A0877
Title: Bayesian sequential batch design with hierarchical Gaussian processes
Authors: Shuang Zhou - Arizona State University (United States) [presenting]
Ping-Han Huang - Arizona State University (United States)
Abstract: Longitudinal studies are often challenged by noisy, irregularly sampled, and sparse data. Traditional functional data analysis relies heavily on frequentist methods, which struggle to incorporate uncertainties in model parameter estimation. Although Bayesian approaches mitigate this by accounting for uncertainty, their use in sequential batch designs essential for information updating and cost efficiency has been largely overlooked. A hierarchical Gaussian process (GP) model that leverages structural relationships between subjects is proposed to reduce uncertainty in trajectory recovery. This model underpins a novel utility function based on Shannon information between posterior predictive distributions, enabling the sequential identification of optimal designs for new subject batches. The approach enhances the quality of model estimation and analysis in sparse data settings, paving the way for adaptive and efficient longitudinal study designs. Numerical results in both simulation studies and health data applications demonstrate that the method outperforms alternatives, such as functional PCA.
