A Bayesian Regression Framework for Circular Models
This thesis develops a comprehensive Bayesian framework to address geometric incompatibilities challenges in analyzing directional data.
Overview
Despite advancements in modern statistics, tools for analyzing directional data remain underdeveloped. This is primarily due to the fundamental geometric incompatibility between circular and linear spaces. In Bayesian inference, progress is further hindered by the lack of intuitive guidance for prior specification.
This thesis develops a comprehensive Bayesian framework to address these challenges. To address the challenge of prior specification, we derive Penalized Complexity (PC) priors for circular distributions, providing a transparent mechanism for incorporating intuitive prior knowledge while favoring contraction toward a simpler base model when information is scarce. To resolve geometric incompatibilities, we eliminate the inferential ambiguity inherent in standard generalized regression model approaches by introducing a regression-oriented adjustment for circular responses. This methodology is then expanded into a joint regression framework capable of accommodating multiple circular and linear response variables and covariates, alongside various functional and random effects. Across the simulation studies, the proposed methods provide stable parameter recovery and accurate posterior predictions across a range of model settings, while the real-world applications demonstrate their flexibility in modelling complex dependence between circular and linear variables. Finally, these developments are implemented in the R package INLAcircular. By structurally aligning with the integrated nested Laplace approximation (INLA), as implemented in the R package R-INLA, the framework enables highly efficient Bayesian inference.
Presenters
Brief Biography
Xiang Ye is a Ph.D. student in Statistics at King Abdullah University of Science and Technology (KAUST), advised by Prof. Håvard Rue.