What is a "good" noise covariance matrix?

✉️ Mailing List Archive (read-only)
1

Dear mne-python experts and users,

following the guidelines of source reconstruction of ERFs, I estimated
noise covariance matrices from empty room noise (neuromag system) for
calculating inverse operator. When looking at the source estimates I got,
it appears that source amplitude can be very variable, not in term of
timecourse patterns (which is good for ERFs) but in term of absolute
amplitude (need to play with "fmult" in mne_analyze visualization tools; I
suppose it's bad for stats).
So I checked if the noise estimation was similar across subjects and
realize I have no criterion to decide if noise covariance was "ok" or
not...
What criterion should I apply?
Should I use then regularization for "bad" subjects?

PS:find attached several noise covariance matrices from my study
PPS: Does it make sense to band-pass the empty room signal with the same
classical band pass applied to the data? Can it improve a bit the thing?

Best,

Baptiste Gauthier

bad?.png
<https://docs.google.com/file/d/0B_eZxstAMJQscGpiOF9VY00yLWc/edit?usp=drive_web>

good?.png
<https://docs.google.com/file/d/0B_eZxstAMJQsY01WdGlJbENHa0U/edit?usp=drive_web>

2014-10-01 14:05 GMT+02:00 Baptiste Gauthier <gauthierb.ens at gmail.com>:

2

Noise Covariance Estimates should always be treated exactly the same as
your data:

You should always do a full visual inspection of the raw data used to
calculate your noise covariance.

All of the operators applied to your real data (e.g. filters and SSP
vectors etc.) should be applied to your noise covariance data as well.

HTH
D

3

Hi Baptiste,

If you have classical ERFs and a 'baseline' I would not rule out computing
the noise cov from baseline segments, In my experience inverse solutions
based on such a 'subject' noise covariance are often more focal. I had
cases where analyses would have failed using an empty room noise cov.
I share your intuition about the classification of the noise covariances
you have sent.
Very roughly you can say that a covariance is better if its matrix plot
looks more diagonal.
As the covariance is used for whitening the data (sensor data + lead field)
you can investigate its quality by computing a whitener and applying it to
the data:

http://martinos.org/mne/stable/auto_examples/plot_evoked_whitening.html

If the majority of signals in the baseline (assumed to represent signals of
non-interest) are not within -1.96 and 1.96 something is wrong. The cov is
actually good if the covariance matrix of the whitened signals looks like
an identity matrix.

Regularization is important when the number of samples used to compute the
noise cov is small.
But it's also important combine different sensort types.

C.f.
http://martinos.org/mne/stable/auto_examples/inverse/plot_make_inverse_operator.html#example-inverse-plot-make-inverse-operator-py

HTH,
Denis

2014-10-01 16:02 GMT+02:00 Baptiste Gauthier <gauthierb.ens at gmail.com>:

Dear mne-python experts and users,

following the guidelines of source reconstruction of ERFs, I estimated
noise covariance matrices from empty room noise (neuromag system) for
calculating inverse operator. When looking at the source estimates I got,
it appears that source amplitude can be very variable, not in term of
timecourse patterns (which is good for ERFs) but in term of absolute
amplitude (need to play with "fmult" in mne_analyze visualization tools; I
suppose it's bad for stats).
So I checked if the noise estimation was similar across subjects and
realize I have no criterion to decide if noise covariance was "ok" or
not...
What criterion should I apply?
Should I use then regularization for "bad" subjects?

PS:find attached several noise covariance matrices from my study
PPS: Does it make sense to band-pass the empty room signal with the same
classical band pass applied to the data? Can it improve a bit the thing?

Best,

Baptiste Gauthier

bad?.png
<https://docs.google.com/file/d/0B_eZxstAMJQscGpiOF9VY00yLWc/edit?usp=drive_web&gt;

good?.png
<https://docs.google.com/file/d/0B_eZxstAMJQsY01WdGlJbENHa0U/edit?usp=drive_web&gt;

2014-10-01 14:05 GMT+02:00 Baptiste Gauthier <gauthierb.ens at gmail.com>:

Dear mne-python experts and users,

following the guidelines of source reconstruction of ERFs, I estimated
noise covariance matrices from empty room noise (neuromag system) for
calculating inverse operator. When looking at the source estimates I got,
it appears that source amplitude can be very variable, not in term of
timecourse patterns (which is good for ERFs) but in term of absolute
amplitude (need to play with "fmult" in mne_analyze visualization tools; I
suppose it's bad for stats).
So I checked if the noise estimation was similar across subjects and
realize I have no criterion to decide if noise covariance was "ok" or
not...
What criterion should I apply?
Should I use then regularization for "bad" subjects?

PS:find attached several noise covariance matrices from my study
PPS: Does it make sense to band-pass the empty room signal with the same
classical band pass applied to the data? Can it improve a bit the thing?

Best,

Baptiste Gauthier

--
Baptiste Gauthier
Postdoctoral Research Fellow

INSERM-CEA Cognitive Neuroimaging unit
CEA/SAC/DSV/DRM/Neurospin center
B?t 145, Point Courier 156
F-91191 Gif-sur-Yvette Cedex FRANCE

--
Baptiste Gauthier
Postdoctoral Research Fellow

INSERM-CEA Cognitive Neuroimaging unit
CEA/SAC/DSV/DRM/Neurospin center
B?t 145, Point Courier 156
F-91191 Gif-sur-Yvette Cedex FRANCE

_______________________________________________
Mne_analysis mailing list
Mne_analysis at nmr.mgh.harvard.edu
Mne_analysis Info Page

The information in this e-mail is intended only for the person to whom it
is
addressed. If you believe this e-mail was sent to you in error and the
e-mail
contains patient information, please contact the Partners Compliance
HelpLine at
MyComplianceReport.com: Compliance and Ethics Reporting . If the e-mail was sent to you in
error
but does not contain patient information, please contact the sender and
properly
dispose of the e-mail.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.nmr.mgh.harvard.edu/pipermail/mne_analysis/attachments/20141001/f224d875/attachment.html

4

I forgot to mention, if you also have EEG data, you cannot use an empty
room noise cov.

2014-10-01 16:31 GMT+02:00 Denis-Alexander Engemann <
denis.engemann at gmail.com>:

Hi Baptiste,

If you have classical ERFs and a 'baseline' I would not rule out computing
the noise cov from baseline segments, In my experience inverse solutions
based on such a 'subject' noise covariance are often more focal. I had
cases where analyses would have failed using an empty room noise cov.
I share your intuition about the classification of the noise covariances
you have sent.
Very roughly you can say that a covariance is better if its matrix plot
looks more diagonal.
As the covariance is used for whitening the data (sensor data + lead
field) you can investigate its quality by computing a whitener and applying
it to the data:

http://martinos.org/mne/stable/auto_examples/plot_evoked_whitening.html

If the majority of signals in the baseline (assumed to represent signals
of non-interest) are not within -1.96 and 1.96 something is wrong. The cov
is actually good if the covariance matrix of the whitened signals looks
like an identity matrix.

Regularization is important when the number of samples used to compute the
noise cov is small.
But it's also important combine different sensort types.

C.f.
http://martinos.org/mne/stable/auto_examples/inverse/plot_make_inverse_operator.html#example-inverse-plot-make-inverse-operator-py

HTH,
Denis

2014-10-01 16:02 GMT+02:00 Baptiste Gauthier <gauthierb.ens at gmail.com>:

Dear mne-python experts and users,

following the guidelines of source reconstruction of ERFs, I estimated
noise covariance matrices from empty room noise (neuromag system) for
calculating inverse operator. When looking at the source estimates I got,
it appears that source amplitude can be very variable, not in term of
timecourse patterns (which is good for ERFs) but in term of absolute
amplitude (need to play with "fmult" in mne_analyze visualization tools; I
suppose it's bad for stats).
So I checked if the noise estimation was similar across subjects and
realize I have no criterion to decide if noise covariance was "ok" or
not...
What criterion should I apply?
Should I use then regularization for "bad" subjects?

PS:find attached several noise covariance matrices from my study
PPS: Does it make sense to band-pass the empty room signal with the same
classical band pass applied to the data? Can it improve a bit the thing?

Best,

Baptiste Gauthier

bad?.png
<https://docs.google.com/file/d/0B_eZxstAMJQscGpiOF9VY00yLWc/edit?usp=drive_web&gt;

good?.png
<https://docs.google.com/file/d/0B_eZxstAMJQsY01WdGlJbENHa0U/edit?usp=drive_web&gt;

2014-10-01 14:05 GMT+02:00 Baptiste Gauthier <gauthierb.ens at gmail.com>:

Dear mne-python experts and users,

following the guidelines of source reconstruction of ERFs, I estimated
noise covariance matrices from empty room noise (neuromag system) for
calculating inverse operator. When looking at the source estimates I got,
it appears that source amplitude can be very variable, not in term of
timecourse patterns (which is good for ERFs) but in term of absolute
amplitude (need to play with "fmult" in mne_analyze visualization tools; I
suppose it's bad for stats).
So I checked if the noise estimation was similar across subjects and
realize I have no criterion to decide if noise covariance was "ok" or
not...
What criterion should I apply?
Should I use then regularization for "bad" subjects?

PS:find attached several noise covariance matrices from my study
PPS: Does it make sense to band-pass the empty room signal with the same
classical band pass applied to the data? Can it improve a bit the thing?

Best,

Baptiste Gauthier

--
Baptiste Gauthier
Postdoctoral Research Fellow

INSERM-CEA Cognitive Neuroimaging unit
CEA/SAC/DSV/DRM/Neurospin center
B?t 145, Point Courier 156
F-91191 Gif-sur-Yvette Cedex FRANCE

--
Baptiste Gauthier
Postdoctoral Research Fellow

INSERM-CEA Cognitive Neuroimaging unit
CEA/SAC/DSV/DRM/Neurospin center
B?t 145, Point Courier 156
F-91191 Gif-sur-Yvette Cedex FRANCE

_______________________________________________
Mne_analysis mailing list
Mne_analysis at nmr.mgh.harvard.edu
Mne_analysis Info Page

The information in this e-mail is intended only for the person to whom it
is
addressed. If you believe this e-mail was sent to you in error and the
e-mail
contains patient information, please contact the Partners Compliance
HelpLine at
MyComplianceReport.com: Compliance and Ethics Reporting . If the e-mail was sent to you
in error
but does not contain patient information, please contact the sender and
properly
dispose of the e-mail.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.nmr.mgh.harvard.edu/pipermail/mne_analysis/attachments/20141001/76d53d6a/attachment.html

5

To answer this question:

Does it make sense to band-pass the empty room signal with the same classical band pass applied to the data? Can it improve a bit the thing?

My experience is that, if you are using empty room data, the band pass makes essentially no difference. With baseline segments it can make a little difference, but even here the difference is minimal between broadband and band passed. Definitely though, at least broad-band pass the empty room and/or baseline though to remove high frequency noise and very low frequency drift (1-50 Hz or 1-100 Hz or something of that nature) and apply any SSP projectors you use in your real data.

Best wishes,
Avniel

6

Whether the band pass makes a difference or not with empty room data in my experience depends on the band that is included... The spatial covariance appears to be dominated by he lower frequencies...

Regardless, I can't think of any reason to not process the noise data segments in the exact same way as the data segments of interest.

Hari

7

Hi Hari,

That is in part because the noise floor of SQUIDs are frequency dependent.
However, the question is not whether the covariance matrix is biased
towards being influenced by low frequencies, but whether or not the
overall shape of the covariance matrix differs across frequencies. Unless
you have multiple sources of environmental noise with different spatial
distributions in sensor space, each of which has a different frequency
dependance, there should be no effect of band pass filtering on the
covariance matrix built from empty room. Hopefully though, the intrinsic
SSP projections that are attached to the raw fif files would address much
of that.

Best wishes,
Avniel

8

That and other "Unless" statements, is why I strongly recommend
treating all your data identically. One cannot know how good each
person's system and default SSPs etc. are, but if you treat all data
the same (which has no downside) you at least can expect the same
responses.

In the plots Baptiste originally sent the "bad?" plot looked as though
there were artifacts in the empty room recording, which created the
the strikingly different plot. Examining the raw data is a must.

HTH
D

9

Hi Dan,

Agreed Dan, though that has to be balanced against potential
interpretability issues with regards to having multiple inverse operators.
Specifically, if one wants to compare across frequency bands of interest,
but the source maps for those bands were made with different inverse
operators (e.g. the data at the same source location comes from different
sensor configurations at different frequencies), the interpretation of
those results becomes difficult.

Best wishes,
Avniel