BLMM

BLMM implements scalable voxelwise linear mixed model estimation for large-scale functional magnetic resonance imaging (fMRI) datasets, enabling inference that accounts for complex covariance and grouping structures.

Key Features:

• Scalability and voxelwise vectorization: Performs vectorization over voxels to accelerate computations for large-n neuroimaging datasets.
• Fisher Scoring estimation: Implements a Fisher Scoring procedure for linear mixed model parameter estimation.
• Fisher information and score derivations: Uses explicit derivations of the LMM Fisher information matrix and score vectors building on Maullin-Sapey and Nichols (2021).
• Missing data handling: Accounts for variable missing data patterns across voxels without resorting to voxel-wise deletion, preserving analysis mask size.
• High-performance computing implementation: Implemented in Python and targeted to high-performance computing clusters to manage computational demands of large sample sizes.
• Large-sample performance: Optimized to manage the computational demands associated with large sample sizes in neuroimaging LMM analyses.

Scientific Applications:

• Large-scale fMRI LMM analyses: Enables linear mixed model analyses on large shared fMRI datasets.
• Voxelwise whole-brain inference: Supports voxelwise inference across whole-brain datasets with complex covariance and grouping structures.
• Analyses with missing data at cortical boundaries: Permits inclusion of voxels affected by missingness near cortical boundaries and brain edges to reduce mask reduction.

Methodology:

Fisher Scoring using derived LMM Fisher information matrix and score vectors, voxelwise vectorization for computational speed-ups, and strategies to account for variable missing data patterns without voxel-wise deletion.