Can missingness depend on observed variables and still be ignorable?

Yes, under the relevant conditions.

Can the assumption be proved from observed data alone?

Usually not; substantive knowledge and sensitivity analysis remain important.

AI Glossary / Statistics and probability / Intermediate

Ignorability

Definition

Ignorability is a condition under which a data-collection, missingness, or assignment mechanism need not be modeled explicitly for a specified inferential task. Its precise requirements depend on the framework. In missing-data likelihood analysis, common sufficient conditions are missing at random and distinct parameters; in causal inference, related terminology concerns treatment assignment being independent of potential outcomes after conditioning on appropriate covariates.

What it means

Ignorable does not mean unimportant or harmless. It means that, under stated assumptions and for a particular analysis, inference can proceed without explicitly estimating one part of the data-generating process.
The observed variables needed to justify that simplification still matter. If missing responses depend on recorded customer segment, an analysis may need to condition on segment even though it does not separately model the response mechanism. Simply analyzing the available values as if they were a random sample can remain misleading.

How it works

For missing data, distinguish the complete-data model from the mechanism determining which values are observed. Let theta govern the outcomes and phi govern missingness. The observed likelihood generally involves both mechanisms and integration over missing values.
Missing at random means, for the relevant missingness pattern, the missingness probability does not depend on unobserved values after conditioning on observed information. It permits missingness to depend on observed variables. Missing completely at random is stronger and rules out dependence on the data values more broadly.
Distinctness means the parameters of the outcome and missingness models are not linked by restrictions that make one informative about the other. For standard Bayesian ignorability results, an appropriate prior factorization, commonly prior independence of the parameter sets, is also needed. State the mode of inference rather than treating one short definition as universal.
Under the relevant conditions, the factor involving missingness can be omitted when obtaining inference for the outcome-model parameters from the observed-data likelihood. This does not authorize arbitrary complete-case analyses, mean replacement, or pretending imputed values were observed without uncertainty.
In causal inference, a common conditional-ignorability statement is that treatment assignment A is independent of the potential outcomes given covariates X. It is often described as no unmeasured confounding or conditional exchangeability. Terminology varies: strong ignorability commonly includes positivity as well.
Positivity requires a positive probability of each compared treatment within the relevant covariate groups. Consistency and other design-specific conditions are also needed to connect observed outcomes with potential outcomes. Ignorability alone is not a complete causal identification argument.
Assess the plausibility of the assumption using substantive knowledge, the collection process, and measured covariates. Observed data alone generally cannot establish that missingness is unrelated to unobserved outcomes or that all confounding has been measured.
Use sensitivity analysis when plausible departures could change conclusions. Alternative missingness models, external information, and targeted follow-up can reveal how much the inference depends on untestable assumptions. Diagnostics can uncover problems without proving that no hidden problem remains.
For AI applications, ask whether labels are observed selectively. Human reviewers may inspect only difficult cases, customers may report only certain failures, or a deployed policy may determine which outcomes become visible. The resulting selection assumptions should be explicit before a model is evaluated or used causally.

Three practical examples

A survey analyst models outcomes conditional on recorded segment when response depends on segment but, by assumption, not on the missing outcome within segment.
An observational study states which measured pre-treatment covariates are assumed sufficient to remove confounding.
An AI team examines whether selective review of model outputs makes its labeled evaluation data systematically different from unreviewed cases.

Think of it this way

Think of a collection rule that can be set aside only after you have used the information that makes it irrelevant to the particular calculation. Ignorability is permission under conditions, not permission to forget how the data arrived.

Worked example

Consider two recorded groups, A and B, each containing 100 cases. For a deliberately transparent example, every A outcome equals 10 and every B outcome equals 30. Suppose 80 A outcomes and 20 B outcomes are observed, with selection otherwise unrelated to outcomes within each group.
The observed-case mean is (80*10 + 20*30)/100 = 14. The full-population mean is (100*10 + 100*30)/200 = 20. The response mechanism depends on the observed group, so missingness can be at random conditional on group without being completely at random.
Using the known equal group sizes and the observed group means gives 0.5*10 + 0.5*30 = 20. The example shows why the group information cannot be discarded merely because a missingness mechanism is called ignorable.
With variable outcomes, group means would themselves have sampling uncertainty. If unobserved high outcomes were less likely to be recorded even within a group, the stated ignorability justification would fail. The arithmetic does not test that assumption.

Illustrative case study

In this fictional case, Juniper Feedback learns that premium accounts answer a service survey more often than basic accounts. Analysts initially report the raw respondent mean. They then model the outcome using account type and other relevant recorded information, documenting the assumption that response is conditionally unrelated to the missing answers. A sensitivity analysis considers dissatisfied nonrespondents, and the final report explains how that possibility affects the conclusion.

Common misconceptions and mistakes

Ignorability is not equivalent to missing completely at random. It does not mean the observed cases can always be analyzed without adjustment. A missingness assumption is not automatically the same as a causal treatment-assignment assumption.
No observed imbalance does not prove no unmeasured confounding. Adding every available variable is not automatically appropriate causal adjustment; post-treatment variables and selection variables can introduce problems. Statistical sophistication does not make an untestable assumption true.

Business use

The concept clarifies which selection assumptions support a survey result, an AI evaluation, or an observational comparison. Document the inferential target, variables used for conditioning, and evidence about collection. Where conclusions depend heavily on unverifiable assumptions, show sensitivity and identify additional data that would be useful.

Check your understanding

Why was the raw mean 14 rather than 20 in the example?

Answer: The lower-outcome group was overrepresented among respondents.

What extra condition is commonly included in strong ignorability for treatment assignment?

Answer: Positivity, alongside the relevant conditional independence assumption.

Related terms

Selection bias
Sensitivity analysis
Missing at random
missing completely at random
conditional exchangeability
positivity

Learn next

Study missing-data likelihoods and missingness mechanisms. Separately learn causal identification assumptions, then examine how data collection and treatment assignment affect the intended inference.

Sources

Gelman and colleagues, Multiple imputation for model checking: https://sites.stat.columbia.edu/gelman/research/published/completed5.pdf Rubin, Inference and missing data (1976): https://academic.oup.com/biomet/article-abstract/63/3/581/270932 The role of exchangeability in causal inference: https://arxiv.org/abs/2006.01799

Previous: IDistance Next: Ignore delete lists