Hey everybody. I am trying to compute the average power of an EEG signal in a specific frequency range. EEG is a time-domail brainwave and is very unlikely to perform perfectly in a single trial. What I have to do is repeat the same experiments for thousands of times and take an average in frequency domain.
I am now able to save a number of PSD data across all trials. For example: 04.png and 06.png show 2 volunteers' EEG in one same experiment. 04db.png and 06db.png show their y-axis transform into bd. Surely, they donot look the same.
And here's my question: If I want to analyze certain frequency bands, how can I average these thousands of PSDs? Can I just make a very simple arithmetic mean with 04&06, or 04db&06db? Or something else?
Hi,
You have many options, it depends on how you want to interpret your data.
Generally speaking, permutation statistics is used to statistically validate your results. If you want to find out, if your PSD differed from the baseline, you have 2 options:
Either average the raw PSD for the baseline and for the task for each participant (04,06 I guess) (average across trials) , and run the permutation test between task and baseline.
The other option is to obtain the normalized PSD (04db/06db I guess) for each participant, and run the permutation test against zero. (in each permutation multiply half of the participants (randomly selected) data by -1 ).
You can use other statistical test, permutation test is commonly used in EEG-analysis. Make sure you correct the multiple comparision either you use permutation test or other.
Here you cand find mor information abot the permutation test:
https://benediktehinger.de/blog/science/statistics-cluster-permutation-test/
and this is a more visual implementation:
https://www.youtube.com/watch?v=-85FAQwD49Q
Hope I helped
Bests
I believe you'd better use Probability Density Functions/Auto-correlation for such computation...
You may select the appropriate density function (uniform, normal, etc.. many) depending on the nature and number of samples ..It is appropriate to use Matlab
Yes, a simple arithmetic mean is fine in any case.
Whether to use raw PSD ( 04&06 in your example) or PDS in decibell units (04db&06db), which is a log-transformation of raw PSD values, depends on the type of statistical comparisons you are planning for your data.
Generally, it is better to use log-transformation of the single-trial PSD before averaging withing a subject. This allows using a t-statistics even at the single-subject level.
Here is an excerpt form my answer to a similar question, explaining why its so:
https://www.researchgate.net/post/Log-transformed_or_absolute_power_for_PSD_comparisons_in_EEGLAB
"In the single-subject data the absolute PSD values follow a chi-square distribution over the sample of trials or FFT epochs, which complicates within-subject statistical tests between experimental conditions. Taking the logarithm of PSD normalizes the data, i.e. log(PSD) values are normally distributed across epochs within a subject. Therefore log-transformation of single-trial PSD allows to use t-statistics for evaluating significance of experimental effects at the single-subject level."