1 Missingness mechanisms and definitions

1.1 Complete overview of missing-data mechanisms

Missing-data mechanisms describe how the act of not observing a value is related to the data values themselves. In practice, the mechanism is not directly observable, but distinguishing classes of mechanisms clarifies which statistical procedures yield valid inferences.

1.1.1 MCAR (missing completely at random)

Under MCAR, the probability that an observation is missing does not depend on either the observed values or the unobserved (missing) values. This is the strongest and most restrictive form. When MCAR holds, analyses that ignore the missingness often remain unbiased under broad conditions because the missingness process is effectively unrelated to the data-generating variables.

1.1.2 MAR (missing at random)

Under MAR, the probability of missingness may depend on observed information but not on the unobserved values. Intuitively, after conditioning on the observed variables, the missingness process carries no additional information about the missing value itself. MAR is weaker than MCAR and is frequently treated as a pragmatic assumption because it allows systematic missingness patterns that are explainable using measured covariates.

1.1.3 MNAR (missing not at random)

MNAR relaxes the MAR restriction by allowing missingness to depend directly on unobserved values. For example, missingness can intensify when the missing outcome would have been large or small, even after conditioning on all observed covariates. MNAR is often harder to handle because valid inference typically requires extra modeling assumptions that link the missingness process to the unobserved data.

1.2 Formal notation and missing-data indicators

A formal definition of MAR uses a missingness indicator along with observed and unobserved components of the dataset.

1.2.1 Observed vs. unobserved components

Let \(Y\) denote the full set of variables of interest, partitioned into observed components \(Y_{\text{obs}}\) and unobserved components \(Y_{\text{mis}}\). Let \(R\) be an indicator matrix (or vector) encoding which entries of \(Y\) are observed.

1.2.2 Dependence structure of missingness

The missingness indicator \(R\) is treated as a random variable whose distribution may depend on both observed and unobserved components of \(Y\), as well as on auxiliary measured information. The key conceptual split is whether dependence on \(Y_{\text{mis}}\) remains after conditioning on \(Y_{\text{obs}}\).

1.2.3 Conditional probability statement for MAR

MAR is commonly expressed as: \[ \Pr(R \mid Y_{\text{obs}}, Y_{\text{mis}}) = \Pr(R \mid Y_{\text{obs}}). \] Equivalently, once the observed data are known, the probability of missingness is conditionally independent of the unobserved values.

2 Assumptions underlying MAR

2.1 What MAR permits and what it excludes

MAR draws a boundary between dependence on information that is already available to the analyst and dependence on the missing values themselves.

2.1.1 Missingness depending on observed covariates

MAR permits missingness mechanisms that vary with observed predictors, observed outcomes at other times, or other measured variables. For example, if dropout in a study depends on baseline characteristics and observed interim measurements, MAR can be consistent with the observed pattern, provided the missing outcome values do not add further predictive power beyond what is already captured by the observed data.

2.1.2 Missingness not depending on unobserved values

MAR excludes dependence on the specific unobserved values in the conditional distribution of missingness. If two units share identical observed data but have different missing outcomes, MAR requires them to have the same probability of being missing. This exclusion is not directly testable from the incomplete data alone, which motivates diagnostics and sensitivity analysis.

2.2 Relationship to ignorability and likelihood-based inference

Many modern procedures for incomplete data rely on separating the outcome model from the missingness model. MAR is a central component of this separation.

2.2.1 Factorization of the observed-data likelihood

In likelihood frameworks, the joint distribution of the data and the missingness indicator can often be factorized into an outcome component and a missingness component. MAR constrains the missingness component so that it depends only on observed data, simplifying the relationship between the observed-data likelihood and the full-data likelihood.

2.2.2 When the mechanism can be ignored

Under additional regularity conditions, MAR allows the missingness mechanism to be treated as “ignorable” for inference about parameters in the outcome model. Practically, this means that maximizing the likelihood for the outcome using only observed data (or using standard missing-data likelihood methods) can yield valid results without explicitly specifying the missingness distribution.

2.2.3 Implications for standard estimators

With MAR in place, many complete-case methods may still fail if the missingness is uneven across subgroups and those subgroups are not balanced by conditioning on observed covariates. MAR supports more flexible alternatives—such as maximum likelihood under correctly specified models or imputation procedures—rather than guaranteeing validity for naive estimators that discard incomplete records.

3 Identifiability and estimability under MAR

3.1 What can be identified from observed data

Even with MAR, not every quantity about the full data-generating process is recoverable from the incomplete observations. Identifiability depends on what is modeled and what structure is assumed.

3.1.1 Parameters estimable without modeling the mechanism

Under MAR and suitable assumptions, parameters of the outcome model can often be estimated without fully specifying the missingness mechanism. The estimability arises because the likelihood contributions involving observed data can be related to the full-data likelihood through the MAR constraint, allowing nuisance parts to drop out for the target parameters.

3.1.2 Limits when unobserved drivers exist

If important predictors of the missingness are not measured (latent variables, unrecorded severity, unobserved patient behavior, and similar hidden drivers), then MAR as stated relative to the available observed variables may not be credible. In such cases, the observed-data distribution may be compatible with multiple full-data models, making certain causal or predictive claims weak without additional assumptions.

3.2 Modeling strategy under MAR

Under MAR, analysts typically proceed by specifying models for outcomes and, where useful, for missingness.

3.2.1 Specifying the outcome model

The outcome model describes the distribution of \(Y\) given covariates and possibly prior outcomes (in longitudinal settings). Its adequacy is crucial: imputation and likelihood methods both rely on it to represent how missing values relate to observed information. Model misspecification can produce biased estimates even when MAR is true.

3.2.2 Specifying the missingness model (when relevant)

For some methods and diagnostics, modeling missingness can be informative. While MAR may render missingness parameters unnecessary for estimation of outcome parameters, the missingness model can still support checks of whether observed predictors adequately explain the missingness pattern, especially when using inverse probability weighting or when assessing plausibility.

4 Methods that rely on MAR

4.1 Multiple imputation under MAR

Multiple imputation (MI) creates several completed datasets by replacing missing values with draws from an imputation model. Under MAR, validity is often tied to using an imputation model consistent with the assumed conditional distribution of missing values given observed data.

4.1.1 Imputation model specification

The imputation model must include relevant predictors of both the outcome and the missingness process (as far as they are observed). It can be built using regression models, generalized linear models, or flexible approaches such as chained equations. The goal is to reproduce the conditional relationships present in the observed data.

4.1.2 Number of imputations and stability

The number of imputations affects Monte Carlo error: more imputations reduce random variability introduced by imputation. In practice, analysts choose a number that balances computational cost and stability, with the required amount depending on the fraction of missing information and the complexity of the imputation model.

4.1.3 Combining estimates (Rubin’s rules)

After analyzing each imputed dataset, results are pooled using Rubin’s rules. These combine within-imputation variance and between-imputation variance to reflect both sampling uncertainty and uncertainty about the imputed values. The pooled inference then approximates the appropriate variability under the MI framework.

Likelihood-based methods use the incomplete-data likelihood directly or indirectly by leveraging MAR to relate observed-data likelihood contributions to parameters of interest.

4.2.1 Full-information maximum likelihood

Full-information maximum likelihood (FIML) maximizes the likelihood based on observed data under a specified outcome model and the MAR constraint. FIML typically uses all available information rather than discarding incomplete cases, and it can produce efficient estimates when the model is correctly specified.

4.2.2 Expectation–maximization intuition

For models with latent structure, the expectation–maximization (EM) algorithm provides intuition: an E-step uses current parameter values to compute expectations involving missing components, while the M-step updates parameters to maximize the expected log-likelihood. Under MAR, the observed-data likelihood remains the target, with missing values handled through their conditional expectations.

4.2.3 Confidence interval construction

Confidence intervals may rely on asymptotic normality, profile likelihood, or variance estimates derived from the observed information. Under standard regularity conditions, these intervals reflect additional uncertainty introduced by missing data through the observed-data curvature rather than treating the dataset as if it were complete.

4.3 Inverse probability weighting (IPW)

Inverse probability weighting uses the missingness indicator to construct a weighted pseudo-population in which missingness is addressed by design, given the MAR assumption.

4.3.1 Propensity models for missingness

An IPW method begins by estimating the probability of being observed given covariates. These fitted probabilities—often called response probabilities or propensities—define the weights applied to observed units to compensate for those not observed.

4.3.2 Weight truncation and diagnostics

When estimated response probabilities are small, weights become large, inflating variance. Truncation or stabilized weights can mitigate instability. Diagnostics often include checking covariate balance across observed/missing groups and examining weight distributions to ensure that no small-probability region dominates the analysis.

4.3.3 Variance estimation under weighting

Because weighting changes the effective sample and introduces additional uncertainty from estimated propensities, variance estimation typically uses robust (sandwich) formulas, bootstrap methods, or other resampling strategies compatible with the weighting procedure.

5 Diagnostic tools and sensitivity analysis

5.1 Assessing plausibility of MAR

While MAR cannot be proven from incomplete data, analysts can evaluate whether the assumption is consistent with observed patterns.

5.1.1 Comparing observed distributions by missingness status

A common diagnostic compares observed covariate and outcome summaries across groups defined by whether data are missing. If major differences remain after conditioning on candidate predictors, it suggests that missingness may relate to unobserved values or that important predictors are omitted.

5.1.2 Modeling missingness as a function of observed variables

Fitting a model for missingness using observed covariates can reveal whether the observed variables provide strong explanatory power for the missingness mechanism. Poor fit or systematic structure in residuals can indicate that the observed variables are insufficient to satisfy MAR in a practical sense.

5.2 Sensitivity to departures from MAR

Sensitivity analysis evaluates how conclusions change when MAR is altered in structured ways, reflecting the fact that MAR may be only an approximation.

5.2.1 MNAR-leaning scenarios and perturbations

Analysts may introduce perturbations that allow missingness to depend weakly on the unobserved values. By varying the strength and direction of the dependence, one can assess whether results are robust or whether they hinge on strict adherence to MAR.

5.2.2 Pattern-mixture and selection-model ideas (high level)

Two broad strategies appear in sensitivity work. Pattern-mixture approaches model the observed data distribution for different missingness patterns and then connect these patterns to infer about missing values. Selection-model approaches specify how the outcome distribution and missingness jointly arise. In both, MAR is recovered as a special case, allowing controlled departures for investigation.

5.2.3 Reporting robustness and limitations

Good practice includes reporting which conclusions appear stable across plausible deviations and which do not. Because sensitivity methods rely on modeling choices, transparency about assumptions and limitations is essential for interpreting the strength of evidence.

6 Practical guidance for analysts

6.1 Choosing variables for imputation and models

The credibility and performance of MAR-based methods depend heavily on variable selection and modeling strategy.

6.1.1 Including predictors of missingness

Variables related to missingness should be included in the imputation and, when relevant, missingness modeling steps. The rationale is that conditioning on these predictors helps achieve the MAR-style requirement: after accounting for observed predictors, the missingness should not provide further information about the missing values.

6.1.2 Handling auxiliary variables

Auxiliary variables—variables not directly targeted in the analysis but associated with missingness or the outcome—can improve imputation quality. They may help preserve relationships in the observed data and reduce bias when the imputation model would otherwise omit important dependence structures.

6.2 Common pitfalls

6.2.1 Overly restrictive imputation models

If the imputation model is too simple relative to the structure of the data, it may fail to reproduce observed relationships such as nonlinearities, interactions, or complex variance patterns. Even with MAR plausibility, such misspecification can distort estimates and standard errors.

6.2.2 Collapsing categories or misspecification

Converting categorical variables into coarse categories can remove signal and weaken conditional modeling. Similarly, using incompatible link functions or ignoring ordinal structure can produce systematic errors, particularly when missingness is concentrated in certain levels of predictors.

6.2.3 Ignoring missingness indicators when needed

In some analyses, missingness indicators (or missing-data-aware terms) can capture systematic patterns not represented by standard covariates. Under MAR, omitting these indicators may be acceptable in principle only if the covariates fully explain the missingness; otherwise, including missingness indicators can improve predictive modeling.

6.3 Implementation checklist

6.3.1 Data preparation and missingness coding

Analysts should ensure correct identification of missing entries and verify that coding conventions (structural zeros vs. missing values, special codes, out-of-range values) are handled consistently. For MAR-oriented methods, the missingness pattern structure must be correctly represented in the indicator \(R\).

6.3.2 Model checking and convergence checks

For MI and likelihood methods, model diagnostics include checking imputation model fit, assessing convergence in iterative algorithms, and verifying that predictive distributions do not behave pathologically (e.g., implausible extrapolations). Convergence failures or unstable parameter updates warrant reconsideration of model specification.

6.3.3 Transparent documentation of assumptions

Reporting should include the assumed missingness mechanism (MAR relative to specified variables), the variables included in imputation or missingness models, and how results were pooled or weighted. Documentation enables readers to evaluate whether the MAR assumption is reasonable in the context of the data.

7 Worked examples (conceptual)

7.1 MAR in a regression setting

Consider a regression where a response \(Y\) is partly missing, and an observed covariate \(X\) is fully recorded. The missingness of \(Y\) can be systematically related to \(X\).

7.1.1 Linear outcomes with covariate-dependent missingness

Suppose the outcome model is \(Y = \beta_0 + \beta_1 X + \epsilon\), and missingness occurs more frequently for larger values of \(X\). If, after conditioning on \(X\), the probability that \(Y\) is missing is the same regardless of the (unobserved) \(Y\) value, the setting aligns with MAR. An MI approach would impute \(Y\) using the fitted conditional distribution given \(X\), while FIML would estimate \(\beta\) by maximizing the observed-data likelihood implied by the regression model.

7.2 MAR in longitudinal or panel data

In longitudinal studies, missingness often arises through dropout or intermittent nonresponse, and it can depend on previous observed history.

7.2.1 Time-varying covariates and dropout

Assume measurements of an outcome \(Y_t\) are taken over time, but some participants drop out. If dropout at time \(t\) depends on earlier observed outcomes and covariates up to time \(t\), but not on the future unobserved outcomes beyond what is captured by that history, MAR may hold relative to the observed trajectory. Likelihood-based methods for repeated measures or MI methods with history-based predictors can then be applied to reconstruct the missing later outcomes.

7.3 MAR with categorical outcomes

Categorical outcomes require conditional models that respect their measurement scale.

7.3.1 Multinomial/ordinal imputation frameworks

If \(Y\) takes multinomial categories or ordered levels, imputation typically uses multinomial logistic models or ordinal regression models to draw plausible missing categories given covariates. Under MAR, these conditional draws are anchored to observed predictors; proper handling of ordering or category structure helps avoid biased category probabilities and supports valid pooling of estimates across imputed datasets.