New Approaches in Survey Statistics: Balanced Sampling, Calibration and Imputation

This thesis explores new approaches and methods in survey statistics. It focuses on balanced sampling, calibration, and imputation. It begins with an introduction to the key concepts of survey methodology, and more specifically, to the three main themes of the thesis. The first part focuses on extending the cube method by incorporating inequality constraints. This allows the selection of samples that are not only balanced on auxiliary variables, but also meet minimum size requirements in specific groups. Applications include, for example, category bounding, controlled matrix rounding or spatial sampling. The second part focuses on calibration and is divided into two articles. The first article addresses the harmonization of survey weights when different variables each have their own weighting system. Two strategies are compared: one based on calibration and the other using optimal transport. The second article uses bagging and principal component decomposition to solve issues arising from high-dimensional calibration. The third part focuses on imputation and introduces a hybrid method that combines SwissCheese, a donor-based balanced hot-deck approach, with missForest, a predictor based on random forests.

Dissertation Committee
Prof. Yves Tillé, University of Neuchâtel, Thesis Director
Prof. Alina Matei, University of Neuchâtel, Jury President
Prof. David Haziza, University of Ottawa, Rapporteur
Prof. Camelia Goga, Marie et Louis Pasteur University, Rapporteur
Dr. Caren Hasler, University of Neuchâtel & University of Zürich, Rapporteur

Thesis Defended on June, 24th 2025

No de thèse : 3198