Generalized entropy calibration for selection bias

Statistical Seminar

Organizer：

Yunan Wu 吴宇楠（YMSC）

Speaker：

Yumou Qiu 邱宇谋 (PKU)

Time：

Fri., 16:00-17:00 , Mar. 20, 2026

Venue：

C654, Shuangqing Complex Building A

Title：

Generalized entropy calibration for selection bias

Abstract:

We propose a unified framework for constructing calibration weights for data with selection bias by maximizing a generalized entropy function subject to carefully chosen calibration constraints. The proposed generalized entropy calibration (GEC) method can be applied to a variety of problems including missing data, causal inference and survey sampling. Compared to widely used augmented inverse propensity weighting (AIPW) methods, the proposed method can integrate information from multiple propensity score and outcome regression models and achieve multiply robust inference under high-dimensional covariates. Traditional calibration methods minimize a distance between calibrated and initial weights. GEC is a novel calibration framework that instead maximizes a generalized entropy function subject to two types of constraints: covariate balancing constraints to incorporate outcome regression models and to improve efficiency and debiasing constraints involving propensity scores. We establish the asymptotic properties of the proposed estimator, including design consistency, asymptotic normality and multiply robustness. Particularly for survey sampling under Poisson design, we develop an optimal entropy function, called contrast-entropy, which minimizes the asymptotic variance among a broad class of entropy functions.