which I want to use the power spectrum - more specifically, I’m interested in the averaged power over all the segments. What I thought I should be doing would be (1) create a STUDY on the EEGlab (i.e. import all the segmented data into one group), and then (2) compute the power. When I epoch the data I was asked by the EEGlab if I wanted to subtract the mean of channels - and that made me again wonder if we really need no baseline correction for the analysis I’d want to run, because I thought the amplitude of segments collected later would probably differ from the segments recorded earlier due to anything (which I believe is why we need a baseline correction for ERPs or time-frequency).

I am really a novice in this so I might be missing out some critical information for you to respond - if so please let me know!
And I apologise if this is too basic a question to ask. Thank you for your time to read this through - and I would appreciate any inputs from any of you.

Many thanks,

Hi Saya. Baseline normalization is often useful but rarely necessary. And it’s not always possible. Furthermore, baseline normalization is possible only for task- or stimulus-related experiments. If you have resting-state/spontaneous activity, then there is no time window you could use as a baseline. You could refer to this video for more discussion.

Is there a reason why you have 5-second epochs? That seems a bit long if this is resting-state. Most people do 2-second epochs. Then you would compute the power spectrum of each segment separately, then average those power spectra together (“Welch’s method”). If you have questions about how to implement this in eeglab, you’ll need to post the eeglab list.

Computing the average over channels is about re-referencing; that’s a separate issue from baselining. Average reference is a good idea for 128 channels.


Hi Mike,

Thank you so very much for your answer. That’s really kind of you - and I have to say I’ve been a massive fan of your books and online videos!

Could you please elaborate on why it is not always necessary to do a baseline normalisation? I do understand it’s not always possible but from your book and the video there certainly are benefits of doing the baseline normalisation when possible so I find myself struggling to understand why it is okay not to do so when we could? That said, it is hard to do the baseline normalisation on my data due to the nature of the setup but if there are advantages that you mentioned in your book and video of doing the baseline normalisation even for the power analysis, I’d definitely be trying to have baseline period during my data collection.

To answer your question - this is for us to investigate what happens in the infant brain while they are doing a standardised task procedure, which in our study is Piagetian A not B task. Basically, infants see an adult hiding an object in one location and are encouraged to search the toy after it’s hidden, but there is typically a 5-sec gap between when they see the object hidden and when they are allowed to search the toy (e.g. the adult hide the toy in one of the two boxes on the table, and they pull the table away from the infant for 5 seconds and then push it towards the infant again so they could reach out for the toy). This study was designed by a behavioural psychologist who doesn’t run an EEG study by themselves but they were interested in neural measures as well so they put the cap on infants while they were doing the behavioural task. So to be honest, it is not specifically designed for collecting EEG data, and I am analysing as I was asked.

But I wasn’t aware of Welch’s method (as I have only done ERPs for my own research) and I guess they were not either, so I’ll certainly ask them if they’d be interested in me trying Welch’s method. Thank you very much for the suggestion, I so much appreciate it!

Many thanks and very best wishes,

OK, so if I understand your experiment, you have three phases: stimulus, delay, search. My first thought for doing the analyses would be to compute power spectra from these different periods and then show them on the same graph, so three lines.

An ideal result would be that the three lines overlap for much of the spectrum and diverge in one or a few narrow frequency ranges. In this way, you would evaluate the power changes for each condition (time window) relative to the other conditions. You don’t need to worry about a baseline normalization.

Hi Mike,

Thank you for your suggestion! That does make sense. What I might worry is that the durations of those three period are different, so I guess maybe I could look at each period (e.g. split each period into, say, 1-sec epochs, and get the average over those epochs as a score for each stage), and compare across three stages - if that makes sense?

Anyway thank you so very much for your help. I really do appreciate it!

Yes, that makes sense. I wouldn’t be too concerned about different durations of the trial periods, as long as they are roughly comparable. If it were 2 seconds vs. 3 hours, then that’s a concern. But a difference of some seconds is unlikely to bias the results either way, particularly if you have enough data.