Testable Implications in Missing Data Models

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

Version 1 of 117 papers30 conceptsfrozen 2026-10-04

In traditional statistics, missing data is often treated as a nuisance to be assumed away. Researchers typically invoke the "Missing at Random" (MAR) assumption—the idea that the reason data is missing depends only on observed information—and proceed with their analysis. The problem is that MAR is famously impossible to verify using the data itself. If the data is missing because of something you didn't see (Missing Not at Random, or MNAR), your results are biased, and you have no way of knowing it.

However, a growing body of research suggests that we don't have to fly blind. By using graphical models and structural constraints, we can identify "testable implications"—empirical fingerprints in the observed data that allow us to falsify or validate our missingness assumptions.

The Myth of Untestability

The long-standing belief was that if you don't have the missing values, you can't possibly know why they are missing. Karthika Mohan and Judea Pearl challenged this by introducing "m-graphs"—graphical models that explicitly represent the causal mechanisms behind missingness.

The intuition is simple: any causal model makes claims about which variables don't affect each other. In an m-graph, these claims translate into conditional independence requirements. If your model says that a missingness indicator $R$ should be independent of a variable $Y$ given $X$, but your observed data shows a strong correlation between them, then your model is wrong. Mohan and Pearl proved that even complex MNAR models, once thought to be untestable black boxes, often produce these "manifest implications" that can be refuted by the data we actually have.

Finding Empirical Fingerprints

How do we find these implications in practice? Jin Tian showed that we can use d-separation—a standard tool in causal inference—to detect when a hypothesized structure is incompatible with the data. If the graph implies that an observed variable and a missingness indicator should be independent, but they aren't, the MAR assumption (or the specific MNAR structure) is falsified.

This idea of "empirical fingerprints" is central to the work of Razieh Nabi and Rohit Bhattacharya. They argue that most identifiable models for missing data impose restrictions on the observed distribution. Even without knowing the functional form of the relationships, the mere structure of the model creates boundaries. If the data falls outside those boundaries, the model is invalid.

Strengthening the Assumptions to Enable Testing

Sometimes, the standard MAR assumption is too weak to test. Iavor Bojinov and Donald B. Rubin introduced a stronger condition called "Missing Always at Random" (MAAR). While MAR is a property of a specific dataset, MAAR is a property of the mechanism itself.

By requiring that missingness depends only on variables that are fully observed (like baseline demographics), researchers can use diagnostic tests—such as comparing conditional means across different missingness patterns—to see if the data behaves as the model predicts. As Bojinov notes, having stable, fully observed covariates acts as an "anchor" that makes these violations visible.

Structural Constraints as Safeguards

Beyond graphs, adding mathematical structure can also make missingness testable. Manfred Jaeger demonstrated that applying restricted parametric structures makes it possible to statistically test the validity of the MAR assumption. Similarly, Daniel O. Scharfstein and colleagues found that applying Markovian restrictions—assuming that missingness at one time point only depends on a limited window of previous events—creates testable implications that reduce computational complexity and allow for empirical verification in long-term studies.

Even in the realm of machine learning, interpretability can serve as a diagnostic tool. Zhi Chen and Cynthia Rudin showed that "glass-box" models like Explainable Boosting Machines (EBMs) can visualize missingness mechanisms. By looking at the learned scores for missing categories, researchers can see if a feature is truly "Missing Completely at Random" or if the missingness itself is carrying a hidden signal that could bias a clinical model.

From Assumption to Verification

The shift in this research area is from assuming a missingness mechanism to diagnosing it. Whether through the graphical rules of Mohan and Pearl, the diagnostic tests of Bojinov and Rubin, or the structural modeling of Franks and Rubin, the goal is the same: to use the data we have to prove that our theories about the data we don't have are consistent with reality.

Papers in this version

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_testable_implications_in_missing_data_mo_bd1cce02,
  title = {Testable Implications in Missing Data Models},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/testable-implications-in-missing-data-models/rdr_bd1cce02a7f84f36ad2805fb27a1e883},
  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_testable_implications_in_missing_data_mo_bd1cce02_v1,
  title = {Testable Implications in Missing Data Models},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/testable-implications-in-missing-data-models/rdr_bd1cce02a7f84f36ad2805fb27a1e883/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}