MerQuR+Statistics Glossary

Statistics Glossary → Missing-data mechanisms (MCAR, MAR, MNAR)

Missing-data mechanisms (MCAR, MAR, MNAR)concept

Rubin's (1976) three-way classification of how the probability that a value is missing relates to the observed and unobserved values in the data.

Under missing completely at random (MCAR), the probability of missingness is unrelated to any variable, and the incomplete cases are like a random subsample of the complete ones. Under missing at random (MAR), missingness can be explained by observed variables but does not depend on the missing value itself; for example, older people skipping the income question more often is MAR if age is known. Under missing not at random (MNAR), the probability of missingness depends on the missing value itself even after the observed variables are taken into account, for example when high earners decline to report their income.

The choice of method depends on the mechanism. Listwise deletion is generally unbiased only under MCAR and loses power. Full information maximum likelihood (FIML), EM and multiple imputation are valid under MAR. Under MNAR no standard method is unbiased by itself; selection models, pattern-mixture models or a δ-shift sensitivity analysis are needed. Little's MCAR test can reject MCAR, but the distinction between MAR and MNAR cannot be tested from the observed data. It is therefore good practice to include auxiliary variables related to missingness and to carry out a sensitivity analysis.

An everyday example

In a depression study some patients miss the follow-up visit. If missing visits are related to how busy patients are at work, and work status is recorded, the missingness can be treated as MAR. But if patients stay away precisely because their depression has worsened, the missingness is MNAR, and an analysis of only those who attended makes the treatment look more successful than it is.

In MerQur

In multiple imputation, 'Missing value analysis: t tests and categorical tables' compares variables by missingness indicator and is added below Little's MCAR row. 'MNAR sensitivity variable and MNAR δ shifts' gives a Rubin-pooled sensitivity (tipping-point) table for each δ, while the main result stays at δ = 0 (MAR). In latent class analysis 'Missing indicators (Listwise deletion / FIML)', and in confirmatory factor analysis and structural equation modelling 'Missing data (Listwise deletion, FIML)', carry out FIML estimation under the MAR assumption.

Analyses where this concept appears

Latent Class Analysis (LCA)Confirmatory Factor Analysis (CFA)Structural Equation Modeling (SEM)Multiple Imputation

Related concepts

Full information maximum likelihood (FIML)Rubin's rules (pooling imputation results)Sensitivity analysis