Multiple Imputation for Categorical Variables in Multilevel Data

Multiple Imputation (MI) is a technique that imputes a set of plausible values for each missing item using an imputation model. An imputation model predicts a value for the missing item given the observed data and should be compatible with any model that is fitted to the multiply imputed data which is known as the substantive analysis model. To hold the compatibility, the imputation model has to capture the complexities in the substantive analysis model such as the structure of the data set and complex terms (higher order, interaction) considering the type of incomplete variable and the number of incomplete variables in the data set. In this thesis, the application of MI for handling missing values in a set of level-1 categorical variables in a two-level data structure where the data can be fitted with a random intercept substantive analysis model is empirically investigated. Simulation studies considering the missing data rate, the number of clusters and the cluster sizes are based on data generated from a real multilevel educational data set.