Score-Driven Modeling of Spatio-Temporal Data.

Area of Science:

• Statistics
• Neuroscience
• Data Science

Background:

• Spatio-temporal data analysis often requires models that account for heavy tails and complex dependencies.
• Functional magnetic resonance imaging (fMRI) data presents unique challenges due to its high dimensionality and inherent noise.
• Identifying spontaneous brain activity is crucial for understanding resting-state brain function.

Purpose of the Study:

• To develop a novel statistical model for analyzing spatio-temporal data with heavy-tailed distributions.
• To apply the model to functional magnetic resonance imaging (fMRI) data for identifying spontaneous brain activations.
• To provide a robust method for estimating dynamic signals in the presence of heavy-tailed noise.

Main Methods:

• Development of a simultaneous autoregressive score-driven model incorporating autoregressive disturbances.
• Signal plus noise decomposition of a spatially filtered process, with a multivariate Student-t distribution for noise.
• Utilizing the score of the conditional likelihood function to drive the dynamics of the space-time varying signal.

Main Results:

• Derivation of consistency and asymptotic normality for maximum likelihood estimators.
• Demonstration of the model's ability to provide robust updates for space-time varying locations in heavy-tailed distributions.
• Successful identification of spontaneous brain region activations in resting-state fMRI data.

Conclusions:

• The proposed score-driven model offers a robust framework for spatio-temporal data analysis, particularly for heavy-tailed distributions.
• The model effectively captures spatial and temporal dependencies in fMRI data, enabling the identification of spontaneous activations.
• This approach advances the analysis of complex neuroimaging data and other spatio-temporal datasets exhibiting extreme values.