Hi,
At first I would thank you for the very informative tutorial and the user-friendly toolbox.
In a within-subject block design study, I recorded EEG for 4 different conditions per subject. I did preporocessing, epoching (45 seconds every block) and averaging (for 4 block repetitions per condition). Now I am willing to compare the power of frequency bands between different conditions. Therefore, I wonder which way is correct for going on:
- first split the averaged file of each conditions to different frequency bands --> compute paired t-test between conditions for the averaged file of each frequency band --> compute PSD and visualize them
OR
- compute PSD of the averaged file of each conditions for different frequency bands --> compute paired t-test between conditions for the PSD of each frequency band --> visualize them
I have also 2 additional questions:
-
basically, to compare the power of the frequency bands in 2 different conditions, do I need to report PSD or the result of comparing averaged files is enough?
-
since we have negative and positive power for different brain areas for each frequency band, is it logically correct to calculate the t-test of “absolute” A and B?
Many thanks in advance.
Cheers,
Mojtaba
1 Like
Hi Mojtaba,
1&2) Only option #2 is available from the interface. The statistical test must be the last element of you pipeline, you would not be able to perform any signal processing on the p or t values.
I would say: PSD for each trial separately > t-test between two conditions (FilesA = the PSD of all the trials for condition A, FilesB = PSD of all the trials for condition B)
You have two options for the PSD: using the full FFT resolution or averaging by larger frequency bands. Both can be good options, but you have to be careful with the way you define your frequency bands. I would recommend you start with the full frequency resolution before going further, it would help you understanding what is happening.
If you don’t see what you’re expecting with the frequency bands, it could be because the bands you defined are not adapted with your data or frequency resolution. You can edit them freely in the additional options of the PSD process.
If you have 3 conditions, maybe you will be interested in running other types of statistical tests.
If you’re familiar with Matlab scripting, you can always export your data from the Brainstorm database as text files, or simply read it from your Matlab scripts and run your own stats.
-
I’m not sure I understand this: Can you describe better what your two options would be?
-
After you calculate the signal power in your different frequency bands with the PSD process, you cannot have negative values. This the sum of squared values.
By “power”, you mean “current amplitude”? In this case, yes, you would run your t-test on the absolute values of the source amplitudes.
Note that the t-test on the source maps are performing independent tests time-by-time. But as you epochs are very long (45s), this might not represent very well your data.
If you are interested in the temporal dynamics of your 45s blocks, I would recommend you run a time-frequency decomposition (using the Hilbert transform or Morlet wavelets), and then to run your t-test on the TF files.
Cheers, Francois
Hi Francois,
Thanks for the response.
1&2) You suggested PSD for each trial, but why not for the average of the trials in a condition? I think statistically I think it is correct.
Moreover, since I am interested in comparing different frequency bands in different conditions. Therefore, isn’t it better to calculate PSD for different frequency bands (editing in PSD options)?
-
With your answer to the above questions, I got my answer.
-
Yes, I meant amplitude.
Best,
Mojtaba
You have many epochs of 45s, split in four conditions, and you want to compare the power in frequency band f1 between condition A and condition B, right?
Over 45s, the brain processes are probably not strictly time-locked. If you average them, you will lose a lot of the information that is not time-locked, including most of the higher-frequency oscillations.
If you estimate the PSD of the averaged signals, you would get something with no relevant information above 5Hz. This is not the appropriate way of proceeding.
Whether you want to run an average or a t-test, you should estimate first the PSD of non-averaged recordings, and then average/test it.
You could do either a difference of averages: average(PSD(trialsA)) - average(PSD(trialsB))
Or a t-test between conditions: t-test(PSD(trialsA), PSD(trialsB))
The two would give you similar results, the test mainly adding a significance level to the difference of means.
When using the option of the PSD “Group in frequency bands”, it first calculates the results with the full frequency definition, and then averages by frequency bands.
This grouping is an additional step of processing, that can easily mask other underlying errors. I would recommend you do both (full spectrum AND frequency bands) for at least one subject, and when you get confident that they both give equivalent correct results, repeat it on the other subjects with just the “frequency bands” version.
Francois
Yes, exactly.
With averaging, the time of the epoch doesn’t change; however, you are absolutely right that with averaging I am losing so much information.
For each subject I have 16 blocks of 45 sec (4 conditions x 4 repetitions), so I will have 16 PSD files per subject. For doing for instance a paired t-test, shall I compute it over all the PSDs (26 subjects x 4 PSD files for one condition), or over the average PSDs of 4 trials of a condition (26 subjects x 1 averaged PSD file)? Or doesn’t matter.
Best,
Mojtaba
This I’m not really sure about… You’d have to ask someone with better skills in statistics.
I would start by running the two and check if the results are any different.
I already asked, and the response was to first averaging and then doing the t-test. One point is that, since it is possible to lose some trials (that I did), and for paired t-test we need the same number of trials; therefore, we will have problem with unequal number of trials.
However, I am also willing to do both by myself.
Many thanks again.
Cheers,
Mojtaba
Hi Mojtaba,
From your questions it sounds that you are new to electrophysiological signals. With MEG and EEG we typically prefer event related designs and not block designs, unless there is a compelling reason to have long blocks. We do not know anything about your experiment and whether you really need such long 45sec blocks or it was just a wrong design.
Be it as it may, Francois hinted on the possibility of losing useful oscillatory signal by averaging trials. For detailed information, read for example the paper ‘Oscillatory gamma activity in humans and its role in object representation’ by Tallon-Baudry in Trends in Cognitive Sciences, 1999. Box 1 is particularly informative. The topic is also addressed in my MIT course ‘A look inside the human brain’ with slides here: http://1drv.ms/17OMR4z (slide 23)
For statistics, it really depends what you want to do. For a random effects analysis, an easy way for t-tests is to average results within each subject, and then carry a single value per subject to the ttest (this is called two-level analysis). You have 4 repetitions with each subject, you should average these. The run t-tests for each pairs of conditions.
You do paired t-tests if measurements can be paired (as in your case, say each subject gave you results for condition A and condition B, then each subject has a useful contrast A-B). In your design you have the extension of paired measurements, you have 4 repeated measurements (each subject was measured in 4 conditions). So you need a repeated-measures ANOVA design. Software packages such as SPSS can effectively run such tests. In this case, the software can handle individual trials for random-effects analysis, so you should not average repetitions, just give the PSD of all 26 subjects, 4 conditions, 4 repetitions.
Best,
Dimitrios