Asymptotics of Nonparametric Estimation under General Non-monotone MAR Missingness: A Nonparametric Maximum Likelihood Approach
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: We're looking at the paper: "Asymptotics of Nonparametric Estimation under General Non-monotone MAR Missingness".
Jane: Here's the gist of the paper. Missing data constitute a pervasive challenge in empirical research The paper derives a general rate of convergence under MAR that requires no modeling of the missingness mechanism and no restriction on the configuration of missing patterns, beyond MAR itself and a natural positivity condition Applying this result to density estimation, it shows that the complete-data density can be estimated at the minimax rate over a Hölder class, up to a logarithmic factor, for any prescribed smoothness level The missingness does not affect the rate and enters only through a constant Theoretical Foundation of Sieve MLE The paper establishes a general sieve MLE estimation rate result under MAR, showing that essentially the same rate can be attained as in the complete data setting, up to the constant 1/√c0 related to the positivity assumption The authors use recent results in Kaji (2026) to remove the stringent assumption of bounded density ratios They derive a new ignoring density estimator that achieves minimax optimal rates up to logarithmic factors in a Hölder class, inspired by the Bayesian estimator discussed in Ghosal and van der Vaart (2017a) The complete-data rate is attained for any smoothness level with only an increased constant related to a natural positivity assumption Density Estimation and Algorithm Implementation The derived result is applied to density estimation, developing a new (ignoring) estimator that achieves minimax optimal rates up to logarithmic factors in a Hölder class This estimator allows for drawing new data points from the estimated complete-data distribution The estimator is implemented using an Expectation Maximization (EM) algorithm that optimizes the parameters of the mixture distribution and operates on the missing data directly Key Contributions to MAR Estimation The paper presents several key contributions, including deriving a general sieve MLE estimation rate result under non-monotone MAR Applying this result to density estimation, it develops a new (ignoring) estimator that achieves minimax optimal rates up to logarithmic factors in a Hölder class The implementation uses an EM algorithm that operates on the missing data directly and is not related to the usual parametric EM algorithms that use implicit imputations of the missing values Simulation Results and Robustness The simulation study assesses the estimator's ability to address challenges illustrated in a motivating example where non-monotone MAR is present The method predicts the 0.1 quantile of X1 with good overall accuracy even when the lower bound of P(M = 0 X = x) varies from 0.01 to 0.1 The estimator performs well in recovering the complete-data density in high-dimension settings, showing robustness against varying levels of missingness Rate Analysis and Complexity The rate analysis demonstrates that MAR enters only as a constant The first inequality in the main result shows that KL(f P0∥Pn) ≤ (3 + 4√MMAR)δ 2n, which is obtained by bounding the exposed tail The random terms are controlled using Chebyshev's inequality, leading to a bound of OP(δ 2n ∨ n-1) for the empirical-process terms The paper concludes that the proposed estimator reaches the minimax rate up to logarithmic factors for any smoothness level, with a rate that degrades only by a constant pertaining to the positivity assumption The work opens the door to develop efficient and well-performing new estimators through closed-form iterations in practice The approach is shown to be much more widely applicable, particularly assuming P to be the set of all Gaussian mixtures Pruning redundant reasoning tokens Missing data constitute a pervasive challenge in empirical research The paper derives a general rate of convergence under MAR that requires no modeling of the missingness mechanism and no restriction on the configuration of missing patterns, beyond MAR itself and a natural positivity condition Applying this result to density estimation, it shows that the complete-data density can be estimated at the minimax rate over a Hölder class, up to a logarithmic factor, for any prescribed smoothness level The missingness does not affect the rate and enters only through a constant How the training signal is built The authors derive a general sieve MLE estimation rate result under MAR, showing that essentially the same rate can be attained as in the complete data setting, up to the constant 1/√c0 related to the positivity assumption They use recent results in Kaji (2026) to remove the stringent assumption of bounded density ratios Where the gains come from The rate analysis demonstrates that MAR enters only as a constant The first inequality in the main result shows that KL(f P0∥Pn) ≤ (3 + 4√MMAR)δ 2n, which is obtained by bounding the exposed tail The random terms are controlled using Chebyshev's inequality, leading to a bound of OP(δ 2n ∨ n-1) for the empirical-process terms Pruning redundant reasoning tokens The authors derive a general sieve MLE estimation rate result under MAR, showing that essentially the same rate can be attained as in the complete data setting, up to the constant 1/√c0 related to the positivity assumption They use recent results in Kaji (2026) to remove the stringent assumption of bounded density ratios Pruning redundant reasoning tokens
Paper summary: Tom: So we're looking at this paper today: "Asymptotics of Nonparametric Estimation under General Non-monotone MAR Missingness: A Nonparametric Maximum Likelihood Approach". Basically, it tackles the problem of missing data in empirical research, which is a big headache.
Jane: Right. The main idea they are pushing here is that you don't need to model *how* the data is missing at random, as long as it follows the general condition of MAR and there's this natural positivity condition.
Lu: So they establish a general rate of convergence using sieve maximum likelihood estimation under MAR, which means it works regardless of whether the missingness pattern is monotone or not, and they don't need to restrict the configuration beyond that basic MAR rule.
Meng: Wait, so no complex modeling of the mechanism itself? That sounds a lot simpler than what we usually have to do when missing data gets messy.
Lalam: Exactly. They use this sieve maximum likelihood approach, which is essentially an ignoring MLE version where you only model the observed marginal distributions, and they show that the rate of convergence is essentially the same as in the complete data setting up to a constant factor.
Tom: That constant factor is important because it means missingness doesn't actually slow down your estimation rate much, it just adds a bit of noise through that constant.
Jane: And when they apply this to density estimation, they get a result where you can estimate the complete data density at the minimax rate over a Hölder class up to a logarithmic factor, for any smoothness level you want.
Lu: That's significant because it means we can actually use this method for estimating the true underlying distribution even when some pieces of data are missing.
Meng: But what about implementation? How does this actually translate into something I can deploy in production, not just a theoretical result?
Tom: They develop a new ignoring estimator, and they implement it using an expectation maximization algorithm that operates directly on the missing data, which is different from those standard parametric EM algorithms that rely on filling in the missing values first.
Jane: That direct operation is key because it lets you draw new data points from the estimated complete-data distribution, which gives you more flexibility in building your model.
Lalam: The paper shows that this estimator achieves minimax optimal rates up to logarithmic factors in a Hölder class, and they use recent results from Kaji two thousand twenty-six to get around some of those older requirements about bounded density ratios <ref:2608.10113#pg2>.
Tom: So, essentially, they've found a way to get the best possible estimation rate without needing super restrictive assumptions on the missingness pattern itself.
Jane: The simulation results they show are pretty robust; even when the lower bound of the probability of missing data changes between zero point one and zero point one their estimator still predicts things like a quantile with good overall accuracy.
Lu: That robustness against varying levels of missingness in high-dimension settings is something I find very compelling for representation learning tasks.
Meng: From a deployment standpoint, the complexity is important; the rate analysis shows that the missingness term only enters as a constant, which means it doesn't blow up our computational time as n gets large.
Tom: Right, and they show that in their main result inequality, KL of the true density to their estimate is bounded by something like three plus four times the square root of the missingness rate times delta squared n.
Jane: So even with missing data, you still get a convergence bound that scales with n at a rate related to delta squared n, which is what we expect from standard estimation methods.
Lu: The random terms in that analysis are controlled using Chebyshev's inequality, leading to a bound of big O of delta squared n or one over n to the power of one for the empirical process terms.
Meng: So, if I'm worried about running this on real traffic, I need to know what the practical cost is beyond just the asymptotic rate.
Tom: The paper concludes that the proposed estimator reaches this minimax rate up to logarithmic factors for any smoothness level, and that it only degrades by a constant related to that natural positivity assumption.
Jane: It opens the door for developing more efficient estimators through closed-form iterations in practice, especially if we assume P is the set of all Gaussian mixtures.
Lu: I think what's most interesting is how this result applies broadly, not just to a specific scenario, but to any setup where MAR holds and you have that natural positivity condition.
Meng: So for production use, it suggests that as long as we stick to the MAR framework, we can aim for these near-optimal rates without getting bogged down in needing an explicit model of why the data is missing.
Tom: It's a solid result because it removes a big modeling hurdle and gives us concrete convergence benchmarks for density estimation under this kind of complex missingness.
Jane: So, to wrap up, this paper tackles the general problem of non-monotone MAR missingness by using sieve MLE to show that we can estimate densities at optimal rates without needing strong assumptions about the missingness pattern itself.
Conclusion: Tom: So, we're wrapping up this paper on asymptotics of nonparametric estimation under general non-monotone MAR missingness, published by those authors.
Jane: It boils down to a general rate of convergence under missing data that doesn't care how the missingness pattern looks, as long as it follows the basic MAR rule.
Lu: Essentially, they use a sieve maximum likelihood estimation approach to show that you can estimate the complete data density at a minimax rate over any smoothness level.
Meng: So, what does this actually mean for us when we're trying to ship models? Does it just mean our training takes less time?
Tom: Not exactly. It means the missingness doesn't slow down your estimation rate much; it only enters as a constant factor in the error bound.
Jane: That constant factor is what they relate to a natural positivity condition, so that’s where the main caveat lies for practical deployment.
Lu: They are essentially saying that you don't need to model the missingness mechanism itself, which removes a huge modeling hurdle for many of our representation learning problems.
Meng: But you still need that natural positivity assumption to get those optimal rates, right? That sounds like another layer of complexity we have to deal with.
Tom: Exactly. And they show this works even when the missingness isn't monotone, which is a big win because most real-world data isn't perfectly ordered like that.
Jane: The simulation results confirm this robustness; the estimator still recovers the complete data density well even when the missingness probability varies a lot in high-dimension settings.
Lu: From what I see, it opens up new avenues for developing estimators because they aren't tied to filling in missing values first.
Meng: So, if we want to deploy this idea, we need to focus on building an EM algorithm that works directly on the missing data without needing those pre-imputations?
Tom: That’s the implementation aspect you have to worry about. The paper shows this approach is more widely applicable than some of the older methods.
Jane: It’s about getting near-optimal performance for density estimation even when you have messy, non-monotone missing data, provided that basic MAR condition holds true.
Lu: The next thing we should look at is how this applies specifically to Gaussian mixture models, as the authors mention that's where it gets really flexible.
University of North Carolina at Chapel Hill · Research Institute for Statistics and Information Science, University of Geneva
stat.ME
Submitted: 2026-08-10
Updated: 2026-11-04
Code: https://github.com/huimin-hu-stat/MAR
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 90/100
The gist: Missing data constitute a pervasive challenge in empirical research Consequently, there is an ever-growing number of methods designed to address this challenge…
Key concepts
- Missing At Random (MAR)
- This describes a situation where the probability of data being missing depends on other observed data, but not on the actual value of the missing data itself. It is a common pattern in empirical research where this assumption simplifies analysis without requiring a complex model for how data might be missing.
- Sieve MLE Estimation Rate
- This refers to deriving the convergence rate for Maximum Likelihood Estimation (MLE) when dealing with missing data under MAR. The paper establishes that this rate is essentially the same as in the complete-data setting, meaning the presence of missing data doesn't fundamentally slow down how fast an estimate converges.
- Hölder Class
- This is a mathematical space used to describe functions (like probability density functions) based on their smoothness. Functions in this class are continuous and have a certain degree of regularity, which allows the authors to prove optimal estimation rates for density estimation.
Terminology
Summary
Missing data constitute a pervasive challenge in empirical research Consequently, there is an ever-growing number of methods designed to address this challenge, with multiple imputation and inverse probability weighting the dominant strategies Despite this, theoretical guarantees remain limited, particularly in the challenging case of non-monotone missing at random (MAR) In this paper, we utilize the theory of sieve maximum likelihood to establish a general rate of convergence under MAR that requires no modeling of the missingness mechanism and no restriction on the configuration of missing patterns, beyond MAR itself and a natural positivity condition Applying this result to density estimation, we show that the complete-data density can be estimated at the minimax rate over a Hölder class, up to a logarithmic factor, for any prescribed smoothness level The missingness does not affect the rate and enters only through a constant
General Theoretical Framework
The paper establishes that a general nonparametric rate result exists under MAR that requires no modeling of the missingness mechanism and no restriction on the configuration of missing patterns, beyond MAR itself and a natural positivity condition. This result shows that the complete-data density can be estimated at the minimax rate over a Hölder class, up to a logarithmic factor, for any prescribed smoothness level. p. 0 The estimator is approximated in practice by a simple expectation-maximization (EM) algorithm operating on the incomplete data directly. p. 0
Consistency of the Estimator
The paper derives a general sieve MLE estimation rate result under non-monotone MAR, showing that essentially the same rate can be attained as in the complete data setting, up to the constant 1/√c0 related to the positivity assumption. This result allows for obtaining a consistent density estimate through an ignoring density estimator that achieves minimax optimal rates up to logarithmic factors in a Hölder class, inspired by Bayesian estimators
Application to Density Estimation
The general result is applied to density estimation, developing a new (ignoring) estimator that achieves minimax optimal rates up to logarithmic factors in a Hölder class. This estimator allows for drawing new data points from the estimated complete-data distribution
Implementation and Simulation
The new ignoring density estimator is implemented using an Expectation Maximization (EM) algorithm that optimizes the parameters of the mixture distribution, operating on the missing data directly. Simulations show that this method performs comparably to the kernel density estimator supplied with the complete data across a wide range of missingness levels
Key Contributions and Results
The main contributions include deriving a general sieve MLE estimation rate result under non-monotone MAR, showing that essentially the same rate can be attained as in the complete data setting, up to the constant 1/√c0 related to the positivity assumption. This result applies to density estimation, developing a new (ignoring) estimator that achieves minimax optimal rates up to logarithmic factors in a Hölder class. p. 8 The estimator is implemented using an EM algorithm that optimizes the parameters of the mixture distribution, operating on the missing data directly. p. 8 The paper proves that for any probability measure P and any density p, H(p0, pˆn) = OP√1/c0δn∨n−1/2
Simulation Results
Simulations demonstrate that the proposed method can address the challenging problem introduced in the motivating example in Section 3.2, predicting the 0.1 quantile of X1 with good overall accuracy. p. 7 Furthermore, it shows capability in recovering complete-data density for d=20 compared to Gaussian Kernel Density Estimator (KDE) with complete data
Conclusion
In this paper, we derived a general result to show the consistency of sieve-MLE approaches under MAR. We then used this to derive a new ignoring density estimator that reaches the minimax rate up to logarithmic factors for any smoothness level, with a rate that degrades only by a constant pertaining to the positivity assumption
References
Balakrishnan, S., Wainwright, M. J., and Yu, B. (2017). Statistical guarantees for the EM algorithm: From population to sample-based analysis
Burgette, L. F. and Reiter, J. p. 6 P. (2010). Multiple Imputation for Missing Data via Sequential Regression Trees
Chen, S. and Yu, C. L. (2016). Parameter estimation through semiparametric quantile regression imputation
Chen, X., Wan, A. T. K., and Zhou, Y. (2015). Efficient quantile regression analysis with missing observations
Chen, Y.-C. (2022). Pattern graphs: A graphical approach to nonmonotone missing data
Chen, Y.-C. and Sadinle, M. p. 2 (2019). Nonparametric pattern-mixture models for inference with missing data
Ch´erief-Abdellatif, B.-E. and N¨af, J. (2025). Parametric MMD estimation with missing values: Robustness to missingness and data model misspecification
Dempster, A. P., Laird, N. M., and Rubin, D. p. 0 p. 0 B. (1977). Maximum likelihood from incomplete data via the EM algorithm
Deng, G., Han, C., and Matteson, D. S. (2022). Extended missing data imputation via GANs for ranking applications
Frahm, G., Nordhausen, K., and Oja, H. (2020). M-estimation with incomplete and dependent multivariate data
Ghosal, S. and van der Vaart, A. p. 2 p. 2 (2017a). Fundamentals of Nonparametric Bayesian Inference
Ghosal, S. and van der Vaart, A. W. (2017b). Fundamentals of Nonparametric Bayesian Inference
Grzesiak, K., Muller, C., Josse, J., and N¨af, J. (2025). Do we need dozens of methods for real world missing value imputation? p. 8
Han, P., Kong, L., Zhao, J., and Zhou, X. (2019). A general framework for quantile estimation with incomplete data
Kaji, T. (2026). The hellinger bounds on the Kullback–Leibler divergence and the bernstein norm
Kremling, G., N¨af, J., and Lederer, J. (2026). Generative modeling under non-monotone MAR missingness via approximate Wasserstein Gradient Flows
Li, S. C.-X., Jiang, B., and Marlin, B. M. (2019). MisGAN: Learning from incomplete data with generative adversarial networks
Liu, Y. and Fan, Y. p. 1 (2023). Biased-sample empirical likelihood weighting for missing data problems: an alternative to inverse probability weighting
Ma, T., Verchand, K. A., Berrett, T. B., Wang, T., and Samworth, R. J. (2024). Estimation beyond Missing (Completely) at Random
Malinsky, D., Shpitser, I., and Tchetgen, E. p. 1 p. 1 J. T. (2022). Semiparametric inference for nonmonotone missing-not-at-random data: The no self-censoring model
Mattei, P.-A. and Frellsen, J. p. 6 (2019). MIWAE: Deep generative modelling and imputation of incomplete data sets
McLachlan, G. J. and Krishnan, T. p. 2 (2008). The EM Algorithm and Extensions
Mealli, F. and Rubin, D. B. (2015). Clarifying missing at random and related definitions, and implications when coupled with exchangeability
Molenberghs, G., Beunckens, C., Sotto, C., and Kenward, M. G. (2008). Every missingness not at random model has a missingness at random counterpart with equal fit
Muzellec, B., Josse, J., Boyer, C., and Cuturi, M. (2020). Missing data imputation using optimal transport
Nabi, R., Bhattacharya, R., Shpitser, I., and Robins, J. (2025). Causal and counterfactual views of missing data models
N¨af, J. and Ch´erief-Abdellatif, B. p. 1-E. (2026). Asymptotics of nonparametric estimation under general nonmonotone MAR missingness: A Bayesian approach
N¨af, J. (2026). A practical guide to modern imputation
N¨af, J., Scornet, E., and Josse, J. (2026). What is a good imputation under MAR missingness? p. 0
Ouyang, Y., Xie, L., Li, C., and Cheng, G. (2023). MissDiff: Training diffusion models on tabular data with missing values
Pervin, M. M., Wang, H., and Kim, J. K. (2026). A calibration framework for inference with partially observed data
Qin, J., Zhang, B., and Leung, D. H. Y. (2009). Empirical likelihood in missing data problems
Rubin, D. B. (1976). Inference and missing data
Saha, S. and Guntuboyina, A. p. 2 (2020). On the nonparametric maximum likelihood estimator for Gaussian location mixture densities with application to Gaussian denoising
Stekhoven, D. J. and B¨uhlmann, P. p. 6 (2011). MissForest—non-parametric missing value imputation for mixedtype data
Sun, B. and Tchetgen, E. J. T. (2018). On inverse probability weighting for nonmonotone missing at random data
Takai, K. and Kano, Y. p. 1 (2013). Asymptotic inference with incomplete data
van der Vaart, A. W. and Wellner, J. p. 2 A. (2023). Weak Convergence and Empirical Processes: With Applications to Statistics
Wainwright, M. J. (2019). High-Dimensional Statistics: A Non-Asymptotic Viewpoint
Wang, Q. and Rao, J. p. 0 N. K. (2002). Empirical likelihood-based inference under imputation for missing response data
Wei, Y., Ma, Y., and Carroll, R. J. (2012). Multiple imputation in quantile regression
Wei, Y. and Yang, Y. p. 0 (2014). Quantile regression with covariates missing at random
Wong, W. H. and Shen, X. (1995). Probability inequalities for likelihood ratios and convergence rates of sieve MLES
Yoon, J., Jordon, J., and van der Schaar, M. p. 2 (2018). GAIN: Missing data imputation using generative adversarial nets
Yu, J., Ying, Q., Wang, L., Jiang, Z., and Liu, S. p. 1 (2025). Missing data imputation by reducing mutual information with rectified flows
Yuan, X. and Dong, X. p. 0 (2019). Weighted empirical likelihood for quantile regression with non ignorable missing covariates
Zhao, S. and Candes, E. (2025). Imputation-powered inference
--- Page 1 ---
The single most informative sentence about the paper's subject and finding is Missing data constitute a pervasive challenge in empirical research.
Missing data constitute a pervasive challenge in empirical research Consequently, there is an ever-growing number of methods designed to address this challenge, with multiple imputation and inverse probability weighting the dominant strategies Despite this, theoretical guarantees remain limited, particularly in the challenging case of non-monotone missing at random (MAR). In this paper, we utilize the theory of sieve maximum likelihood to establish a general rate of convergence under MAR that requires no modeling of the missingness mechanism and no restriction on the configuration of missing patterns, beyond MAR itself and a natural positivity condition
The paper derives a general sieve MLE estimation rate result under non-monotone MAR, showing that essentially the same rate can be attained as in the complete data setting, up to the constant 1/√c0 related to the positivity assumption. This result shows that for any prescribed smoothness level, the complete-data density can be estimated at the minimax rate over a Hölder class, up to a logarithmic factor
The new ignoring density estimator is implemented using an Expectation Maximization (EM) algorithm that optimizes the parameters of the mixture distribution, operating on the missing data directly Simulations show that this method performs comparably to the kernel density estimator supplied with the complete data across a wide range of missingness levels
The main contributions include deriving a general sieve MLE estimation rate result under non-monotone MAR, showing that essentially the same rate can be attained as in the complete data setting, up to the constant 1/√c0 related to the positivity assumption. This result applies to density estimation, developing a new (ignoring) estimator that achieves minimax optimal rates up to logarithmic factors in a Hölder class
In this paper, we derived a general result to show the consistency of sieve-MLE approaches under MAR.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Nonparametric Density Estimation under General MAR Missingness: The system can estimate complete data density at minimax rate up to a logarithmic factor for any prescribed smoothness level, which is comparable to the complete-data case rate
up to a logarithmic factor.
-
Robust Estimation Against Complex Missingness Patterns: The estimation framework is robust against non-monotone missing at random (MAR) patterns and does not require modeling the missingness mechanism or making assumptions about dependencies between variables.
-
Efficient Density Estimation via Sieve MLE: The system utilizes an EM algorithm operating directly on incomplete data to find an approximate maximizer, which avoids the
often difficult task of specifying a model for the missingness mechanism
and allows forclosed-form iterations.
-
High-Dimensional Density Recovery: The estimator achieves nearly optimal rates in high dimensions, with the final convergence rate being related to n minus 2 beta over 2 beta plus d times log n squared t.
-
Improved Model Selection: The system can select the number of components in a Gaussian mixture model using a BIC or AIC criterion, and the paper suggests that
a penalty is needed that factors in the information loss of the missing values,
indicating an improved model selection mechanism compared to standard methods.
Sources
- Nonparametric Pattern-Mixture Models for Inference with Missing Data
- Parametric MMD Estimation with Missing Values: Robustness to Missingness and Data Model Misspecification
- Do we Need Dozens of Methods for Real World Missing Value Imputation?
- Generative Modeling under Non-Monotone MAR Missingness via Approximate Wasserstein Gradient Flows
- Estimation beyond Missing (Completely) at Random
- Asymptotics of Nonparametric Estimation under general non-monotone MAR missingness: A Bayesian Approach
- A Practical Guide to Modern Imputation
- What Is a Good Imputation Under MAR Missingness?
- A Calibration Framework for Inference with Partially Observed Data
- Imputation-Powered Inference
Related papers
- Learning Robust Treatment Rules for Censored Data
- Neural Estimation of Summary Statistics for Simulation-Based Inference with Non-Asymptotic Guarantees
- Spectral divide-and-conquer MCMC for long stationary time series
- Improving the Efficiency of Subgroup Analysis in Randomized Controlled Trials with TMLE
- Personalized Federated Vector Autoregression with Personalization Diversity
- Doubly Robust and Efficient Calibration of Prediction Sets for Right-Censored Time-to-Event Outcomes