A comparison of methods for creating multiple imputations of nominal variables

Many variables that are analyzed by social scientists are nominal in nature. When missing data occur on these variables, optimal recovery of the analysis model's parameters is a challenging endeavor. One of the most popular methods to deal with missing nominal data is multiple imputation (MI). This study evaluated the capabilities of five MI methods that can be used to treat incomplete nominal variables: multiple imputation with chained equations (MICE) using polytomous regression as the elementary imputation method; MICE based on classification and regression trees (CART); MICE based on nested logistic regressions; the ranking procedure described by Allison (2002); and a joint modeling approach based on the general location model. We first motivate our inquiry with an applied example and then present the results of a Monte Carlo simulation study that compared the performance of the five imputation methods under conditions of varying sample size, percentage of missing data, and number of nominal response categories. We found that MICE with polytomous regression was the strongest performer while the Allison (2002) ranking procedure and MICE with CART performed poorly in most conditions