Dear users,
I am writing to you because I find some issue with the statistics on source reconstruction, indeed, I have 40 images of sources (averaged in time) and when trying to make statistics on them using the positive ttest agains 0, the results is quite different from the average of those 40 files, furthermore, when I change the statistical threshold it doesnt change the activation of the T sources (because the T are huge but they shouldn’t be as it is a resting study on 40 different randomly taken time windows averaged in time each). I had the same issue with the MNE inversion and the NAI inversion,
Thank you much if you have any idea that could help me understand this issue,
Faithfully,
Astrid Kibleur
Hi Astrid,
First of all, maybe the measure your are taking from the recordings is not adapted to your question.
The difference of averages and t-tests are computing measures for each time point independently. You should rather compute a summary measure from your time windows (a PSD for instance), and use this instead of the original recordings.
A t-test against zero is useful to test if a contrast is different from zero. You would need two condition to contrast.
Otherwise, there is no reason for which your recordings or PSD values should be equal to zero (this would mean that you have no signal at all).
Very high t values for the test you are currently running are expected, simply because you signals are not equal to zero.
Dear Francois, thank you much for your answer,
I already used averaged measure of my time windows, however, the use of PSD would be even a better option indeed because it should give me information on source activation split in different frequency bands I guess? However it doesnt work because 10 s is too short for this PSD measurement (at least this is the error I get in Brainstorm), what would you advise to use to still get such information using Brainstorm?
I get why the signals have high t values, but I thought that it could be thresholded to visualise only activation high enough and reproducible accross the 40 time windows, Therefore I’ll stick to means to see the activation,
Best regards,
Astrid
However it doesnt work because 10 s is too short for this PSD measurement (at least this is the error I get in Brainstorm)
You should not have any problem computing a PSD on a 10s window. You can probably use estimator windows of 1 or 2 seconds.
Can you please copy-paste here the full error message you get?
To summarize: you are taking 40 random windows of 10s each in a longer resting session, average each of them in time, then re-average them together?
This is somewhat equivalent to average your entire session in time. The problem with this approach is that your average amplitudes are only driven by the very low frequencies in your recordings, that are not necessarily coming from the brain. You should at least do this analysis on signals filtered in a relatively narrow frequency band. In the example below, they used averages but only for the alpha band:
http://neuroimage.usc.edu/brainstorm/Tutorials/DeepAtlas
A more practical way of getting similar results is to compute the PSD of the full resting session at once. You can find examples of this approach here:
http://www.sciencedirect.com/science/article/pii/S1053811915003183
We will write a Brainstorm tutorial corresponding to this article in the next few months, including a recommended processing pipeline for resting MEG recordings.
Cheers,
Francois
Dear Fracois,
Thank you again for your help, the psd was okay actually, I used the 40 time windows because it is a common trick to study DMN and have an intra individual information, usually, I use the power of the source reconstruction, and therefore, the sign of the t test results should be directional and indicate the effect of condition, is it right with Brainstorm too? Finally for the group analysis, I wondered if I could perform a t test on the 40 * nb of patients for my two experimental conditions, indicating the factor of non interest (the patients’ variance) in covariates linking each of the 40*2 images from 1 patient? Is it possible with brainstorm? Thank you again!
Best,
Astrid
Hi Astrid,
usually, I use the power of the source reconstruction, and therefore, the sign of the t test results should be directional and indicate the effect of condition
If you are comparing two conditions yes. If you are testing power values against zero as you were mentioning before, the t values will be strictly positive.
or the group analysis, I wondered if I could perform a t test on the 40 * nb of patients for my two experimental conditions, indicating the factor of non interest (the patients' variance) in covariates linking each of the 40*2 images from 1 patient?
In Brainstorm, your only option would be to compute one summary metric per patient and per condition (the average of the 40 windows?), and then test the difference between the averages of the two conditions across patients.
If you are interested in more advanced statistics, you can export all your source maps and process them with R or SPM. Two tutorials explain how to export data to SPM:
http://neuroimage.usc.edu/brainstorm/ExportSpm8
http://neuroimage.usc.edu/brainstorm/ExportSpm12
If you know exactly how to code your test in Matlab, you could also implement it yourself as a process (there is also plenty of documentation about this if you're interested).
I'm sorry if my answers are not very useful, my skills in statistics are relatively poor. If you have more questions, I'd recommend you seek help from someone else for this part.
Cheers,
Francois