[FieldTrip] NAI noise normalization beamforming

Schoffelen, J.M. (Jan Mathijs) janmathijs.schoffelen at donders.ru.nl
Thu Sep 29 14:09:57 CEST 2022

Hi Nina,

The NAI is computed by dividing the estimated source power by an estimate of the projected sensor noise, where the sensor noise is assumed to be a scaled identity matrix. Furthermore, it is assumed that the strength of this noise can be approximated by the computations in line 193-200 of ft_inverse_lcmv (for instance). In other words, if you specified ‘lambda’ as a regularization parameter, then this noise strength is assumed to be the max of the lowest singular value of the data covariance, and the regularization parameter. If you specified this lambda to be a percentage, e.g. ‘5%’, the number that corresponds to lambda amounts to a percentage of the trace of the covariance matrix. However, irrespective of how you specified the regularization, and even if you specified cfg.lambda to be 0, if there’s a systematic difference in some property of the covariance matrix across groups (it could be something as exotic as the number of removed spatial components after an ICA, or perhaps even a slightly larger amount of subject movement, or distance from the sensors), then it could indeed be that you observe a ‘flip’ of your effect when you compare power versus when you compare NAI.

In general, I would recommend against using NAI in its native FT implementation overall, due to what I wrote above.

Best wishes,

On 22 Sep 2022, at 13:57, Merkel, Nina via fieldtrip <fieldtrip at science.ru.nl<mailto:fieldtrip at science.ru.nl>> wrote:

Dear Fieldtrip Community,
I am using NAI normalization for my resting state MEG data. It does something to my data that I don't quite understand😳.
Before normalization I have the average power of groupA > groupB (also true for the sensor data). After normalization the pattern turns, while the variance of groupA decreases and the variance in groupB increases. I am not exactly sure what that means. Is the change in power completely explainable by the higher variance of groupA? Are there any limitations of using NAI? Is there a situation where I should not use it?
I am especially wondering about the variance since during my whole analysis it was quite clear that this a a major difference between the groups.
I am posting this question since I was not able to figure out how exactly the NAI is estimated during source reconstruction.
Does anyone have more information about the computation of NAI?
Thanks so much
fieldtrip mailing list

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20220929/909c7b9d/attachment.htm>

More information about the fieldtrip mailing list