New Methods in Experimental Design for Subsampling Big Data

This thesis develops novel subsampling methods to address challenges in big data analysis. It proposes experimental design-based techniques that overcome key limitations of existing subsampling approaches by allowing multiple models and accounting for potential model misspecification. The benefits are demonstrated through simulation and real-world applications where improved efficiency and precision in parameter estimation are observed. The proposed methods provide a scalable and rigorous framework for analysing complex, high-dimensional data. This research advances experimental design and promotes model-robust subsampling for more reliable data-driven decision-making. The developed methods are implemented via the R package NeEDS4BigData, supporting practical applications across big data settings.