Recovering Distributions from Nonignorable Missing Data

This direction is merged into another direction: non-random missingness. Its last version stays readable and citable here.

Version 1 of 118 papers18 conceptsfrozen 2026-10-04

In most data science workflows, missing data is treated as a nuisance to be imputed via simple averages or "Missing at Random" (MAR) assumptions. However, real-world data often suffers from nonignorable missingness, also known as Missing Not at Random (MNAR). This occurs when the reason a value is missing is tied to the value itself—for instance, high-income individuals refusing to report their salary, or the sickest patients dropping out of a clinical trial.

When missingness is nonignorable, standard imputation methods fail because they cannot account for the systematic selection bias. Recent research has focused on identifying when these distributions can be recovered and developing efficient algorithms to do so without introducing bias.

The Geometry of Recoverability

The first hurdle in handling MNAR data is determining if the true distribution is even "identifiable"—that is, whether we can uniquely determine the underlying population parameters from the incomplete data we observe.

Jin Tian has pioneered the use of graphical models, specifically m-graphs, to solve this. In Recovering Probability Distributions from Missing Data, Tian demonstrates that by mapping the causal relationships between variables and their "missingness indicators," we can systematically determine if a joint or marginal distribution can be reconstructed. This moves the field away from "guessing" if data is recoverable and toward a rigorous, algorithmic proof of recoverability based on the structure of the data-gathering process.

Efficient Estimation without the Curse of Dimensionality

Even if a distribution is theoretically recoverable, the computational cost can be prohibitive. Traditional likelihood-based methods often struggle with the "curse of dimensionality" when many different patterns of missingness exist.

To combat this, Chunrong Ai and Oliver Linton proposed a method using the Generalized Method of Moments (GMM). In A Simple and Efficient Estimation Method for Models with Nonignorable Missing Data, they show that by matching observed moments rather than maximizing full likelihoods, researchers can achieve high accuracy in finite samples without the exponential complexity of previous approaches. Similarly, Aaron Courville and Yoshua Bengio demonstrated that organizing missingness patterns into a Minimum Spanning Tree can speed up Gaussian Mixture Model training by 20x, making it feasible to "fill in the blanks" for massive datasets.

Bridging Deep Learning and Statistical Inference

Modern machine learning requires training complex models like Variational Autoencoders (VAEs) on incomplete data. The challenge is that the missingness mechanism itself must often be modeled alongside the data.

Vaidotas Simkus and Michael U. Gutmann introduced Variational Gibbs Inference (VGI) to address this. As detailed in Variational Gibbs inference for statistical model estimation from incomplete data, VGI combines the flexibility of Gibbs sampling with the speed of variational inference. This allows the model to iteratively refine its estimate of the missing values and the model parameters simultaneously, reducing the complexity of handling missing data patterns from exponential to linear.

For unnormalized models where the partition function is unknown, Masatoshi Uehara and Jae Kwang Kim developed fractional imputation estimators like FINCE. These tools allow for consistent parameter estimation even when the data is MNAR, by weighting imputed values to correct for the systematic bias inherent in the missingness.

Causal Effects and Cautious Modeling

In clinical and social science settings, the stakes of nonignorable missingness are highest. If patients drop out because a treatment isn't working, a simple analysis will overestimate the treatment's success.

Hua Chen and Zhi Geng addressed this in Semiparametric Inference of the Complier Average Causal Effect with Nonignorable Missing Outcomes. Their semiparametric approach identifies causal effects without requiring the researcher to perfectly specify the mathematical form of the dropout mechanism—a major step forward in robustness.

When the missingness mechanism is truly untestable, Julia Plass and Thomas Augustin advocate for partial identification. Instead of forcing a single, potentially biased point estimate, their "cautious" framework provides a range of credible values. This acknowledges the inherent uncertainty of MNAR data, providing a more honest foundation for policy decisions in small area estimation.

Handling Sparse Missingness Patterns

In high-dimensional settings, we often encounter "sparse support," where many theoretical patterns of missingness never appear in the sample. Trung Phung and Rohit Bhattacharya recently proposed Recursive Equations For Imputation Of Missing Not At Random Data With Sparse Pattern Support, which uses recursive logic to identify and impute values even when the specific combination of missing variables is rare. This ensures that MNAR techniques remain viable even as the number of features grows.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_recovering_distributions_from_nonignorab_9c2e5fae,
  title = {Recovering Distributions from Nonignorable Missing Data},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/recovering-distributions-from-nonignorable-missing-data/rdr_9c2e5faee9eb43dc83e9b2e84a95ca2c},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_recovering_distributions_from_nonignorab_9c2e5fae_v1,
  title = {Recovering Distributions from Nonignorable Missing Data},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/recovering-distributions-from-nonignorable-missing-data/rdr_9c2e5faee9eb43dc83e9b2e84a95ca2c/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}