I have a question that’s been bugging me about the data I choose to use for a PSD analysis (e.g., Welch method).
If I have resting-state data (or any EEG data for that matter) that I’ve previously epoched during preprocessing (e.g., to remove noisy EEG segments etc.), I’m wondering whether it’s appropriate to then re-join it into a single long segment for the purpose of running a PSD?
The potential benefit of this would be that I could then use larger windows for the PSD to run across (as opposed to being limited to 2 or 3 second epochs). However, I’m unsure whether having some discontinuity in the data (i.e., as some epochs would have been removed during cleaning) might affect the PSD algorithm and produce spurious results? My feeling is that it shouldn’t have much of an effect, as I can’t see how it would alter the spectral power in any way? But maybe I’m missing something?
Anyway, just curious if those with more experience in this area might be able to shed any light on this!
Hi Brian. Epoching resting-state data is a bit arbitrary. The benefits include smoothing the spectrum and minimizing cancelation from nonstationarities. But you don’t have to cut the data into epochs.
As for concatenating the epochs: yes, that will create edges at the boundaries of non-adjacent epochs. It’s difficult to say whether that will be really detrimental, because it depends on the size of the edges, which in turn depends on the data at the boundaries. Also, the discontinuities are a relatively small proportion of the total signal. It’s analogous to how outliers have less negative impact in large compared to small datasets.
Anyway, my advice is just to try it both ways! Plot the spectra on top of each other and see how much they differ. Keep in mind that any time you change an analysis parameter, the results will necessarily change numerically; the important question is whether you would draw a different conclusion about the data based on the analysis parameter.
Thank you Mike, yes that makes sense!
I tried plotting some of these to visually compare both approaches, and the results are very similar (e.g., attached example from one subject with the top figure showing the result of the PSD after first concatenating all epochs, and the bottom running over each 3s epoch separately (and then averaging).
Very nice. Qualitatively the same.