Time frequency statistics paired t test or paired permutation?


I just read tutorials and if I understand well to run statistics between two paired group on time frequency maps I have to necessarly do non parametric test, so I have to do non normalized times frequency maps for each trials for each condition for each subject?

Or it is possible to simply do normalized time frequency maps for each trials for each condition for each subject and then do my paired t test on average of all trials of normalized time frequency maps for each condition between my condition across my subjects ?

I tried both for few subject, both worked with almost same results.. i want to know what is the correct way to do ?

thank you,


Hi Marie,

Which tutorial are you refering to?
This page suggests possible pipelines for group analysis of time-frequency maps:

For a single subject analysis, you can test single trials between experimental conditions. For group analysis, you test subject-level averages.


Hi Fran├žois,

Thank you, and I was refering to this tutorial : https://neuroimage.usc.edu/brainstorm/Tutorials/Statistics#Example_6:_Nonparametric_test_on_time-frequency_maps

I have my answer I can run parametric test on yime frequency, but my new question is:
do i have to normalize each subject averages with ERS/ERD or Z-score ?

If when I previously run my time-frequency maps I selected "1/f" to nromalized and "Save averaged time-frequency maps", this average is not normalized already ?

thank you again,


What you need when comparing multiple subjects is to standardize your data with respect to a baseline, in order to bring all your patients in the same range of values (Z-score or other).
The "1/f" normalization is more a display option than a proper standardization.

Maybe @pantazis has additional suggestions?

Hi Marie,

For time frequency data, the statistical tests will evaluate each time-frequency point separately. Thus a 1/f normalization will not influence the statistical results because it is the same as multiplying the data of each test with a fixed number (for example, for 10 Hz each data point will be multiplied by 1/10 before performing say a t-test. This will not influence the test). In contrast, a normalization such as ERS/ERD will convert the values to % changes from baseline. This will scale data from each subject separately, providing a meaningful normalization especially if some some subjects produce more variable responses than others. So I agree with Francois suggesting an ERS/ERD rather than 1/f normalization.