Inference and Imputation for Missing Not At Random Data

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

Version 1 of 1176 papers210 conceptsfrozen 2026-10-04

In data science, missing values are often treated as a nuisance to be "filled in" before the real work begins. However, the reason why data is missing is frequently as informative as the data itself. Most classical methods rely on the assumption that data is Missing At Random (MAR), where the cause of missingness can be explained by other observed variables.

But in the real world, we often face Missing Not At Random (MNAR) scenarios—also known as nonignorable missingness. This occurs when the probability of a value being missing is tied to the value itself. For example, high-income individuals may be less likely to report their salary, or a sensor might fail only when a temperature threshold is exceeded. In these cases, the observed data is a systematically distorted version of reality.

The Geometry of Missingness Graphs

To solve MNAR problems, we must move beyond simple imputation and toward causal modeling. Karthika Mohan, Judea Pearl, and Jin Tian introduced Missingness Graphs (m-graphs) to formalize this challenge. By using Directed Acyclic Graphs (DAGs) to model the relationship between the underlying variables and the "missingness mechanism," they proved that many MNAR distributions are actually "recoverable."

The intuition is that if we can map the causal path of why data disappears, we can use the remaining observed correlations to mathematically "undo" the bias. This framework shifts the focus from guessing missing values to identifying which statistical relations are theoretically consistent despite the gaps. Razieh Nabi and James M. Robins have further refined these causal views, arguing that without a structural understanding of the missingness mechanism, any attempt at inference is essentially an unverified assumption.

Modeling Missingness as a Modality

If the missingness is coupled with the data, then the "mask" (the matrix of 0s and 1s indicating what is missing) should be treated as a first-class citizen in our models.

Traditional generative models like VAEs often fail here because they assume the mask is independent of the latent features. To fix this, Sahra Ghalebikesabi and colleagues developed PSMVAE, which uses pattern-set mixture models to cluster data based on why it is missing. Similarly, Jialei Chen and Pengyang Wang proposed the GNR model, which treats the data and the missingness mask as two equal modalities in a parallel architecture. By learning the joint distribution of the data and its own absence, these models can reconstruct the true underlying distribution even when the observed samples are heavily biased.

From Imputation to Robustness

A growing trend in the field suggests that we should stop trying to "fill in" missing values and instead start modeling the uncertainty they create.

Alireza Aghasi and Saeed Ghadimi introduced RIGID, a framework that treats missing features as sources of statistical uncertainty rather than empty slots. Instead of a single "best guess" imputation, they use a distributionally robust optimization approach to find solutions that work across all plausible values the missing data could have taken.

This "probabilistic" view is also championed by Łukasz Struski and Jacek Tabor, who represent incomplete data points as Gaussian distributions. By training models (like SVMs) on these distributions rather than on raw points, the model naturally learns to be less confident in regions where data is frequently missing, preventing the systematic over-fitting that plagues standard MNAR imputation.

Temporal and Clinical Contexts

MNAR is the default state in clinical time-series data. A doctor doesn't order a blood test at random; they order it because they suspect something is wrong. This makes the timing of the measurement a proxy for the patient's health.

To capture these complex temporal dependencies, Jinsung Yoon and Mihaela van der Schaar developed the Multi-directional Recurrent Neural Network (M-RNN). Unlike standard RNNs, M-RNN looks both forward and backward in time and across different variables simultaneously. This allows the model to see that a missing heart rate reading might be explained by a stable blood pressure reading elsewhere in the stream. For even faster processing, Wenjie Du and David Côté introduced SAITS, which uses self-attention to weigh the importance of observed values across the entire time series, effectively "attending" to the most relevant historical data to fill current gaps.

The Path Forward: Identifiability and Domain Knowledge

The recurring lesson across these 75 papers is that MNAR data is theoretically "unidentifiable" without extra information. Whether that information comes from:

  1. Auxiliary variables that predict missingness, as suggested by Lijuan Wang.
  2. Structural assumptions about the causal graph, as explored by Judea Pearl.
  3. Physics-integrated constraints, such as the musculoskeletal models used by Souvik Chakraborty to estimate unobservable muscle activity.

The field is moving away from generic "black-box" imputation and toward models that explicitly account for the mechanism of data loss. By treating missingness as a signal rather than noise, researchers are finally beginning to unlock the true distributions hidden within incomplete datasets.

Papers in this version

Showing 60 of 176 papers.

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_inference_and_imputation_for_missing_not_da3f8eaa,
  title = {Inference and Imputation for Missing Not At Random Data},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/inference-and-imputation-for-missing-not-at-random-data/rdr_da3f8eaaf2d1450988249e920369e844},
  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_inference_and_imputation_for_missing_not_da3f8eaa_v1,
  title = {Inference and Imputation for Missing Not At Random Data},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/inference-and-imputation-for-missing-not-at-random-data/rdr_da3f8eaaf2d1450988249e920369e844/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}