A0647
Title: A spatial autoregressive graphical model
Authors: Pariya Behrouzi - Wageningen University and Research (Netherlands) [presenting]
Sjoerd Hermes - Wageningen University (Netherlands)
Joost van Heerwaarden - Wageningen University (Netherlands)
Abstract: There is a notable gap in the statistical literature for methods capable of modeling asymmetric multivariate spatial effects, particularly in settings where spatial relationships vary across categorical labels. In such scenarios, observations at a location arise from both within- and between-location effects, with the latter often exhibiting asymmetry due to heterogeneous interactions between different location types. A novel Bayesian spatial graphical model is proposed that integrates multivariate spatial autoregressive structures with Gaussian graphical models. This integration allows capturing asymmetric spatial dependencies that are modulated by a categorical feature at each location. These feature-dependent spatial effects relax the usual symmetry assumptions commonly imposed in spatial models. However, the added flexibility comes with a trade-off: Spatial effects are not identifiable without prior knowledge of the system or additional parameter constraints. The model's performance is evaluated via simulation studies, and its practical utility is demonstrated on an intercropping dataset, where asymmetric spatial effects naturally arise from interactions between different crop types. The proposed method is implemented in the R package SAGM.
