Informative Missingness

• Informative missingness is a phenomenon where the absence of data itself provides insights into latent states or outcomes.
• It underpins methodologies in healthcare, time series, semi-supervised, and reinforcement learning by explicitly modeling nonrandom data absence.
• Integrating missingness indicators in algorithms enhances prediction accuracy and causal inference by leveraging otherwise ignored information gaps.

Informative missingness refers to the circumstance where the fact that a data entry is missing is itself informative about the data-generating process. In this setting, the probability that an observation is missing depends on unobserved or partially observed data, such as latent health status, true outcome, or the underlying label. This phenomenon stands in contrast to classical missing data frameworks, such as Missing Completely At Random (MCAR) and Missing At Random (MAR), where missingness is considered ignorable for likelihood-based inference.

Informative missingness is pervasive in modern statistical learning and applied machine learning, especially in healthcare, time series, semi-supervised learning, multimodal data integration, and reinforcement learning. Exploiting rather than ignoring or simply imputing informative missingness has been shown to improve both predictive accuracy and causal inference in a wide range of scientific and operational settings.

1. Conceptual Foundations and Formal Definitions

Informative missingness has precise definitions within the Rubin–Little taxonomy:

• Missing Completely At Random (MCAR): $P(M \mid X_\text{obs}, X_\text{mis}) = P(M)$; the missingness mechanism is independent of all data.
• Missing At Random (MAR): $$P(M \mid X_\text{obs}, X_\text{mis}) = P(M \mid X_\text{obs})$$; missingness depends only on observed quantities.
• Missing Not At Random (MNAR): $P(M \mid X_\text{obs}, X_\text{mis})$ depends on unobserved values ($X_\text{mis}$) or the label/latent state itself. In this regime, the pattern of missingness is non-ignorable and may carry information about the system state or the target variable (Habiba et al., 2020, Wu et al., 4 Dec 2025, Rockenschaub et al., 2024).

Informative missingness is synonymous with MNAR. Examples include:

• Diagnostic tests ordered only for sicker patients (EHR time series),
• Labels missing preferentially for ambiguous or hard-to-classify cases (semi-supervised learning) (Wu et al., 4 Dec 2025, McLachlan et al., 3 Dec 2025),
• Sensor failures correlated with environmental extremes (sensor networks) (Habiba et al., 2020).

2. Mathematical Models and Algorithmic Strategies

2.1 Likelihood-based and Mixture Models

The joint modeling of both data and missingness mechanisms is essential for valid inference and prediction. A canonical strategy in finite mixture and latent variable models is to augment the likelihood with an explicit missingness term (Wu et al., 4 Dec 2025, McLachlan et al., 3 Dec 2025): $$L(\theta, \phi) = f(X, Y \mid \theta) \cdot g(M \mid X, Y; \phi)$$ where $g$ parameterizes the missingness process, often with logistic or entropy-driven link functions. EM or ECM algorithms iteratively update both parameter blocks.

Sufficiently informative modeling can yield an expected error for the semi-supervised EM lower than that of the fully supervised model if the informative gain from the missingness exceeds the information loss due to label unavailability (Wu et al., 4 Dec 2025).

2.2 Bayesian and Probabilistic Graphical Models

Bayesian hierarchical methods and mixed-mode mixture models have been developed to jointly capture observed values and missing-data indicators, particularly for time series and EHR data (Mikalsen et al., 2020, Mikalsen et al., 2019). These approaches model the observed variables and missingness masks as part of a unified generative process, typically with conjugate priors for stability under high missing rates.

2.3 Deep Learning Approaches

Specialized architectures such as GRU-D and Neural ODEs have been proposed to model informative missingness in irregular time series (Habiba et al., 2020). GRU-D employs decay-parameterized dynamics on both inputs and masks, enabling the network to learn the temporal significance of observed and missing signals. Neural ODE variants propagate state and missingness masks in continuous time, supporting seamless adaptation to irregular and missing-skewed data streams.

Multimodal fusion models, such as MMNAR-aware causal representation learning, encode missingness patterns via learnable embeddings and condition modality attention and gating on these patterns, providing end-to-end trainable leveraging of nonrandom missingness (Liang et al., 21 Sep 2025).

3. Applied Domains and Methodological Innovations

3.1 Semi-supervised Learning and Label-Missingness

Label missingness in SSL is often MNAR. Explicit modeling of the missing-label mechanism dramatically improves classification accuracy, especially when labeling is biased toward "easy" (or "hard") cases. Inverse propensity weighting (IPW) using propensity scores $p(r=1 | x,y)$ or entropy-driven logistic regressions corrects bias in empirical risk minimization and data augmentation-based SSL (Sportisse et al., 2023, McLachlan et al., 3 Dec 2025). Likelihood-ratio tests can diagnose informative missingness (Sportisse et al., 2023).

3.2 Multivariate and Clinical Time Series

In EHR time series, the pattern of measurements—what is observed and when—is tightly coupled to clinical reasoning and the underlying disease process (e.g., lab ordering policies, test frequency). Models such as TCK$_{IM}$ and its extensions use ensemble Bayesian mixture models that encode missingness as part of the clustering structure, yielding robust performance in high-missingness clinical data (Mikalsen et al., 2020, Mikalsen et al., 2019). Patient-level missingness signatures (IMM) can be learned as embedddings and used to condition imputers and predictors (Ghosheh et al., 2024).

3.3 Multimodal Data Integration

In multimodal settings, naive information gain estimation is confounded by informative missingness. The ICYM$^2$I framework corrects for biased estimates of predictive/unique/shared/complementary information by employing MAR-consistent IPW, producing unbiased scores for data fusion and attribution (Choi et al., 22 May 2025).

3.4 Reinforcement Learning and Sequential Decision-Making

Markov Decision Processes with observation missingness (miss-MDPs) require explicit modeling of the missingness mechanisms. If missingness is MNAR (e.g., self-censoring), belief updates, value computation, and policy optimality all depend on correctly learning the missingness function. PAC-optimal planning is possible only in subclasses exhibiting sufficient identifiability (e.g., no self-censoring, indicator independence) (Wendland et al., 12 May 2026).

3.5 Longitudinal and Outcome-Dependent Processes

Pattern-mixture and selection models, including extensions of the Heckman two-step, have been adapted to quantile regression and binary outcome settings where dropout or missingness is outcome-dependent (Marino et al., 2015, Doretti et al., 14 Nov 2025). Correction terms are derived via relative risks (for logistic-logistic setups) or via class-weighted mixture components for heterogeneity due to drop-out. These adjustments yield consistent parameter estimates under outcome-dependent (informative) missingness.

4. Informative Missingness in Algorithm Design and Evaluation

4.1 Missing Indicator Method and Feature Engineering

Augmenting datasets with explicit missingness indicators (MIM) is effective across a variety of supervised learners for capturing informative missingness, particularly for categorical attributes or when missing rates exceed certain thresholds. MIM is theoretically harmless under MCAR and yields empirically significant improvements under MNAR, provided overfitting is controlled. Selective MIM (SMIM), applying an FDR filter for indicator inclusion, is preferable in high-dimensional settings (Ness et al., 2022, Lenz et al., 2022).

4.2 Imbalanced Learning and Data Augmentation

In imbalanced, scientific, or engineering datasets, methods such as OverNaN extend SMOTE and ROSENaN oversampling to operate directly on incomplete vectors, preserving or mimicking the class-conditional missingness pattern during synthetic sample generation. These approaches outperform classic impute/discard strategies and maintain structural missingness signals as part of the feature space (Barnard, 12 May 2026).

4.3 Expert-Guided Interpretable Classification

Physically grounded generative models for both observed features and detection patterns allow for decomposed goodness-of-fit scoring, explicitly isolating the contributions from detection/non-detection events and observed measurements for inspection and decision-making in critical domains such as seismic monitoring (Cohen et al., 16 Apr 2026).

5. Empirical Performance, Impact, and Limitations

Empirical studies across domains confirm that modeling informative missingness leads to significant gains in classification, prediction, and policy learning accuracy, often exceeding gains attainable through more labeled data or sophisticated imputation alone. This is most pronounced at moderate class overlap, high missingness, and when there is strong dependence between missingness and the target or latent state (McLachlan et al., 3 Dec 2025, Wu et al., 4 Dec 2025, Reich et al., 2010, Habiba et al., 2020).

However, modeling must be aligned with assumptions on the missingness process (MAR vs MNAR) and carefully validated, as misspecification can induce bias rather than reduce it (Rockenschaub et al., 2024). Under non-ignorable shifts in missingness during deployment, overreliance on informative missingness learned in the training regime can lead to decreased performance, mandating sensitivity analysis and robust model selection.

6. Directions for Theory, Methods, and Practice

Key open questions and future research directions in informative missingness include:

• Development of approaches for non-ignorable (MNAR) missingness in high-dimensional, nonparametric, or causal settings, especially in reinforcement learning, representation learning, and longitudinal analysis.
• Extension to multi-modal and multi-task frameworks with differential missingness structures for heterogeneous data sources (Liang et al., 21 Sep 2025, Liang et al., 23 Apr 2026).
• More efficient uncertainty quantification and interpretability of models that couple missingness structure and observed data, including tools for identifying when missingness patterns are truly informative (Cohen et al., 16 Apr 2026).
• Sensitivity analysis and robust estimation methods for model- and policy-deployment under missingness distribution shifts (Rockenschaub et al., 2024, Wendland et al., 12 May 2026).

Leverage of informative missingness is a paradigm shift from treating missingness as a statistical nuisance to embracing it as a source of information. Successfully doing so requires explicit statistical modeling, integration with modern machine learning architectures, and context-specific domain insight. The emerging literature establishes informative missingness as an essential consideration in any practical statistical or machine learning pipeline where data collection, labeling, or observation processes are nonrandom.