Hi,
I am trying to do cluster-based permutation tests on my MEG data (Neuromag, 306 channels). As I don’t have a reason to prefer one sensor type of another (mags vs grads) and there is no adjacency for all 306 sensors, I run all analyses separately for both sensor types, which is a little annoying. So when I stumbled upon the method epochs.as_type(), it seemed like an easy way to combine the sensor types. Also in the accompanying example, it says that using this function is a way to run statistics on both sensor types at the same time.
This sounds to good to be true, so I wondered whether this is really as easy. Can I just map my gradiometers to (virtual) magnetometers, and then run a cluster-based permutation test on all 306 channels, using the adjacency matrix of the magnetometers, or is there more to it?
The magnetic field will have a distinctly different pattern across the axial magnetometers than the planar gradiometers. The magnetometers “see” a source as the magnetic field coming out of the head (red colors) on the sensors at one side of the source and the magnetic field going into the head (blue colors) on the sensors on the other side of the source. In contrast, the gradiometers see the source as an increase in the gradient of the magnetic field on the sensors right above the source.
Conceptually, it does not make much sense to perform a cluster-based permutation test on both magnetometers and gradiometers at the same time, because effects will be assigned to different clusters based on which sensor type is recording it. Also, there is no clear adjacency between magnetometers and gradiometers. While the magnetometer sits on top of two gradiometers, they measure the field at different orientations. The whole concept of a “cluster”, where “nearby” sensors measure “mostly the same thing” is invalid. Note that this is not the case for EEG, where the sensors do see sources in such a way that clusters can be defined.
The recommended solution is to perform source estimation and perform the test on the estimated sources, which do behave in the manner assumed by the “cluster” concept at the core of the test.
However, if that is not possible for some reason, I would say that the gradiometers behave the most like EEG sensors and in a manner most suitable for a clustering approach. If you really want to include data from the magnetometers, you could indeed use the epochs.as_type() method to use that data to create slightly more “precise” virtual gradiometer signals and perform the test on that. When you do this, please do explain the procedure in the Methods section of the paper, as this is not a very common operation, and your reasons for not going the typical route of source estimation.
