MerQuR+Statistics Glossary

Statistics Glossary → Auxiliary variables (imputation model)

Auxiliary variables (imputation model)concept

Variables that are not part of the main analysis but are used as predictors in the multiple imputation of missing data because they are related to the incomplete variable or to the reason it is missing.

Multiple imputation predicts missing values many times from a model based on the observed data, producing a completed data set each time. Its validity rests on the missing-at-random (MAR) assumption, i.e. that the probability of being missing can be explained by observed variables alone. Adding variables that are strongly related to the incomplete variable, or that predict missingness, to the imputation model makes this assumption more plausible and improves the accuracy of the imputed values. These variables need not appear in the main analysis model.

Auxiliary variables can reduce bias in the estimates and shrink standard errors. Not every addition helps, however: variables weakly related to the incomplete variable complicate the imputation model needlessly, and variables that are themselves heavily missing or nearly identical to analysis variables (multicollinearity) can cause convergence problems. An important rule is that the imputation model must contain all variables in the main analysis (including the dependent variable) and their interactions; otherwise relationships are pulled towards zero.

An everyday example

Some respondents in a survey did not report their income, and the researcher is studying the relationship between income and health. Education and occupation are not the subject of the analysis, but they are strongly related to income and also explain in which groups the non-reporters are concentrated. Adding these two as auxiliary variables to the imputation model makes the imputed incomes far more realistic.

In MerQur

In Multiple Imputation the 'Auxiliary variables' parameter selects variables that do not enter the analysis but act as predictors in the imputation model.

Analyses where this concept appears

Multiple Imputation

Related concepts

Multiple imputation by chained equations (FCS/MICE) and PMM