A SAS Macro for Applying Multiple Imputation to Multilevel Data

Single-level multiple imputation procedures (e.g., PROC MI) are not appropriate for multilevel data sets where observations are nested within clusters. Analyzing multilevel data imputed with a single-level procedure yields variance estimates that are biased toward zero and may yield other biased parameters. Given the prevalence of clustered data (e.g., children within schools; employees within companies; observations within people), a general approach is needed for handling missing data in multilevel data sets. This paper describes a SAS macro, MMI_IMPUTE, that performs multiple imputation for clustered data sets with two levels. The macro uses a Bayesian implementation of the mixed linear model to generate imputations for lower-level incomplete variables, and uses single-level procedures similar to those used in PROC MI to generate imputations for cluster-level variables.