GCM Help Pages
Main Before you start Computing GCMs Viewing GCMs Interpreting GCMs Advanced More Help	Before you start Granger Causality Mapping is an exploratory technique that can be applied to any fMRI data-set scanned under any kind of experimental protocol. That is, the GCMPlugin will certainly not stop you from computing a map as long as the input parameters are logically valid. Interpretation of the attained result is discussed in the section Interpreting GCMs. However, there are some steps you should undertake even before computing GCMs, in order for the result to make sense later. More specifically, some steps you might take in the preprocessing of your data and the selection of your VOIs might have a deleterious effect on resulting GCMs. Preprocessing Slice scan time correction (SSTC) is of crucial importance for GCM. It should always be performed and, of course, performed correctly. Particularly, selecting the wrong slice scan order, e.g. Ascending instead of Descending, can make SSTC fail in its attempts to correct slice timing discrepancies (note that for SIEMENS scanners Ascending interleaved 1 vs. Ascending interleaved 2 is a subtle though important distinction). SSTC should also be perfomed before anything else, particularly before 3D motion correction (3DMC) (note that this is the default order for BVQX 1.6.0 onward, but not in earlier versions). This is because 3DMC will interpolate spatially, putting data originally scanned in one slice into another one (especially when a target in another run is chosen for simultaneous cross-run alignment). When SSTC (particularly for interleaved slice scan orders) is performed on this 'temporally mixed' data, timing in the preprocessed data will be distorted. When 'slice-shaped' artefacts appear in your GCMs, you'll know you probably have an SSTC-related problem. Temporal preprocessing is also an important step for GCM. Linear trends and slow wave components should be removed at the voxel level by linear trend removal (LTR) and temporal high pass filtering (THP). The VAR models used in the computation of GCMs are only strictly correct for temporally stationary time-series. Non-stationarities, such as linear trends and slow-wave components, should therefore be removed. Temporal smoothing is not recommended on data used for GCM, since it generally deteriorates the temporal resolution of your data. There are not objections to modest spatial smoothing. Selecting VOIs VOIs selected as a reference region for GCM can have any shape or size, it's up to you. However, one recommendation, based mainly on conceptual arguments, is to keep them small (small being between, say, 100 and 300 VMR voxels). Large VOIs will average a lot voxels, risking a loss of temporal detail. More importantly, although a large cluster of voxels may have been activated in the same contrast, some of those voxels may be differently functionally (or effectively) connected to other parts of the brain. Computing GCMs for VOIs averaged over functionally different regions will yield maps that show a mixed average of the involved functional networks. In other words, VOIs that are too large may clutter your GCMs and make interpretation harder.

GCM Help Pages

Main
Before you start
Computing GCMs
Viewing GCMs
Interpreting GCMs
Advanced
More Help

Before you start

Granger Causality Mapping is an exploratory technique that can be applied to any fMRI data-set scanned under any kind of experimental protocol. That is, the GCMPlugin will certainly not stop you from computing a map as long as the input parameters are logically valid. Interpretation of the attained result is discussed in the section Interpreting GCMs. However, there are some steps you should undertake even before computing GCMs, in order for the result to make sense later. More specifically, some steps you might take in the preprocessing of your data and the selection of your VOIs might have a deleterious effect on resulting GCMs.

Preprocessing

Slice scan time correction (SSTC) is of crucial importance for GCM. It should always be performed and, of course, performed correctly. Particularly, selecting the wrong slice scan order, e.g. Ascending instead of Descending, can make SSTC fail in its attempts to correct slice timing discrepancies (note that for SIEMENS scanners Ascending interleaved 1 vs. Ascending interleaved 2 is a subtle though important distinction). SSTC should also be perfomed before anything else, particularly before 3D motion correction (3DMC) (note that this is the default order for BVQX 1.6.0 onward, but not in earlier versions). This is because 3DMC will interpolate spatially, putting data originally scanned in one slice into another one (especially when a target in another run is chosen for simultaneous cross-run alignment). When SSTC (particularly for interleaved slice scan orders) is performed on this 'temporally mixed' data, timing in the preprocessed data will be distorted. When 'slice-shaped' artefacts appear in your GCMs, you'll know you probably have an SSTC-related problem.

Temporal preprocessing is also an important step for GCM. Linear trends and slow wave components should be removed at the voxel level by linear trend removal (LTR) and temporal high pass filtering (THP). The VAR models used in the computation of GCMs are only strictly correct for temporally stationary time-series. Non-stationarities, such as linear trends and slow-wave components, should therefore be removed. Temporal smoothing is not recommended on data used for GCM, since it generally deteriorates the temporal resolution of your data. There are not objections to modest spatial smoothing.

Selecting VOIs

VOIs selected as a reference region for GCM can have any shape or size, it's up to you. However, one recommendation, based mainly on conceptual arguments, is to keep them small (small being between, say, 100 and 300 VMR voxels). Large VOIs will average a lot voxels, risking a loss of temporal detail. More importantly, although a large cluster of voxels may have been activated in the same contrast, some of those voxels may be differently functionally (or effectively) connected to other parts of the brain. Computing GCMs for VOIs averaged over functionally different regions will yield maps that show a mixed average of the involved functional networks. In other words, VOIs that are too large may clutter your GCMs and make interpretation harder.