Missingness Aware Estimator Selection

Choose the missing-data estimator only after stating why values are absent and what assumption makes the target estimand recoverable.

Solution archetype #

640

Problem family

Uncertainty, Evidence & Inference Failure

Problem subfamily

Sampling, Selection, Missingness & Generalization

The Diagnostic Story

Symptom: The report notes how many observations were dropped but says nothing about why those observations were missing. Complete-case analysis is the default not because its assumptions are defensible but because it is what the software does. Multiple imputation was applied as a generic repair, but the imputation model omits key predictors. Different analysts, each making defensible choices, produce different conclusions and cannot agree on which one to trust.

Pivot: State the missingness mechanism assumption before choosing an estimator, not after. Inventory which values are absent and why, verify that the identifying condition for the chosen method holds or at least is plausible, and plan sensitivity bounds for the scenarios where the mechanism assumption cannot be justified.

Resolution: Incomplete data are handled in a way that better preserves the intended population and contrast rather than silently substituting a complete-case population. Stakeholders can see which conclusions depend on which mechanism assumptions. False certainty decreases when non-ignorable missingness remains plausible and that possibility is documented alongside the central estimate.

Reach for this when you hear…

[longitudinal health study] “We dropped everyone who missed a follow-up visit, but those people dropped out because they got sicker — that's not random, that's outcome-dependent.”

[market research] “The survey non-responders were exactly the customers most dissatisfied with us — listwise deletion gave us a satisfaction score from our fans.”

[sensor network analysis] “Sensors fail more often in extreme conditions, so the missing readings are precisely the ones that matter most for the safety model.”

When This Archetype Applies

Partial catalog groundingSome structural conditions are represented by existing abstractions, but no sufficient condition set is fully represented.

Diagnostic problem

An analysis must estimate a target quantity from incomplete data, but the chosen estimator may be valid only under a missingness mechanism that has not been stated, checked, or sensitivity-tested.

Show the applicability expression

Applicability expression5 distinct conditions

Unobserved dataandDifferential missingnessandUnjustified default imputationandIncomplete cases alter estimandandOutcome-dependent absence process

Algebraic1∧2∧3∧4∧5

groundedpartly groundedopen

5 conditions, all required.

5Required in every casenumbered 1–5

These hold no matter which pattern applies.

1

Unobserved data · grounded

Some outcomes, covariates, units, events, waves, readings, or follow-up values are unobserved. more

and — also required

2

Differential missingness · grounded

Missingness patterns differ or may differ across groups, periods, arms, devices, locations, or channels. more

and — also required

3

Unjustified default imputation · open

Deletion, complete-case analysis, simple filling, or software-default imputation is considered without mechanism justification. more

and — also required

4

Incomplete cases alter estimand · open

Excluding or differentially weighting incomplete cases can change the target estimand. more

and — also required

5

Outcome-dependent absence process · open

The absence process can depend on observed covariates, prior values, rules, failures, behavior, or unobserved outcomes. more

Other requirements and context (1)

Why these sit outside the expression

Supporting context — it may accompany or help interpret the situation, but it is not a load-bearing condition in a sufficient diagnostic set.

• Supporting contextA result will inform a high-stakes claim, policy, product decision, scientific inference, safety conclusion, or resource allocation.

Coverage

2 of 5 conditions grounded · 3 open.

Read the methodology · Download the trigger-logic data

Mechanisms / Implementations

• Doubly Robust Missingness Adjustment: Combines outcome modeling with response weighting so estimates can remain consistent if one of the two model components is correctly specified.
• Full-Information Maximum Likelihood Path: Uses likelihood-based estimation with incomplete observed data when model and missingness assumptions are appropriate.
• Inverse-Probability Weighting Model: Weights observed cases by modeled response probability to reduce bias from differential observation when covariates support the response model.
• MCAR Diagnostic Test and Balance Review: Compares complete and incomplete cases and uses MCAR-oriented tests where appropriate while treating non-rejection as limited evidence rather than proof.
• Missingness Indicator Matrix: Creates response indicators and pattern tables that show which records, variables, waves, or sensors are absent.
• Multiple Imputation Workflow: Creates multiple plausible completed datasets, analyzes each, and combines estimates while preserving imputation uncertainty under the stated assumption.
• Pattern-Mixture Sensitivity Model: Models outcomes by missingness pattern and varies unobserved departures to explore MNAR-sensitive conclusions.
• Process-Based Missingness Audit: Uses field knowledge, collection logs, device records, administrative rules, or interview protocols to infer why data became absent.
• Selection-Model Sensitivity Analysis: Models the response process jointly with the outcome to examine how non-ignorable missingness would affect estimates.
• Tipping-Point Analysis: Shows how extreme missing outcomes or response-process assumptions would need to be before the substantive conclusion changes.

Related Abstractions

Abstractions this archetype builds on — directly (a source ingredient) or as a related pattern. Links follow the typed catalog namespace.

Built directly on (2)

• Missing Data Mechanisms (MCAR, MAR, MNAR): MCAR, MAR, MNAR.
• Statistical Inference: Reasoning from a finite, noisy sample back to the underlying population or process while explicitly quantifying the uncertainty that sampling introduces.

Also references 12 related abstractions

• Causality: Cause-effect relationships.
• Confidence Intervals: Range of plausible values.
• Confounding: Hidden variable interference.
• Counterfactuals: Alternate hypothetical scenarios.
• Data Integrity: Accuracy and consistency preserved.
• Robustness: Maintain functionality under stress.
• Sampling (Representativeness): Representative subset selection.
• Selection Bias: Skewed sampling.
• Sensitivity Analysis (in Operations Research): Analyze impact of parameter variation.
• Statistical Power: Probability of detecting effect.

▸ Show 2 more

Variants

Narrower or domain-specific specializations that share this archetype's core structure. Recognized variants are established; candidate variants are provisional.

Complete-Case Defensibility Gate · subtype · recognized

A conservative variant that allows complete-case or listwise deletion only when missingness and estimand-change risks are explicitly checked.

MAR-Conditioned Imputation or Weighting · subtype · recognized

A variant that treats missingness as ignorable only conditional on observed covariates and then uses imputation, likelihood, weighting, or hybrid adjustment.

MNAR Sensitivity-Bounded Inference · risk or failure variant · recognized

A variant that treats non-ignorable missingness as unresolved and reports conclusions across explicit sensitivity scenarios rather than pretending the mechanism is harmless.

Longitudinal Attrition Adjustment · domain variant · candidate

A repeated-measurement variant that adjusts for dropout, intermittent missing waves, or follow-up loss while preserving time-dependent estimands.

Editorial Notes

Problem Classification

Classification: Uncertainty, Evidence & Inference Failure → Sampling, Selection, Missingness & Generalization

Problem kernel: estimator assumptions do not match the missingness mechanism

Rationale: Earliest causal condition: An analysis must estimate a target quantity from incomplete data, but the chosen estimator may be valid only under a missingness mechanism that has not been stated, checked, or sensitivity-tested.

Independent corroboration: The earliest necessary condition in the frozen evidence is: An analysis must estimate a target quantity from incomplete data, but the chosen estimator may be valid only under a missingness mechanism that has not been stated, checked, or sensitivity-tested. That is a sampling selection missingness and generalization problem because Observed cases differ systematically from the target because entry, dropout, missingness, case choice, or reuse beyond the sampled domain is ungoverned.

Review outcome: Independent reviewer agreement; high confidence.