Learning and Inference Under Relaxed Missingness Assumptions

117 concepts108 papers1 proposal

When building machine learning models or conducting causal inference, we rarely enjoy the luxury of complete datasets. Real-world data is riddled with holes. Historically, the standard remedy has been to make convenient assumptions about why data is missing, typically sorting the problem into Donald Rubin’s classic taxonomy:

  • Missing Completely at Random (MCAR): The missingness is entirely independent of any data values (e.g., a sensor occasionally drops packets due to random thermal noise).
  • Missing at Random (MAR): The missingness depends only on observed data (e.g., older patients are less likely to fill out a digital survey, but age is fully recorded).
  • Missing Not at Random (MNAR): The missingness depends on the unobserved values themselves (e.g., high-income individuals systematically decline to report their earnings).

While MCAR and MAR allow for clean mathematical guarantees, they are frequently violated in practice. If a model assumes missingness is a random coin flip when it is actually driven by the underlying features, the model learns a fundamentally biased distribution. A growing body of research focuses on relaxing these rigid assumptions—developing methods to diagnose missingness mechanisms empirically, guarantee safety under MNAR scenarios, and understand the implicit mechanics of how models handle incomplete data.

Diagnosing the Missingness Mechanism

The core challenge of missing data is that you cannot easily prove whether data is MAR or MNAR from the dataset alone. Because the missing values are unobserved, verifying if the missingness depends on them is fundamentally impossible without external information.

To bypass this mathematical dead end, Iavor Bojinov et al. introduce a testable alternative in Diagnosing missing always at random in multivariate data. They define a condition called Missing Always at Random (MAAR). Instead of requiring that missingness depends on unobserved data, MAAR requires that the missingness indicators depend strictly on variables that are fully observed across the entire dataset.

By shifting the dependency to fully observed covariates, the authors construct three empirical diagnostic tests. For the first time, researchers can statistically test whether their missingness assumptions hold before deploying downstream models, signaling exactly when a dataset requires rigorous sensitivity analysis rather than standard imputation.

Breaking the Dependency in MNAR Settings

When data is MNAR, standard imputation or weighting methods collapse because the "missingness mask" itself carries predictive information about the missing values. To build robust models, algorithms must explicitly decouple or account for this statistical dependency.

Density Ratio and Weighting Adjustments

One way to handle MNAR data is to reweight the observed data to match the true, unconfounded population distribution. In Density Ratio Estimation and Neyman Pearson Classification with Missing Data, Josh Givens et al. introduce M-KLIEP, a robust density ratio estimator designed for MNAR environments. M-KLIEP corrects the selection bias introduced by the missingness mechanism, restoring statistical consistency and allowing classifiers to maintain strict error control even when critical features are systematically withheld.

For complex datasets where variables are missing in arbitrary, non-sequential combinations—known as non-monotone missingness—traditional likelihood methods become computationally intractable. Jianing Dong et al. tackle this in Balancing Weights for Non-monotone Missing Data. The authors use functional basis expansions to construct balancing weights that align the observed data distributions across different missingness patterns, bypassing the need to explicitly specify the intractable MNAR likelihood function.

Information-Theoretic De-biasing

Early deep generative imputation models, such as GAIN (Jinsung Yoon et al., GAIN: Missing Data Imputation using Generative Adversarial Nets), relied heavily on the assumption that data was MCAR or MAR to provide formal guarantees. When applied to MNAR data, these models inadvertently encode the missingness bias into the imputed values.

To solve this, Jiahao Yu et al. developed MIRI (Missing Data Imputation by Reducing Mutual Information with Rectified Flows).

Standard Imputation (Biased under MNAR):
[True Data Distribution] ---> [Missingness Mask] ---> [Imputed Data retains Mask Bias]

MIRI Imputation Framework:
[True Data Distribution] ---> [Rectified Flows] 
                                    |
                        (Minimize Mutual Information)
                                    v
[Missingness Mask] ------------> [MIRI Imputation] ---> [Unbiased Data Reconstruction]

MIRI uses an ordinary differential equation (ODE) objective powered by Rectified Flows to systematically minimize the mutual information between the generated data values and the missingness mask. By breaking this statistical link, MIRI prevents the imputation engine from learning the bias of the missingness mechanism, yielding clean data reconstructions even in high-dimensional tabular and image datasets.

Causal Inference and Uncertainty Quantification

Missing data is particularly disruptive in causal inference, where missing outcomes or covariates can easily be mistaken for treatment effects.

Robust Causal Estimators

In observational studies, the Difference-in-Differences (DiD) design is widely used to estimate treatment effects over time. However, if outcomes are missing at random (MAR), standard DiD estimators fail. Lorenzo Testa et al. address this in Efficient Difference-in-Differences Estimation when Outcomes are Missing at Random by deriving "multiply robust" estimators. These estimators maintain efficiency and accuracy even if some of the underlying propensity score or outcome regression models are misspecified, providing a mathematical safety net for observational studies.

When clinical trials face both patient noncompliance and missing outcomes, the problem compounds. Trang Quynh Nguyen et al. formalize a solution in Identification of complier and noncomplier average causal effects in the presence of latent missing-at-random (LMAR) outcomes: a unifying view and choices of assumptions. They introduce Latent MAR (LMAR), a framework where outcomes are assumed to be missing at random conditional on baseline characteristics, treatment assignment, and the unobserved compliance type of the participant, unifying several disparate causal identification strategies.

Preserving Predictive Coverage

Beyond point estimation, missing data invalidates standard uncertainty quantification. Conformal prediction typically guarantees that a model's prediction interval will contain the true value with a user-defined probability (e.g., 95%). However, as Margaux Zaffran et al. demonstrate in Conformal Prediction with Missing Values, missing features alter the conditional variance of the prediction error, destroying these coverage guarantees.

The authors introduce Missing Data Augmentation (MDA). MDA works by artificially masking the calibration dataset across various missingness patterns during the conformalization step. This forces the prediction intervals to adapt dynamically to the specific pattern of missingness present at test time, preserving exact distribution-free coverage guarantees without needing to retrain the underlying predictive model.

Demystifying Naive Imputation

While researchers design sophisticated algorithms to handle missingness, practitioners frequently rely on "naive" hacks, such as zero-imputation (replacing all missing values with zeros). Surprisingly, these simple methods often perform remarkably well in empirical benchmarks. Recent work has begun to uncover the mathematical reasons behind this phenomenon.

In Naive imputation implicitly regularizes high-dimensional linear models, Alexis Ayme et al. prove that for high-dimensional linear models where the number of features $d$ vastly exceeds the sample size $n$ ($d \gg \sqrt{n}$), zero-imputation acts exactly like an implicit ridge regularizer. The bias introduced by shifting missing values to zero shrinks the model coefficients, reducing variance and stabilizing predictions in overparameterized regimes.

However, this mathematical luck does not automatically translate to deep learning. Joonyoung Yi et al. demonstrate in Why Not to Use Zero Imputation? Correcting Sparsity Bias in Training Neural Networks that zero-imputation causes neural networks to develop a severe sparsity bias.

Because neural network layers compute weighted sums, a sample with many missing values (imputed as zeros) will naturally produce smaller activations than a complete sample. The network shifts its predictions based on the quantity of missing data rather than the actual feature values. To counteract this, the authors introduce Sparsity Normalization (SN), which dynamically scales layer activations based on the sample's missingness ratio, restoring the network's ability to learn true feature representations across varying levels of data completeness.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_missingness_mechanism_assumption_00f8d92e,
  title = {missingness mechanism assumption},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/missingness-mechanism-assumption/rdr_00f8d92e5f4746178e6e7731fb8f96a5},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_missingness_mechanism_assumption_00f8d92e_v1,
  title = {missingness mechanism assumption},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/missingness-mechanism-assumption/rdr_00f8d92e5f4746178e6e7731fb8f96a5/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake