db-Correcting Subjects When Creating an Average Condition Analysis

[Christopher Clifford]

Dr. Cohen,

I've read your book on analyzing neural time series data, and I'm trying to apply some of the techniques for my own analysis.  I'd like to do an analysis of power-based connectivity differences across subjects exploring two different frequency bands for a single time window (trying to find an overlap between two different erp components to infer functional connectivity). I'm looking to apply this analysis to three different conditions in my data.

In order to do this, I'm attempting to apply a trail-averaged baseline corrected power normalization (db) to each subject, then sort each subject's different trials into aggregate "condition sets" where I can then try and look at the time series of the correlation between the two different electrode montages to see if my time window will be a good fit for analyzing the individual differences in the correlation.  Does this sound like a proper way both to apply normalization and power-based connectivity analysis?  I've done analysis to typical measures of ERP before, but something like this has a bit more parts so I would like to make sure I've got everything straight. 

Thank you for all the books, lectures, and code you have provided!

[Mike X Cohen]

Hi Christopher. Do you mean that you will correlate the power values at each time point over trials for each condition? If so, you don't need to apply a baseline normalization. Normalizations tend to produce unexpected effects at the single-trial level. Plus, the correlation itself is a normalized measure. 

Also keep in mind that time-frequency power has reduced temporal precision compared to the ERP because of the filtering. So it's possible that two neighboring ERP peaks will get mixed together in the time-frequency plane. 

Mike

--
You received this message because you are subscribed to the Google Groups "AnalyzingNeuralTimeSeriesData" group.
To unsubscribe from this group and stop receiving emails from it, send an email to analyzingneuraltimes...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/analyzingneuraltimeseriesdata/2be96849-bccf-4fd8-b402-a6947a1f4c04%40googlegroups.com.

--

Mike X Cohen, PhD
Fresh look: mikexcohen.com

[Christopher Clifford]

Dr. Cohen,

What I mean to say is I will be correlating the averaged power per stimulus type per subject at all time points in the specified latency window.  In this way with three different stimulus types presented in my experiment during the viewing for every subject, I will be averaging power only from like-stimulus trials and then ultimately have three different power time series per subject.  I would only be applying the normalization to this averaged stimulus type power.  However, since the correlation is a normalized measure, I suppose that it wouldn't matter.  Does the first part sound like an appropriate way to measure power-based connectivity?

Thank you for the point on the ERP peak mixing.  Do you have a recommended way to test/look for this occurring other than visual inspection of the data?

Thank you for your time,

To view this discussion on the web visit https://groups.google.com/d/msgid/analyzingneuraltimeseriesdata/CAABvs%2BufUc1DeZNc2U%3Dk%2BFFBYCSy97hPO%2BZjP_7F-HPaBZV4cQ%40mail.gmail.com.

--

Christopher Clifford

Doctoral Student

Florida International University, Department of Psychology

Brain and Behavior Development Lab

[Mike X Cohen]

I see. I think the potential concern for power correlations over time is the time window. Because the power time series is smooth and has only low-frequency characteristics, a correlation over a short window of time will be unreliable, and inflated due to the temporal autocorrelation.

To be honest, I don't think there is a good solution here. Power time series correlation is designed for long windows. But do you need to compute power connectivity? Phase synchronization over trials will give higher temporal precision, although it can still be tricky to interpret in the time of the ERP, because strong but independent phase reset can give rise to spurious synchronization values. Perhaps one idea is to look at phase synchronization in time windows defined by the ERP but only in the non-phase-locked part of the signal.

Please call me Mike, btw. I got my PhD double-digit years ago, yet somehow I have never gotten comfortable with being called "Dr. Cohen." The only worse epithet is "professor Cohen"  ;)

To view this discussion on the web visit https://groups.google.com/d/msgid/analyzingneuraltimeseriesdata/CAOUye8peCgX1nMTWuSx3pUOphR8_gKu81%2Biw8e8_jMDKEz%2BPVg%40mail.gmail.com.

[Christopher Clifford]

Mike,

I see what you mean regarding the time window.

Maybe it would be better if I took a step back to explain what I'm trying to look for, and perhaps you could advise me on the approach if you wouldn't mind.

I'm looking at two ERP components in some infant data collected at 9 months.  I'm looking at the n290 (a sort of infant-precursor to the n170) and NC.  I've done some scalp plots over time to pinpoint the montage sites I want to use (central-occipital and central-parietal respectively) and so now I'm hoping to see if I can find a way to determine if the n290 activity has any kind of functional connectivity with the NC activity.  Would you still advice at trying to look at the phase synchronization in the combined time window? 

Thanks,

To view this discussion on the web visit https://groups.google.com/d/msgid/analyzingneuraltimeseriesdata/CAABvs%2BsktQazQMLO5DFfPjCeaxRXuRBCYYGfXh%3DL6b1%3Db%3D6myg%40mail.gmail.com.

[Mike X Cohen]

Hi Chris. I don't know what an NC is, but I gather that it's later in time and maximal at a different electrode? Perhaps you can try defining ERP time windows and electrodes based on the trial-average, and then get the time-domain data from each trial within those windows, and then correlate the two. It's a pretty simple approach, but seems to address your question. Two things come immediately to mind:

1) It might be useful to compute the energy in the time window instead of the average voltage. That will help account for polarity flips. You can compute energy as, for example, root-mean-square.

2) It might be a good idea to take a third time/space window as a control. If the entire brain is more energetic on some trials (e.g., because of attention), then the N290 and NC might be spuriously correlated, and having a third region (e.g., a frontal channel?) can be used to partial out this unrelated shared variance.

My other suggestion is to look through the literature and see how other people have tried to solve this problem. It's been almost 15 years since I've done anything with ERPs except for looking at them as quality-control indicators, and I generally don't teach, write about, or advise on ERP analysis.

Mike

To view this discussion on the web visit https://groups.google.com/d/msgid/analyzingneuraltimeseriesdata/CAOUye8oX69yWfUn7v9DDEujfzzxihNJ2aOSzB8_EjYpxZF1sgQ%40mail.gmail.com.