Analysis in sensor space

This chapter describes how to perform statistical analyses of EEG/MEG data. This requires transforming data from SPM M/EEG format to image files (NIfTI format). Once the data are in image format the analyses for M/EEG are procedurally identical to 2nd level analyses for fMRI. We therefore refer the reader to the fMRI section for further details of this last step.

In the drop down “Images” menu, select the function Convert to images. This will open the batch tool for conversion to images. You will need to select the input dataset, that can be either a mat-file on disk or a dependency from a previous processing step.
Then you need to set the ‘mode’ of conversion. M/EEG data in general case can be up to 5-dimensional (3 spatial dimensions, time and frequency). SPM statistical machinery can only handle up to 3 dimensions. Although this is a purely implementational limitation and the theory behind SPM methods can be extended to any dimensionality, in practice high-dimensional statistical results can be very hard to interpret not least due to our inability as humans to visualise them. Furthermore, unconstrained high-dimensional test would incur very severe penalty for multiple comparisons and should in most case be avoided. Thus, our purpose is to reduce our data dimensionality to be 3 or less. The three spatial dimensions in which the sensors reside can be reduced to two by projecting their locations onto a plane. Further reduction of dimensionality will involve averaging over one of the dimensions. The choices for ‘mode’ option correspond to all the different possibilities to average over a subset of data timensions. Some of the options are only relevant for time-frequency data where the frequency dimension is present.
‘Conditions’ options makes it possible to only convert data for a subset of conditions in the file. This is especially useful for batch pipeline building. The conversion module outputs as a dependency a list of all the generated NIfTI images. These can be used as input to subsequent steps (e.g. statistical design specification). By including the ‘Convert2images’ module several times in batch each condition can have a separate dependency and enter in a different place in the statistical design (e.g. for two-sample t-test between two groups of trials).
The ‘Channels’ option makes it possible to select a subset of channels for conversions. These can be either selected by modality (e.g. ‘EEG’) or chosen by name of by a list in a mat-file (e.g. to average over all occipital channels).
‘Time window’ and ‘Frequency window’ options limit the data range for conversion which is especially important if the data are averaged over this range. Make sure you only include the range of interest.
Finally the ‘Directory prefix’ option specifies the prefix for the directory where images will be written out. This is important if several different groups of images are generated from the same dataset (e.g. from different modalities or different channel groups).

Output

When running the tool a directory will be created at the dataset location. Its name will be the name of the dataset with the specified prefix. In this directory there will be a nii-file for each condition. In the case of averaged dataset these will be 3D images (where some dimensions can have size of 1). In the case of an epoched dataset there will be 4D-NIfTI images where every frame will contain a trial.

Averaging over time or frequency

Although 2D scalp images averaged over time or frequency dimension can be created directly in conversion to images, they can also be generated by averaging over part of the Z dimension of previously created 3D images. This is done via ‘Collapse time’ tool in the ‘Images’ menu.

Masking

When you set up your statistical analysis, it might be useful to use an explicit mask to limit your analysis to a fixed time window of interest. Such a mask can be created by selecting Mask images from “Images” dropdown menu. You will be asked to provide one unsmoothed image to be used as a template for the mask. This can be any of the images you exported. Then you will be asked to specify the time (or frequency) window of interest and the name for the output mask file. This file can then enter in your statistical design under the ‘Explicit mask’ option or when pressing the ‘small volume’ button in the ‘Results’ GUI and choosing the ‘image’ option to specify the volume.

Smoothing

The images generated from M/EEG data must be smoothed prior to second level analysis using the Smooth images function in the drop down “Images” menu. Smoothing is necessary to accommodate spatial/temporal variability between subjects and make the images better conform to the assumptions of random field theory. The dimensions of the smoothing kernel are specified in the units of the original data (e.g. [mm mm ms] for space-time, [Hz ms] for time-frequency). The general guiding principle for deciding how much to smooth is the matched filter idea, which says that the smoothing kernel should match the data feature one wants to enhance. Therefore, the spatial extent of the smoothing kernel should be more or less similar to the extent of the dipolar patterns that you are looking for (probably something of the order of magnitude of several cm). In practice you can try to smooth the images with different kernels designed according to the principle above and see what looks best. Smoothing in time dimension is not always necessary as filtering the data has the same effect. For scalp images you should set the ‘Implicit masking’ option to ‘yes’ in order to keep excluding the areas outside the scalp from the analysis.
Once the images have been smoothed one can proceed to the second level analysis.