The General Linear Model (GLM) is the most commonly used statistical analysis method for fMRI data. By adapting its design matrix/model, it becomes a highly flexible tool suitable for specific applications such as parametric modulation, psychophysiological interaction analysis (PPI), and longitudinal analysis. Additionally, the GLM can be used beyond affirmative analysis as a “denoising” tool.
It is important to note that the GLM approach relies on several practical and theoretical preconditions. While the specific fMRI software used for analysis can help users avoid some practical errors, it is still the user's responsibility to ensure that the conditions for the GLM are met and that the results of the analysis are sensible and can be properly interpreted.
This paragraph outlines some common issues related to the practical application of the GLM. More in-depth discussions of the theoretical preconditions of the GLM can be found in the relevant literature.
Data
Functional Coverage
For a proper multi-subject analysis, it is crucial not only to transform all functional data into the same reference space (e.g., MNI) but also to ensure sufficient spatial coverage of the data.
Technically, any voxel not covered in all functional datasets of the study will be excluded from the analysis (in the case of a “Random Effects”/RFX GLM).
This can be easily tested using the “Verify Functional Coverage” function within the Options menu. Here, users can load the MDM file of a group analysis and check how well the associated data covers the same voxels. As an output of this analysis, a volume map showing the coverage of each individual VTC file and a probabilistic coverage map are created.
Functional coverage issues can arise from signal intensity problems, corrupted files, residual motion, coregistration errors, normalization problems, and other factors.
Quality of the Single VTC file
Problems in a multi-subject GLM analysis may also arise from noisy and low-quality data included in the analysis. If you would like to inspect the quality of your VTC files (signal to noise, outliers), you can use the "VTC inspector" plugin available from the Brain Innovation helpdesk: https://helpdesk.brainvoyager.com/642193-VTC-Inspector
Model
Mismatching design matrices (SDMs)
Due to issues such as naming inconsistencies or different condition orders within the protocol files linked to your functional data, you may encounter non-matching names or orders of predictors in the SDM files. In such cases, the Multi-Study GLM will not function correctly, and BrainVoyager will report an error message. You will need to check each design matrix (SDM) file and correct the predictors accordingly.
Zero-filled predictors
You may have created a design matrix that contains a predictor that includes no information - only zeros. This can happen, for example, when you add "error predictors" to the design matrices of all your subjects. If a subject never makes an error, the error predictor would be zero-filled. Mathematically, this creates a problem for the GLM, unless you run a dummy-coded multi-study GLM analysis where the same predictor has non-zero values in other runs for the same subject: https://helpdesk.brainvoyager.com/271888-Comparison-of-Different-Runs-or-Sessions-in-RFX-Analyses-Dummy-Coding-Design-Matrices
To assist with the analysis of your design matrix (SDM) files, you can use the “SDM Inspector” plugin available from the Brain Innovation helpdesk: https://helpdesk.brainvoyager.com/636555-SDM-Inspector