Asymptotics of Nonparametric Estimation under General Non-monotone MAR Missingness: A Nonparametric Maximum Likelihood Approach
Abstract
Missing data are pervasive in empirical research, and the missing at random (MAR) condition has been central to their analysis since its introduction fifty years ago. Yet theoretical guarantees remain limited, particularly when missingness is non-monotone. Where they exist, they are confined to simplified settings such as monotone or block-wise missingness, or rest on restrictive assumptions about the missingness mechanism. In this paper, we establish a general rate of convergence result for an ignoring sieve MLE that holds without these restrictions. It requires only a natural positivity condition and a condition strictly weaker than (possibly non-monotone) MAR, which we refer to as ``MAR-like MNAR''. We moreover show that, even if this condition is violated, the resulting bias may still be bounded by a measure of deviation from MAR-like MNAR, which is of order ε under Huber-type ε-MNAR-contamination of a MAR mechanism. Our result can be used broadly to design new algorithms for data generation and imputation. We illustrate this through a density estimator for the Hölder class and show that it achieves the complete-data minimax rate for any prescribed smoothness level, up to a logarithmic factor. The estimator is approximated in practice by a simple expectation-maximization (EM) algorithm and performs comparably to the kernel density estimator supplied with the complete data. We illustrate the utility of the new estimator in several applications, including
Related
- Asymptotics for likelihood ratio tests of boundary points with singular information and unidentifiable nuisance parameters
- Estimation and Inference for Win Measures with Multiple Ordinal Endpoints Subject to Missingness
- Nonparametric hazard rate estimation with associated kernels and minimax bandwidth choice
- nethist: An R package for Nonparametric Graphon Estimation via Network Histograms
- CovEsts: Nonparametric Autocovariance Estimation and Analysis
BibTeX
@misc{zou2026asymptotics,
title = {{Asymptotics of Nonparametric Estimation under General Non-monotone MAR Missingness: A Nonparametric Maximum Likelihood Approach}},
author = {Yating Zou and Huimin Hu and Gitte Kremling and Jeffrey Näf},
howpublished = {arXiv},
year = {2026},
month = nov,
doi = {10.48550/arxiv.2608.10113},
eprint = {2608.10113},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2608.10113},
}
RIS
TY - UNPB
TI - Asymptotics of Nonparametric Estimation under General Non-monotone MAR Missingness: A Nonparametric Maximum Likelihood Approach
AU - Yating Zou
AU - Huimin Hu
AU - Gitte Kremling
AU - Jeffrey Näf
PY - 2026
DA - 2026/11/04
PB - arXiv
DO - 10.48550/arxiv.2608.10113
UR - https://arxiv.org/abs/2608.10113
AB - Missing data are pervasive in empirical research, and the missing at random (MAR) condition has been central to their analysis since its introduction fifty years ago. Yet theoretical guarantees remain limited, particularly when missingness is non-monotone. Where they exist, they are confined to simplified settings such as monotone or block-wise missingness, or rest on restrictive assumptions about the missingness mechanism. In this paper, we establish a general rate of convergence result for an ignoring sieve MLE that holds without these restrictions. It requires only a natural positivity condition and a condition strictly weaker than (possibly non-monotone) MAR, which we refer to as ``MAR-like MNAR''. We moreover show that, even if this condition is violated, the resulting bias may still be bounded by a measure of deviation from MAR-like MNAR, which is of order ε under Huber-type ε-MNAR-contamination of a MAR mechanism. Our result can be used broadly to design new algorithms for data generation and imputation. We illustrate this through a density estimator for the Hölder class and show that it achieves the complete-data minimax rate for any prescribed smoothness level, up to a logarithmic factor. The estimator is approximated in practice by a simple expectation-maximization (EM) algorithm and performs comparably to the kernel density estimator supplied with the complete data. We illustrate the utility of the new estimator in several applications, including
ER -