Choice of multivariate autoregressive model order affecting real network functional connectivity estimate

Area of Science:

• Neuroscience
• Biophysics
• Computational Neuroscience

Background:

• Estimating functional connectivity in neural networks is vital for understanding brain function.
• Multivariate Autoregressive (MVAR) models are powerful tools for this analysis.
• Accurate MVAR model order selection is critical for reliable connectivity estimates.

Purpose of the Study:

• To identify the most robust procedure for selecting the correct model order in MVAR analysis.
• To evaluate the impact of signal-to-noise ratio, filter settings, and sampling rates on functional connectivity estimates.
• To compare different methods for assessing time-frequency connectivity.

Main Methods:

• Simulated realistic cortical sources derived from magnetoencephalography (MEG) recordings.
• Comparison of various model order selection criteria (e.g., MDL, SBC).
• Validation in real cases using resting-state and task-based MEG data, comparing MVAR with non-parametric methods.

Main Results:

• Incorrect MVAR model order significantly distorts functional connectivity estimates.
• Minimal Description Length (MDL) and Schwartz Bayesian Criterion (SBC) are the most robust model order selection methods.
• Partial Directed Coherence (PDC) demonstrated superior performance for time-frequency connectivity estimation in both simulated and real data.
• MVAR connectivity estimates are sensitive to filter settings in real-world applications.

Conclusions:

• A robust procedure for MVAR model order selection has been established.
• Validation of MVAR results by comparison with classical methods is recommended for real-world data.
• Correct MVAR model order selection and appropriate band filtering are essential for accurate neural network connectivity estimation.