When faced with missing values in a dataset, the standard machine learning playbook is straightforward: fill in the blanks by minimizing reconstruction error. Whether using Mean Squared Error (MSE), matrix factorization, or training a deep generative model to match the observed data density, the goal is to make the imputed values as close to the "true" values as possible.
However, as Nhat-Hao Pham et al. demonstrate in Correlation visualization under missing values: a comparison between imputation and direct parameter estimation methods, a low root mean square error (RMSE) does not guarantee that the underlying relationships between variables are preserved. In fact, optimizing purely for reconstruction accuracy can be catastrophic for downstream causal discovery.
This tension arises because standard imputation algorithms rely on full conditional distributions—using every available feature to predict a missing value. As Trent Kyono et al. point out in MIRACLE: Causally-Aware Imputation via Learning Missing Data Mechanisms, correlation-based models do not distinguish between true causal mechanisms and coincidental statistical patterns. When an algorithm uses an auxiliary variable to fill a gap, it can inadvertently introduce artificial correlations, creating what Borja Seijo-Pardo et al. call "statistical ghosts" in Analysis of imputation bias for feature selection with missing data. These hallucinated dependencies distort the joint distribution, leading downstream causal discovery tools to identify false-positive edges (see Anchored Causal Inference in the Presence of Measurement Error by Basil Saeed et al.; Gene Regulatory Network Inference in the Presence of Dropouts: a Causal View by Haoyue Dai et al.).
Why Standard Imputation Distorts Causality
To understand why minimizing reconstruction error breaks causal discovery, consider a simple causal graph where two independent causes influence a single outcome ($X \rightarrow Y \leftarrow Z$). In graphical modeling, $Y$ is known as a collider. Because $X$ and $Z$ are independent, knowing something about $X$ tells you nothing about $Z$.
If values of $Y$ are missing, a standard regression or imputation model will try to reconstruct $Y$ using both $X$ and $Z$. By forcing the imputed values of $Y$ to be a smooth, deterministic function of its parents to minimize prediction error, the imputer inadvertently couples $X$ and $Z$ in the filled-in data. This induces a spurious correlation between the two independent causes, violating the true independence structure of the system.
As Erdun Gao et al. argue in MissDAG: Causal Discovery in the Presence of Missing Data with Continuous Additive Noise Models and Vy Vo et al. explore in Learning Directed Graphical Models with Optimal Transport, statistical distribution recovery is fundamentally misaligned with causal mechanism preservation. An imputer optimized for data density might perfectly mimic the marginal distributions, but it will routinely violate the conditional independence constraints required to orient causal edges.
In time-series data, this problem manifests as artificial temporal dependencies. Smooth interpolations or forward-filling methods are often misinterpreted by discovery algorithms as meaningful physical dynamics rather than artifacts of the imputation process (Time-Series Causality with Missing Data by Mohamed A. Naiel et al.; Causal Discovery for Irregularly Time Series with Consistency Guarantees by Weihong Li et al.). Furthermore, Maya B. Mathur demonstrates in Pitfalls of imputing using incomplete auxiliary variables that automated tools like MICE introduce systematic errors when they blindly condition on all available variables, because the mathematically correct distribution for imputation is often a specific subset dictated by the true, underlying causal graph.
The Chicken-and-Egg Dilemma
This creates a classic paradox highlighted by Vy Vo et al. in Optimal Transport for Structure Learning Under Missing Data and Weihong Li et al. in Causal Discovery for Irregularly Time Series with Consistency Guarantees:
- To accurately impute missing data without introducing bias, we need to know the true causal graph so we only condition on a variable's valid causal parents or its Markov blanket.
- To learn the true causal graph, we need a complete dataset free of imputation artifacts.
If we handle these steps sequentially—imputing first, then learning the graph—we trigger a destructive feedback loop where imputation errors reinforce structural errors in the recovered graph.
[Naive Imputation] ──(Introduces Spurious Correlations)──> [Distorted Causal Graph]
▲ │
└──────────────────(Reinforces Erroneous Priors)─────────────┘
Breaking the Loop with Causally-Aware Imputation
To break this cycle, recent research shifts from sequential pipelines to joint frameworks that force data completion and structure learning to cooperate.
Graph-Guided Conditioning
Instead of using all variables, frameworks like the one proposed by Yang Liu 0206 et al. in Improving the imputation of missing data with Markov Blanket discovery use Markov Blanket discovery to restrict the imputer's conditioning set. By isolating only the direct causes, effects, and spouses of a missing variable, they prevent extraneous information from leaking spurious correlations into the dataset.
Simultaneous Learning and Filtering
The MIRACLE framework by Trent Kyono et al. dynamically learns the missingness mechanism alongside the causal graph, using the graph as a filter to block uncausal feature correlations during imputation. Similarly, Erdun Gao et al. integrate the missing data problem directly into an Expectation-Maximization (EM) loop for structural learning, ensuring that the distribution recovery respects continuous additive noise models.
Stochastic Correction Maps
In non-linear settings where EM scales poorly, Vy Vo et al. introduce an optimal transport framework (OTM). Instead of relying on deterministic reconstruction, they pair a global imputer with a stochastic correction network. This network maps initial noisy imputations to values that strictly adhere to the structural constraints of the currently estimated causal graph, treating structure learning as a density-fitting problem via Wasserstein distance.
[Missing Data] ──> [Global Imputer] ──> [Stochastic Correction Map] ──> [Causally Valid Imputation]
▲
(Enforces Graph Constraints)
│
[Current Causal Graph]
Causal Priors in Time-Series
For irregular time-series, frameworks like CUTS (CUTS: Neural Causal Discovery from Irregular Time-Series Data by Yuxiao Cheng et al.) and ReTimeCausal (Weihong Li et al.) use graph-informed neural networks. They leverage the history of inferred causal parents to predict gaps in their descendants, turning the causal graph into a powerful structural prior for data recovery.
Bypassing Imputation Entirely
When the risk of introducing structural bias is too high, some researchers argue for bypassing imputation altogether. For instance, Haoyue Dai et al. prove in Gene Regulatory Network Inference in the Presence of Dropouts: a Causal View that in single-cell RNA sequencing data, a strategy of "test-wise deletion"—simply ignoring samples where conditioned variables are zero—can perfectly recover gene regulatory networks without the hallucinations inherent to data completion.
Alternatively, Josh Givens et al. develop marginal score matching techniques in Score Matching with Missing Data that allow models to learn the underlying graphical model directly from incomplete data, avoiding the pitfalls of heuristic zero-imputation.
Ultimately, this body of work demonstrates that data completion cannot be decoupled from downstream semantics. To preserve the true mechanisms of a system, imputation algorithms must look beyond pixel-perfect reconstruction and respect the underlying laws of cause and effect.
Key Foundational Work
Anchored Causal Inference in the Presence of Measurement Error (2020) — contributed 2 concept elements to this direction.
Analysis of imputation bias for feature selection with missing data (2018) — contributed 2 concept elements to this direction.