[FieldTrip] Phase coherence analysis

Thu Mar 7 19:53:42 CET 2024

Hi Christina,

As you know, I would never use single trial coherence estimates, because these estimates are way too biased.  But I could imagine that you are thinking of a multitaper estimate (i.e. compute the coherence on a long single trial using very heavy taper smoothing), or use a Welch type of estimation (cut the long trial into many overlapping segments, perform FFT, compute coherence). This could work provided your trials are sufficiently long.

Then, if I read your questions well, you want to combine coherence at - say 1.4 Hz - in trial 1, with coherence at - say 1.6 Hz - in trial 2? To me, naively, this sounds a bit fishy, because what does this metric actually aim to quantify?

Thinking out loud, however, you may be able to get away with it. Here are some thoughts. From the theory we know that coherence is (the normalised) cross spectral density function, where the latter is the frequency domain version of the cross-covariance function (which in the time domain quantifies as a correlation-like measure the linear relation between two time series as a function of time lag between the two. Now, if you have time series with a clear relationship based on band-limited periodicities of the underlying time series, it makes sense to go into the frequency domain for sensitivity, and the coherence spectrum will show a peak at the relevant frequency. (if on the other hand you would compute coherence between a time-shifted white noise time series and the same not time-shifted version, you’d get a flat coherence spectrum, with high values of coherence all over the spectrum). Now, if this is all taken at face value, one could aggregate coherence estimates across trials at the per trial peak frequencies, because it indeed quantifies the relationship between the dependent time series (EEG) and a periodic driver (the audio envelope). Whether the reviewers of your paper are going to like this, I am not sure.

An alternative to consider (which also would probably confuse a reviewer) could be to keep the phase difference information per trial (i.e. rather than coherence per trial, compute the cross-spectral density + power), and only take the absolute value (and normalise with the product of the power) after you aggregate across trials. Here you would probably need to phase shift the phase differences as a function of frequency before they are aggregated, in order to mimick the frequency domain equivalent of a fixed latency shift in the time domain, but it might buy you some signal-to-noise (and it would reduce the bias in the estimate, if done properly).

Either way, I think that in a written report the approach should be motivated well.

With respect to your question 2) I don’t see a problem.

Best wishes,
Jan-Mathijs

On 27 Feb 2024, at 03:28, Christina Vanden Bosch der Nederlanden via fieldtrip <fieldtrip at science.ru.nl<mailto:fieldtrip at science.ru.nl>> wrote:

Dear FieldTrip!

My team is working on a project where we are trying to see if phase coherence is calculated while people listen to a busy scene with 4 sounds playing at the same time. We wanted to examine whether the sound participants were told to listen to (attended) had greater phase coherence between the EEG signal and the attended sound's envelope compared to coherence between the EEG signal and the other, non-attended, sounds in the scene.

Is it a violation of the assumptions of the ft_connectivity analysis pipeline I use to calculate phase coherence if I calculate phase coherence 1) on a single trial basis and 2) multiple times using the same EEG signal but different envelopes of the sounds in the scene? Of course, we still have 100s of trials (500-700), but the goal is to log the phase coherence values for each trial and each pairing for the envelope's peak frequency (all the sounds have different peak modulation frequencies, they are not just speech sounds) and group the data at a later time according to coherence (at each stimulus' peak mod freq) for attended and unattended, perhaps calculated as a difference score from when that envelope was attended or not.

I know the phase coherence metric evaluates the consistency of coherence at particular frequencies across trials of the same or similar stimuli, but we really can't do that here since the peak frequencies are so different across each stimulus type. So I am wondering if there are big issues with calculating the data on a single trial basis and grouping the data after that.

Thanks for your insights!

Christina
--
Christina M. Vanden Bosch der Nederlanden
Assistant Professor
Department of Psychology
University of Toronto Mississauga
christinavb at gmail.com<mailto:christinavb at gmail.com>
Website: https://www.utm.utoronto.ca/lamalab/<https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fsites.google.com%2Fview%2Fchristinavbdn&data=05%7C02%7Cfieldtrip%40science.ru.nl%7Cdad1c1c1f3614356c9e108dc3ed7e915%7C084578d9400d4a5aa7c7e76ca47af400%7C1%7C0%7C638454344243407458%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C0%7C%7C%7C&sdata=e%2FRbS9sjwax225t%2BmDaDCNsa8wxE4ZJLUK2S4bZHqJ8%3D&reserved=0>
_______________________________________________
fieldtrip mailing list
https://mailman.science.ru.nl/mailman/listinfo/fieldtrip<https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fmailman.science.ru.nl%2Fmailman%2Flistinfo%2Ffieldtrip&data=05%7C02%7Cfieldtrip%40science.ru.nl%7Cdad1c1c1f3614356c9e108dc3ed7e915%7C084578d9400d4a5aa7c7e76ca47af400%7C1%7C0%7C638454344243407458%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C0%7C%7C%7C&sdata=RI%2B1i0K8OKk53q3W65mTiNnQo0g%2FuEH1Baz8b6euB7c%3D&reserved=0>
https://doi.org/10.1371/journal.pcbi.1002202

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20240307/5cb08260/attachment.htm>