Analyzing Functional Data with a Mixture of Covariance Structures Using a Curve-Based Sampling Scheme

Abstract

Motivated by distinct walking patterns in real-world free-living gait data, this paper proposes an innovative curve-based sampling scheme for the analysis of functional data characterized by a mixture of covariance structures. Traditional approaches often fail to adequately capture inherent complexities arising from heterogeneous covariance patterns across distinct subsets of the data. We introduce a unified Bayesian framework that integrates a nonlinear regression function with a continuous-time hidden Markov model, enabling the identification and utilization of varying covariance structures. One of the key contributions is the development of a computationally efficient curve-based sampling scheme for hidden state estimation, addressing the sampling complexities associated with high-dimensional, conditionally dependent data. This paper details the Bayesian inference procedure, examines the asymptotic properties to ensure the structural consistency of the model, and demonstrates its effectiveness through simulated and real-world examples.