Choosing the appropriate baseline correction in TF analysis with Morlet wavelet

[leontion]

Dear Mike, 

I wonder if you could give me some advise with the following: 

I have a question regarding the selection of baseline correction in TF analysis with Morlet Wavelet. 

My EEG data are based on a cognitive experiment with a rather odd design and I 'm quite confused about the baseline I should use.

I have followed all the steps from the book and the lecturelets but I believe the [-500 -200] baseline won't work in my case because the epochs are extracted based on self-paced responses of the subjects.

Subjects in the experiment were exposed to 1 min blocks of the stimulus, during which they had to produce as many responses as they could by pressing a key (that is used as the time 0 in epoch extraction). I need to get -1250 to 0 with  TF analysis. 

I think that if I get a baseline preceding each epoch, in many cases it will be affected by the task activity from the previous epoch. There is a fixation cross interval between each 1 min block so I was wondering if it is a good idea to extract a baseline from there. In this case all the epochs extracted from the same block (number varies between 1 to 15 epochs per block) will have the same baseline I guess.  

I also have 2 min of fixation cross recorded before the main task begins (from an  eyes open - eyes closed session). Maybe I could extract a baseline from there or it's a bad idea (last task block is like half an hour later from eyes open session)? 

Best regards, 

Mina

PS Thanks for the uploaded lecturelet!

[Mike X Cohen]

Hi Mina. A few ideas come to mind. 

You could try not doing any baseline normalization (in this case you should normalize the wavelet in the frequency domain before point-wise multiplying with the data). If you are not looking for power suppressions, this might be OK. 

If you have multiple conditions, you could use a condition-average baseline. That way, any differences will still be apparent in the baseline period. 

You could use the entire epoch as the baseline. I usually don't like this approach because sustained changes in power will dominate the baseline estimate. 

I think taking the baseline from a separate and long rest period is generally not a good idea. There are many differences in cognitive set between task and rest.

I hope that helps. Sorry I can't give a simple answer ;)  If you are still stuck, perhaps you can attach a picture of a TF plot so we get a better sense of what the results look like. Sometime that helps give more specific advice.

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.
Visit this group at http://groups.google.com/group/analyzingneuraltimeseriesdata.
For more options, visit https://groups.google.com/d/optout.

--

Mike X Cohen, PhD
mikexcohen.com

[Mina Marmpena]

Hi Mike, 

I was thinking your suggestions and I wanted to upload some plots to get some more advice cause I'm quite stuck with it, but I had to go through the preprocessing again because I had some serious mistakes and very small epochs. 

I actually have two conditions: the main one (M), and a control (C). I tried to implement some of your suggestions. 

 

You could try not doing any baseline normalization (in this case you should normalize the wavelet in the frequency domain before point-wise multiplying with the data). If you are not looking for power suppressions, this might be OK. 

Trying to see how the above idea works I added the following line of code immediately after the fft of the wavelet

cmwX = cmwX./max(cmwX); 

I kept the dB conversion of the tf without dividing with the baseline. 

The result is a plain dark red surface without any variation at all. Do you think there's a bug the way I used the code or I forgot to change/add something important (I use the code from the lecturelets)?

If you have multiple conditions, you could use a condition-average baseline. That way, any differences will still be apparent in the baseline period.

I 'm not sure what is the proper way to do that. Here's is the steps I assume, pls let me know if I got it wrong. 

1) Get the TF of all the concatenated trials (both conditions)

2) Get the mean power over trials for each freq.

3)  Select an interval [-0.5 0] and avg across time for each freq as baseline vector

4)  Divide the TF of each condition with the baseline vector separately during dB conversion.

Approach A: Results for Fz (33 subjects for M condition and 38 for C)

​ 

 

​

You could use the entire epoch as the baseline. I usually don't like this approach because sustained changes in power will dominate the baseline estimate.

Approach B: I tried this approach as well in two ways: First I did as  in approach A but instead of taking a [-0.5 0] baseline I got the whole interval [-2.5 0]

​

I think taking the baseline from a separate and long rest period is generally not a good idea. There are many differences in cognitive set between task and rest.

Approach C: I understand you are not very into this approach, so I did not use the initial eyes open eyes closed distant trials, but the small 2 secs trials of fixation cross I have in the beginning of each trial (less than 10-50s far before each epoch). I TF, mean and got the baseline vector from these fixation cross trials from both conditions averaged and then divided the mean TF of each condition separately. 

​

The problem is that the response [time=0] comes in self paced mode within each trial and I was worried that if I get my baseline from an interval before -500 ms it might contain task power. Now the response in both conditions is given by pressing a key and I read that the intention for that action is sometimes observed in [-500 0]. So again I am not so sure about how taking a baseline from this interval can affect the TF since it probably contains this intention.   

I really appreciate your help. 

Best, 

Mina

[Mike X Cohen]

Hi Mina. If the non-dB plot is all red, I suspect the color scaling is off. Absolute power cannot be negative, so if you used something like set(gca,'clim',[-3 3]) it will produce all red. Try plotting without setting the colorscaling. 

As for the baselines, first of all, notice that all of your baseline options generally show the qualitatively same pattern, which is good. I think option A is the best. But without knowing more about the design of the experiment, it's difficult to be more confident about a recommendation. 

If the response is made at time=0, then the baseline period should end before time=0, so allow some separation with the response-preparation activity. Perhaps something like [-500 -300] or even earlier if possible.

Mike

[Mina Marmpena]

Hi Mike,

Thank you for your insightful observations,

I did the trick for the non-dB plot (it makes perfect sense as you put it), and I got the two following tf plots ().

It seems very similar for both conditions. Moreover, higher frequencies so no energy.  

Left: Main condition, Right: Control condition

I also tried the [-500 -300] baseline from the averaged conditions:

It seems to me that the interpretations can be very different depending on the baseline choice, something you explain in detail in your book, so I want to ask if I can overcome this ambiguity with the statistical analysis. 

I would be very interested into doing within subject statistics first and then follow up with group analysis. 

I wanted to ask you if I it is good idea to use the second baseline (averaged both conditions trials [-500 - 300]) as a baseline for the single trial vs baseline statistics first.  

Thank you sincerely, 

Mina

[Mike X Cohen]

Hi Mina. The plots look very nice ;) 

One option to think about is performing the statistics on the raw (non-dB transformed) data. You could then show 3 plots: the top two power plots, and a t-map of the differences. Then you wouldn't need to worry about baseline normalization at all. As long as the two groups of subjects were tested on the same equipment and all other aspects of data processing were the same, it should be OK.

In general, single-trial normalizations should be avoided, particularly dB. If you want to apply single-trial baseline, there is a paper from Grandchamp and Delorme (2011) about this issue. 

Mike

[Mina Marmpena]

Thank you Mike!

I think I'll come back for some questions about the statistical analysis, after I study a bit more of these chapters in your book and the paper you suggested. 

I see other people have already posted questions about it and I think it's a great opportunity we have to understand a bit more about this complicated procedure with your help.