Meta-Imputation Methods

Updated 10 July 2026

Meta-imputation is a framework where the decision to impute is driven by higher-level information, making the process adaptive rather than a fixed preprocessing step.
The approach encompasses adaptive triggering, ensemble combinations, and meta-learned priors, thereby enhancing robustness and efficiency in handling missing data.
Empirical evaluations show that meta-imputation methods can reduce error rates and improve downstream prediction performance compared to traditional imputation techniques.

Searching arXiv for recent and foundational papers on meta-imputation and related formulations. Meta-imputation denotes a family of methods in which imputation is not treated as a fixed preprocessing step, but as a higher-level operation that is selected, combined, adapted, or corrected in response to context, task structure, or inferential goals. In the literature, this includes instance-dependent decisions about whether to impute at all, supervised models that learn to combine multiple candidate imputations, meta-learned priors that adapt to new matrices or time series, cross-study borrowing schemes in multilevel or meta-analytic settings, and inference procedures that use imputation as an efficiency device while retaining explicit bias correction (Liu et al., 2016, Azad et al., 3 Sep 2025, Iwata, 2021, Zhao et al., 17 Sep 2025).

1. Conceptual scope and recurrent formulations

The literature shows that “meta-imputation” is used in multiple, partially overlapping senses. The common theme is that the imputation mechanism is controlled by another layer of information: ambiguity in classification, behavior of several base imputers, task-specific adaptation across datasets, cross-cluster parameter pooling, or an explicit inferential correction.

Formulation	Core mechanism	Representative papers
Adaptive triggering	Impute only when observed data are insufficient	(Liu et al., 2016)
Ensemble combination	Learn a supervised combiner over multiple imputers	(Azad et al., 3 Sep 2025)
Task-adaptive meta-learning	Learn priors or latent codes that adapt to new matrices or series	(Iwata, 2021, Naour et al., 2023, Nie et al., 20 Jan 2025)
Cross-study borrowing	Pool imputation-model parameters across clusters or studies	(Muñoz et al., 2023, Santis et al., 24 Jun 2026)
Inference-oriented imputation	Impute quantities needed for valid or more efficient inference	(Pasaniuc et al., 2013, Kim et al., 2015, Burgwinkel et al., 8 Jul 2026, Zhao et al., 17 Sep 2025)

This breadth matters because the object being “meta-learned” differs across settings. In some work it is the decision to invoke imputation; in others it is the weighting of candidate imputations, the prior over latent representations, or the correction applied after blackbox imputation. A plausible implication is that meta-imputation is better understood as a design pattern than as a single algorithm family.

2. Instance-adaptive and credal formulations

A canonical example of instance-adaptive meta-imputation is Credal Classification with Adaptive Imputation (CCAI), introduced for incomplete pattern classification (Liu et al., 2016). The method first attempts direct classification using only observed attributes. For each class $\omega_g$ , it computes a prototype and a normalized distance using only available dimensions, then converts that distance into a basic belief assignment over $\{\omega_g,\Omega\}$ . Before combining the classwise evidence, it evaluates a distinguishability degree $\chi_i$ , defined as the ratio between the second-largest and largest class support. If $\chi_i \le \epsilon$ , the observed attributes are treated as sufficient and no imputation is performed; if $\chi_i > \epsilon$ , the missing values are deemed crucial and a second-stage imputation procedure is triggered (Liu et al., 2016).

The second stage is explicitly class-conditioned. For each class, a self-organizing map is trained on class-specific training samples, the $K$ nearest SOM prototypes are found for the incomplete pattern, and a classwise imputed vector is constructed by exponentially weighted averaging. This yields up to $c$ alternative completions, one per class hypothesis. Each completion is then classified again, producing classwise basic belief assignments that are discounted according to imputation reliability and fused by a rule that can allocate belief mass not only to singleton classes but also to meta-classes, that is, disjunctions of classes (Liu et al., 2016).

The decisive innovation is that uncertainty induced by imputation is not collapsed into a single hard label. The object is allowed to belong with different masses of belief to specific classes and meta-classes, and these meta-classes absorb conflicts between alternative imputations. In the paper’s experiments, CCAI reduces misclassification relative to always-impute baselines such as KNNI, FCMI, and SOMI, while avoiding unnecessary imputations on “easy” cases; for example, in the 2D experiment with a fully missing second dimension, SCCAI Step 2 only achieved error $4.83\%$ with $Ri_2=19.33\%$ , whereas adaptive CCAI obtained error $5.83\%$ with $\{\omega_g,\Omega\}$ 0 and lower cost than always-impute KNNI (Liu et al., 2016).

This formulation makes meta-imputation a decision problem: whether missing values matter enough, for a given instance, to justify completion. It also makes uncertainty representation central rather than incidental.

3. Learned combinations of imputers and task-aware weighting

A second major formulation treats meta-imputation as supervised combination of several base imputers. Meta-Imputation Balanced (MIB) defines a set of base imputers $\{\omega_g,\Omega\}$ 1 and a meta-model that predicts the true value of an artificially masked cell from the vector of candidate imputations $\{\omega_g,\Omega\}$ 2 and feature metadata $\{\omega_g,\Omega\}$ 3 (Azad et al., 3 Sep 2025). In the paper’s implementation, the base imputers are mean imputation, median imputation, mode imputation, k-Nearest Neighbors imputation, Matrix Factorization based imputation, Autoencoder-based imputation, and GAIN. The meta-learner is a linear regression trained with Mean Squared Error on 10% MCAR synthetic masking under 5-fold cross-validation (Azad et al., 3 Sep 2025).

The mechanism is stack-like rather than heuristic. Each masked cell becomes a supervised training point, the linear model learns weights $\{\omega_g,\Omega\}$ 4 for imputers and coefficients $\{\omega_g,\Omega\}$ 5 for feature metadata, and the final imputed value is a data-driven combination rather than an unweighted average. The paper emphasizes that this differs both from best-single-imputer model selection and from deterministic averaging, because the combination is fit to known ground truth on masked entries (Azad et al., 3 Sep 2025).

Empirically, MIB is presented as robust rather than uniformly dominant. On Heart Disease, it achieved the best direct metrics, with Masked MAE $\{\omega_g,\Omega\}$ 6 and Masked RMSE $\{\omega_g,\Omega\}$ 7. On Diabetes Health Indicators, it attained Masked RMSE $\{\omega_g,\Omega\}$ 8 and the best downstream Random Forest RMSE $\{\omega_g,\Omega\}$ 9. On Gallstone, direct RMSE was weaker than KNN and XGBoost imputation, but under Linear Regression MIB achieved the lowest Prediction RMSE $\chi_i$ 0 (Azad et al., 3 Sep 2025). The paper’s interpretation is that ensemble stability across datasets and downstream models is the main advantage.

GEDI extends the task-aware view further by optimizing imputation jointly with downstream label prediction (Chen et al., 2022). It combines a Transformer-based feature encoder, which models column-wise contextual relationships, with a learned sample-similarity graph and GCN, which models row-wise similarities. Its meta-learning component does not operate across many datasets in a MAML sense; instead, it treats each feature’s imputation loss as an auxiliary task and learns weights $\chi_i$ 1 so that the weighted joint loss improves the target prediction task. The result is a within-task meta-imputation scheme in which features influential for prediction receive larger imputation emphasis (Chen et al., 2022).

A common misconception is that better average reconstruction must imply better downstream performance. GEDI explicitly contests that assumption: equal treatment of all features can produce negative transfer, and end-to-end training with meta-weighted auxiliary losses can outperform both two-step impute-then-predict pipelines and unweighted multitask objectives (Chen et al., 2022).

4. Meta-learned priors and fast adaptation for matrices and time series

A third formulation learns how to impute new structured objects by meta-learning from many related tasks. In meta-learning for matrix factorization without shared rows or columns, the task is a partially observed matrix $\chi_i$ 2, and the model learns a neural network that maps a matrix to prior means $\chi_i$ 3 and $\chi_i$ 4 for its factor matrices (Iwata, 2021). These priors are then adapted by differentiable gradient-descent MAP updates, and the meta-objective is the expected imputation error on held-out entries. Because the architecture uses exchangeable matrix layers and permutation-invariant aggregation, it handles matrices with different sizes and with no shared users or items across tasks (Iwata, 2021). On ML100K, ML1M, and Jester, the proposed method achieved the lowest test MSE on missing entries, with $\chi_i$ 5, $\chi_i$ 6, and $\chi_i$ 7, respectively (Iwata, 2021).

TimeFlow applies the same principle to irregular time series by representing each series as a continuous implicit neural function $\chi_i$ 8 (Naour et al., 2023). The shared parameters $\chi_i$ 9 define the base INR, while a per-series latent code $\chi_i \le \epsilon$ 0 is adapted by a small number of gradient steps and passed through a hypernetwork to produce layer-wise modulation parameters. This turns imputation into a meta-learned adaptation rule rather than a fixed forward pass. The paper reports that TimeFlow achieves state-of-the-art performance on imputation benchmarks, with particularly strong gains at low observation rates, and that second-order meta-learning is materially better than REPTILE or encoder-based alternatives (Naour et al., 2023).

Collaborative Imputation of Urban Time Series through Cross-city Meta-learning extends this idea to cross-city transfer (Nie et al., 20 Jan 2025). MetaTSI uses delay-coordinate embeddings, multi-scale Fourier-enhanced INRs, hierarchical modulation, and masked instance normalization to learn a city-agnostic prior over urban time series. Shared base parameters $\chi_i \le \epsilon$ 1 are meta-trained across a diverse urban dataset from 20 global cities, while a low-dimensional latent code $\chi_i \le \epsilon$ 2 is adapted for each sensor or city. In the cross-city setting, MetaTSI-1D achieved the best average performance across 15 cities, and few-shot adaptation to unseen cities and unseen series remained superior to TimeFlow and Functa, including at observation rates as low as $\chi_i \le \epsilon$ 3 (Nie et al., 20 Jan 2025).

Across these papers, the imputer is not merely a function $\chi_i \le \epsilon$ 4. It is a task-conditional procedure: learn a prior over imputers, then specialize it rapidly to a new matrix or series.

5. Cross-study, multilevel, and meta-analytic borrowing

In evidence synthesis and multilevel data, meta-imputation often means that the imputation model itself is pooled across studies or clusters. Multiple imputation of incomplete multilevel data using Heckman selection models constructs cluster-specific outcome and selection models, then meta-analyzes their parameters in a second stage (Muñoz et al., 2023). Cluster-specific estimates $\chi_i \le \epsilon$ 5 and their covariance matrices are first obtained from copula-based Heckman fits; then multivariate random-effects meta-analysis and univariate random-effects meta-analysis are used to estimate marginal parameters $\chi_i \le \epsilon$ 6 and heterogeneity $\chi_i \le \epsilon$ 7. Cluster-level imputation parameters are finally drawn from a shrinkage posterior centered between the cluster-specific estimate and the meta-analytic mean, with clusters exhibiting systematic missingness borrowing almost entirely from the pooled distribution (Muñoz et al., 2023). This makes the meta-analysis step a prior-generating layer for imputation.

A related strategy appears in the treatment of partially missing within-study correlations in bivariate and surrogate meta-analysis (Santis et al., 24 Jun 2026). There the missing quantity is not a cell in a feature matrix but the within-study correlation $\chi_i \le \epsilon$ 8. The paper maps $\chi_i \le \epsilon$ 9 to $\chi_i > \epsilon$ 0, fits a Beta regression with covariates $\chi_i > \epsilon$ 1, and generates multiple imputations by drawing regression parameters and then drawing $\chi_i > \epsilon$ 2 from a Beta law. Rubin’s rules are then used after fitting the substantive multivariate meta-analysis model (Santis et al., 24 Jun 2026). The methodological point is that the imputation model is tailored to the geometry of the missing parameter and to the meta-analytic likelihood in which it is used.

Other multilevel formulations pursue latent-structure borrowing rather than explicit meta-analysis. Multilevel singular value decomposition for mixed data decomposes the data matrix into between-group and within-group components and performs SVD on both parts, yielding single imputation procedures for quantitative, categorical, and mixed multilevel data, with a distributed computation scheme for hospital-partitioned datasets (Husson et al., 2018). Multivariate Gaussian Linear Cluster-Weighted Modeling, in turn, treats fully observed auxiliary variables as a mechanism for cluster identification and local Gaussian regression when imputing missing continuous outcomes, and is explicitly presented as a fully Bayesian mixture model well suited to multiple imputation (Masmela-Caita et al., 2023).

These approaches share a structural premise: heterogeneity is not nuisance to be averaged away. It is modeled as the very object through which imputation borrows strength.

6. Inference-oriented meta-imputation

A distinct line uses imputation to enable or correct inference rather than to reconstruct a generic completed table. Gaussian imputation from summary association statistics models a vector of SNP $\chi_i > \epsilon$ 3-scores as approximately multivariate normal with covariance given by LD, partitions the vector into typed and untyped SNPs, and imputes untyped $\chi_i > \epsilon$ 4-scores by conditional expectation, $\chi_i > \epsilon$ 5 (Pasaniuc et al., 2013). Because it operates directly on GWAS or meta-analysis summary statistics, it is described as meta-imputation in the sense that imputation happens at the meta-analysis level rather than at the genotype level. In simulations, it recovered 84% of the effective sample size for common variants and 54% for low-frequency variants, increasing to 87% and 60% when summary LD information from target samples was available (Pasaniuc et al., 2013).

Statistical matching using fractional imputation addresses a different inferential problem: $\chi_i > \epsilon$ 6 and $\chi_i > \epsilon$ 7 are never jointly observed, and identification is supplied by an instrumental variable assumption rather than conditional independence (Kim et al., 2015). For each missing value, the method generates multiple candidate imputes and attaches fractional weights $\chi_i > \epsilon$ 8 proportional to $\chi_i > \epsilon$ 9, then solves weighted estimating equations. Here imputation is not a completed dataset in Rubin’s sense but a weighted empirical distribution over plausible missing values (Kim et al., 2015).

Outcome reporting bias in meta-analysis produces missing outcomes rather than missing cells, and a recent multiple imputation approach handles this by combining a random-effects imputation model with a logistic selection model and importance weighting across imputations (Burgwinkel et al., 8 Jul 2026). The same inferential orientation motivates Imputation-Powered Inference, which uses arbitrary blackbox imputers only to create surrogate scores and then debiases them with complete cases through a one-step update $K$ 0 (Zhao et al., 17 Sep 2025). Under MCAR blockwise missingness, IPI yields valid and efficient M-estimation; under the weaker first-moment MCAR condition, it improves subpopulation inference while retaining statistical validity and supplies explicit diagnostics for the key assumption (Zhao et al., 17 Sep 2025).

In these settings, meta-imputation is best viewed as an inferential instrument. The imputer contributes efficiency; validity comes from conditioning, weighting, or debiasing layered on top of it.

7. Assumptions, uncertainty representations, and methodological boundaries

The surveyed literature differs sharply in what is assumed about missingness. MIB is evaluated only under MCAR with a fixed 10% masking rate and does not simulate MAR or MNAR (Azad et al., 3 Sep 2025). TimeFlow is deterministic, trained with plain MSE, and does not explicitly model missingness mechanisms or uncertainty (Naour et al., 2023). MetaTSI emphasizes cross-city generalizability, but formal guarantees under specific missingness mechanisms are left open (Nie et al., 20 Jan 2025). By contrast, the multilevel Heckman framework is explicitly constructed for MNAR through correlated selection and outcome equations (Muñoz et al., 2023), while IPI guarantees population inference under MCAR and subpopulation inference under first-moment MCAR (Zhao et al., 17 Sep 2025).

Uncertainty is likewise encoded in markedly different ways. CCAI propagates ambiguity through belief masses on meta-classes rather than collapsing it to a point label (Liu et al., 2016). MIB produces a point estimate from a linear combiner, with no explicit uncertainty model (Azad et al., 3 Sep 2025). Matrix-factorization meta-learning and TimeFlow perform task adaptation but still return point predictions after optimization (Iwata, 2021, Naour et al., 2023). Bayesian LCWM yields posterior predictive draws (Masmela-Caita et al., 2023). Beta-regression imputation, Heckman multilevel imputation, and ORB adjustment use multiple imputation and Rubin-style combination (Santis et al., 24 Jun 2026, Muñoz et al., 2023, Burgwinkel et al., 8 Jul 2026). Fractional imputation replaces multiple completed datasets with weighted candidate values in a single augmented dataset (Kim et al., 2015). IPI uses asymptotic linearization and bootstrap approximations for the variance of a debiased one-step estimator (Zhao et al., 17 Sep 2025).

Several limitations recur. Learned ensemble methods can be strong empirically yet depend on how synthetic masking matches real missingness (Azad et al., 3 Sep 2025). Meta-learned adaptation presumes transferable task structure; when tasks are too heterogeneous or too few, generalization degrades (Naour et al., 2023, Nie et al., 20 Jan 2025). Selection-model approaches depend on valid exclusion restrictions or correctly chosen missingness models (Muñoz et al., 2023, Burgwinkel et al., 8 Jul 2026). Summary-statistic imputation depends on accurate reference LD and degrades for low-frequency variants (Pasaniuc et al., 2013).

The literature therefore presents meta-imputation less as a single statistical doctrine than as a recurring methodological move: place a higher-level mechanism above ordinary imputation. That mechanism may decide when to impute, learn how to weight candidate imputations, adapt a prior to a new task, borrow information across clusters or studies, or debias imputation-driven estimators. Across these variants, the defining shift is the same: missing-data completion is no longer a fixed preprocessing routine, but an object of modeling, control, and inference in its own right.