
On Computation and Robustness Issues in Spatial Statistics
This thesis develops and evaluates robust statistical methods for the analysis and modeling of spatial data, with a focus on improving inference reliability in the presence of outliers and computational challenges.
Overview
First, the effect of spatial covariance matrix ordering on parameter estimation via low-rank approximation is examined, demonstrating that careful ordering can yield higher computation efficiency without affecting accuracy. Next, the Maximum Lq-Likelihood Estimator (MLqE) is developed for replicated spatial datasets, showing superior robustness over the Maximum Likelihood Estimator when outliers are present. To extend its applicability to large-scale problems, a mixed-precision strategy, combined with a hierarchical precision scheme, is introduced to accelerate MLqE computations. Finally, a new method, the Maximum Composite Lq-Likelihood Estimator (MCLqE), is proposed by integrating MLqE with composite likelihood, enabling robust parameter estimation for single-replicate spatial data. Together, these contributions broaden the practical utility of robust spatial inference in both methodological and computational dimensions.
Presenters
Brief Biography
Sihan Chen joined the PhD program in Statistics at King Abdullah University of Science and Technology (KAUST) in 2021, under the supervision of Prof. Marc G. Genton. Prior to that, he received his Bachelor’s and Master’s degrees from The Chinese University of Hong Kong. Sihan's doctoral research focused on developing robust and scalable statistical methods for spatial data analysis. His research interests include spatial statistics, robust estimation, high-performance computing, and applied data science.