Do mixed model and multiple imputations work together for longitudinal data analysis with missing values?

Abstract

Multiple imputations are a simulation-based approach to filling in missing values so that complete-data analysis methods may be employed. In contrast, mixed models are used to analyze data with missing dependent variable values. Studies frequently generate missing values for the dependent and independent variables. Including a large number of independent variables with missing values in mixed model analysis will exclude significant number of data vectors from analysis which creates a major problem of drawing meaningful conclusions. We used multiple imputations to impute missing values for all independent variables in a longitudinal study setting and used the complete data matrix to fit mixed model independently for each of the imputed datasets. We present a method to combine multiple estimates and inferential statistics generated from multiply imputed datasets using the same mixed model with random intercepts and slopes. We account for the within and between imputations variability in a mixed model setting and incorporating several variance-covariance components. We conclude that in some situations it is desirable to use both methods together to draw conclusion.