A1611
Title: A 3D semiparametric spatial autoregressive model for multisubject data
Authors: Jinxuan Bai - University of Southampton (United Kingdom) [presenting]
Zudi Lu - City University of Hong Kong (China)
Chao Zheng - University of Southampton (United Kingdom)
Ian Galea - University of Southampton (United Kingdom)
Aravinthan Varatharaj - University of Southampton (United Kingdom)
Abstract: Spatial data analysis is essential for medical imaging but poses substantial methodological challenges. Neuroscience studies of the blood-brain barrier (BBB) produce massive three-dimensional datasets, demanding spatial models that are both interpretable and computationally tractable with statistical efficiency for medical research. To address this need, a three-dimensional semiparametric spatial autoregressive (3D-S2AR) model is introduced. The framework captures local voxel-level variation through a spatial autoregressive component based on three-dimensional distance-weighted matrices, while simultaneously modeling large-scale smooth nonparametric trends. A key feature of the framework is its explicit accommodation of multi-subject neuroimaging data, allowing each participant's spatial domain to differ across subjects to reflect individual variation in brain geometry, without requiring a common spatial domain. Crucially, it achieves computational efficiency even for datasets with millions of spatial locations. To mitigate the computational burden, a profile quasi-likelihood estimation method is developed and consistency and asymptotic normality of the resulting estimators are established.
