Missing at random (MAR) is an assumption about how missing values arise in statistics. It states that, conditional on the information observed, the probability of a particular pattern of missingness does not depend on the values that are missing. Despite its name, MAR does not require missingness to be haphazard or unrelated to the data: it permits systematic differences in missingness explained by observed variables. The assumption is central to likelihood-based analysis and missing-data imputation, because under additional conditions it allows inference without explicitly modeling the process that caused the missing values.
Formal definition
Let denote the complete data that would be available if nothing were missing, and let be a corresponding vector of observation indicators, with when is observed and when it is missing. For a particular missingness pattern , write the complete data as
where the two components contain the values observed and missing under that pattern.
Let the missingness mechanism have conditional probability
where denotes parameters governing missingness. MAR requires that, for the relevant pattern and observed values,
Thus, holding the observed information fixed, changing the unobserved values does not change the probability of that missingness pattern.
The pattern-specific notation matters: which values count as observed depends on . Informal statements such as “missingness is independent of the missing data given the observed data” convey the main idea, but can obscure this dependence.
A common special case
Suppose is fully observed, may be missing, and indicates that is observed. MAR becomes
Equivalently,
expressing conditional independence between observation status and , given .
Where , this implies
Within a group defined by the same , the observed outcomes therefore have the same conditional distribution as the outcomes in the full group. The distribution of among observed cases may nevertheless differ substantially from its distribution in the full population.
Realized and everywhere MAR
Formal treatments distinguish two versions:
- Realized MAR requires the restriction for the missingness pattern and observed values actually encountered, across the possible completions of those data.
- Everywhere MAR requires the restriction for every possible missingness pattern and every compatible set of data values.
The distinction matters when specifying the scope of ignorability results. A condition sufficient to factor the likelihood for an observed dataset is not automatically sufficient for every repeated-sampling claim about an analysis procedure.
Comparison with other missingness mechanisms
MAR is commonly distinguished from two other assumptions.
| Mechanism | Dependence allowed |
|---|---|
| [[missing-completely-at-random | Missing completely at random]] (MCAR) |
| Missing at random (MAR) | Missingness may depend on observed values, but not additionally on missing values. |
| [[missing-not-at-random | Missing not at random]] (MNAR) |
MCAR is a stronger restriction than MAR. In the simple setting with fully observed and partially observed ,
MAR does not imply that the observed cases form a simple random sample of the original dataset. It also does not mean that the missing proportion is small or that missing and observed cases have the same marginal distribution.
Example
Consider a survey in which age is recorded for everyone but income is sometimes missing.
- If an independent random device determines whether income is recorded, the mechanism is MCAR.
- If recording income depends on recorded age, but not on income within age groups, the mechanism is MAR.
- If people with higher incomes are less likely to report income even among people of the same recorded age, the mechanism is MNAR relative to the available information.
Under the MAR scenario, respondents and nonrespondents may have different average incomes because they have different age distributions. An unadjusted mean of reported incomes can therefore exhibit bias even though MAR holds.
Why MAR matters for inference
The modern framework for distinguishing missingness mechanisms and conditions for ignoring them was developed by Donald Rubin in 1976. Its central insight is that the distribution of the complete data and the mechanism determining what is observed are distinct components of a statistical model.
Suppose models the complete data and models missingness. The likelihood based on the observed values and the missingness pattern is
with summation replacing integration for discrete missing values.
Under MAR, the missingness factor does not depend on the missing values and can be taken outside the integral:
If and are distinct parameters, meaning their joint parameter space is a Cartesian product without constraints linking them, likelihood inference about can disregard the missingness factor. The relevant observed-data likelihood is then
This is the principal sense in which missingness can be ignorable. Ignoring the missingness mechanism is not the same as ignoring the missing observations or analyzing only complete records.
For Bayesian inference, a standard additional requirement is an appropriate factorization of the prior distribution, commonly independence between the data-model and missingness-model parameters:
MAR alone therefore does not establish ignorability for every inferential purpose. Correct model specification and the relevant parameter conditions remain necessary, and repeated-sampling properties of particular estimators require separate justification.
Analysis methods under MAR
Observed-data likelihood
Maximum likelihood estimation can use the observed-data likelihood directly, integrating over missing values rather than filling them with fixed guesses. Computation may use the expectation–maximization algorithm or other numerical methods.
For example, a joint model for several measurements can use each record’s observed components, including records that are incomplete. This use of partial information distinguishes observed-data likelihood from complete-case analysis.
Multiple imputation
Multiple imputation generates several plausible completed datasets using predictive distributions for missing values. Each dataset is analyzed, and the results are combined to reflect both within-imputation uncertainty and uncertainty about the missing values.
Under a MAR-based analysis, the imputation model needs to represent the relevant observed predictors of missingness and of the missing values. Variables included in the substantive analysis, nonlinear relationships, and interactions may also matter. MAR does not protect an analysis against a poorly specified imputation model.
Replacing every missing value with one predicted value is not equivalent to multiple imputation. Treating such replacements as known observations generally fails to represent missing-value uncertainty and can distort standard errors.
Inverse-probability weighting
In the simple setting with fully observed and partially observed , let
Under MAR and suitable positivity and integrability conditions,
Here the weighted contribution is evaluated for observed outcomes; missing outcomes contribute zero. This identity motivates inverse-probability weighting: observed cases receive weights inversely proportional to their probabilities of being observed.
The method requires an adequate model for observation probabilities. Very small probabilities can produce large weights and unstable estimates.
Complete-case analysis
Deleting every record containing a missing value is not generally justified by MAR. It can alter the distribution of variables relevant to the target quantity and reduce precision.
Complete-case analysis can nevertheless be valid under particular conditions. For example, when predictors are fully observed, only the outcome is missing, and observation depends on but not on conditional on , complete cases retain the same conditional outcome distribution given . A correctly specified conditional regression model may then be estimable from those cases. This does not automatically justify estimating marginal population quantities from their unadjusted distribution.
Assessment and limitations
MAR is generally not testable from observed data alone
The essential restriction concerns values that are not observed. Without additional assumptions, different MAR and MNAR models can produce the same distribution of observed data. Consequently, a good model fit or a nonsignificant diagnostic test does not establish MAR.
Relationships between missingness indicators and observed variables can provide evidence against MCAR. They do not determine whether missingness also depends on unobserved values after conditioning on the observed information.
This is a problem of statistical identification, not merely one of sample size: arbitrarily large observed samples need not distinguish competing assumptions about unseen outcomes.
MAR depends on the available information
MAR is relative to the variables observed and included in the conditioning information. A mechanism that depends on an omitted variable may appear MNAR when that variable is unavailable, but may satisfy MAR if it is measured for everyone and adequately accounts for the dependence.
Collecting auxiliary information can therefore make MAR more plausible. Simply adding many predictors, however, does not prove the assumption or guarantee a correct analysis model.
Positivity and support
MAR alone does not ensure that all missing values can be learned about from observed cases. If a subgroup has no observed outcomes, its outcome distribution cannot generally be identified nonparametrically from the data.
In the simple outcome-missingness setting, a common requirement is
throughout the covariate support relevant to the target population. Otherwise, inference for some groups necessarily relies on additional information or model-based extrapolation.
Sensitivity to departures from MAR
Because observed data alone generally cannot establish MAR, sensitivity analysis can examine how conclusions change under specified departures from it. One approach shifts the conditional mean of missing outcomes away from the value predicted under MAR:
Here gives equality of the two conditional means, while nonzero values describe particular departures. Equality of means is weaker than the full distributional equality implied by MAR, so this parameterization addresses only one aspect of possible nonrandom missingness.
More general sensitivity analyses specify models for outcome-dependent missingness or for outcome distributions within missingness patterns. Their additional assumptions describe aspects of the missing values that the observed data do not identify.