Date/Time
Date(s) - 17/09/2026
8:00 am - 9:30 am
Category(ies)
Time: Thursday, 17/9/2026, 8:00 – 9:30am New York time
Title: Agnostic Model-Assisted Estimation with Machine Learning for Survey Data
Speaker: David Haziza (University of Ottawa)
Co-authors: Ziming An, Mehdi Dagdoug, and Yves Tillé
Abstract
Model-assisted estimation uses auxiliary information and prediction models to improve the efficiency of estimators of finite-population parameters under a probability sampling design. Classical approaches typically rely on linear working models, whereas recent work has incorporated nonparametric and machine learning methods. However, many existing methods and theoretical developments are tied to specific prediction algorithms, making it difficult to formulate general conditions for valid design-based inference. We propose an agnostic framework for model-assisted estimation that accommodates a broad class of data-adaptive prediction methods. The proposed approach relies on cross-fitting to weaken the dependence between the model-fitting and estimation steps and to obtain estimators with standard first-order behavior. We establish conditions for the consistency of finite-population mean estimators and develop practical methods for variance estimation and confidence-interval construction. Simulation studies illustrate the finite-sample performance of the proposed procedures under several prediction methods and sampling designs. We also discuss extensions to more general finite-population parameters defined through estimating equations, with particular attention to distribution functions, and present simulations for these settings.