Multiple Imputation for Missing Data Analysis in Proportional Odds Models for Ordinal Response Variables

Abstract

Although multiple imputation (MI) of missing data has been getting popularity in educational research, previous research mainly focuses on normally distributed continuous variables. There is a great need to impute ordinal categorical variables. Further, since there exist various methods for MI, it is unclear which one should be used for specific empirical data. The purpose of this study is to compare and illustrate the implementation of both MI for a single ordinal variable and multiple imputation by chained equations (MICE) for multivariate variables in ordinal logistic regression to predict mathematics proficiency levels. This study helps researchers better understand and implement the methods through comparing various proportional odds (PO) models and the results of these models with different numbers of imputations. For demonstration purposes, the empirical data from the High School Longitudinal Study (2009) are used for the missing data analysis.