Date/Time
Date(s) - 20/10/2026
4:00 pm - 5:30 pm
Category(ies)
Date/time: Tuesday, October 20th, 2026, 4:00pm – 5:30pm New York time
Title: Robust imputation procedures in the presence of influential survey data based on adaptative tuning constants
Speaker: Sixia Chen
Registration Link: Webinar Registration – Zoom
Abstract:
Missing data due to nonresponse is a common issue in survey sampling and, without appropriate treatment, may lead to substantial bias in point estimation. Item nonresponse is typically handled through imputation, and under a Missing at Random (MAR) mechanism with a correctly specified imputation model, nonresponse bias can be eliminated. However, even when this condition holds, the presence of influential units among respondents may severely affect the efficiency of estimators. Influential units arise frequently in practice, particularly in business and biomedical surveys, where the distribution of the study variable may be highly skewed. These units may have a disproportionate impact on estimates due to high leverage, large residuals, or large sampling weights. Unlike nonresponse, which primarily induces bias, influential units mainly inflate the variance of estimators, leading to unstable results and large mean squared error. A common approach to mitigate the effect of influential observations is to use robust regression methods at the imputation stage (e.g., M-estimators; see Maronna et al., 2019; Andersen, 2008). These methods typically rely on a fixed tuning constant to control the level of robustness. However, while such procedures may perform well for describing the behavior of the inlier population, they may introduce substantial bias when estimating finite population quantities. In particular, a fixed tuning constant does not adapt to the sample size and may lead to an unfavorable bias–variance trade-off, especially in the presence of asymmetric outliers. This paper addresses the problem of constructing imputation procedures that properly account for influential units while maintaining good bias and efficiency properties for finite population estimation.
We propose robust imputation procedures based on adaptive tuning constants, where the level of robustness is allowed to vary with the sample size and the structure of the data. The key idea is that, as the sample size increases, the variance of classical non-robust estimators decreases, and the need to downweigh influential observations diminishes. An adaptive tuning constant reflects this behavior by allowing the estimator to gradually revert toward its non-robust counterpart. We investigate three approaches: Conditional Bias-Based Approach: This approach extends the method of Beaumont et al. (2013), which uses the conditional bias of a unit to quantify its influence. The tuning constant is adapted based on this measure, allowing influential units to be downweighted in a data-driven manner. This idea has been further developed in Favre-Martinoz et al. (2016) and Chen et al. (2024) in related contexts. Adaptive -Estimation Approach: We consider an M-estimation framework in which the tuning constant is allowed to vary across units and is determined using information on their influence (e.g., conditional bias). This leads to a robust regression estimator that adapts to the degree of contamination in the data. Mean Squared Error Minimization Approach: We propose selecting the tuning constant by minimizing an estimate of the mean squared error of the resulting imputed estimator. This provides a principled way to balance bias and variance in finite population estimation. In addition, we develop resampling-based procedures to obtain valid statistical inference, as estimating the mean squared error of robust estimators with adaptive tuning constants is a challenging problem.
Bio:
Dr. Sixia Chen is an Associate Professor of Biostatistics at the University of Oklahoma Health Sciences Center and holds the Presbyterian Health Foundation Presidential Professorship. He is an Elected Fellow of the American Statistical Association and an Elected Member of the International Statistical Institute. Dr. Chen directs the Novel Methodologies Unit within the BERD Core of the Oklahoma Shared Clinical and Translational Resources and the Neurosurgery Biostatistics Unit. His research focuses on survey sampling, data integration, missing data, and statistical methodology. He has published more than 130 peer-reviewed articles in biomedical and statistical journals, including Biometrika, The Annals of Applied Statistics, and Statistica Sinica, and serves as an Associate Editor for several leading statistical journals.
