Bayesian Modeling and Imputation for Missing Mixed Ordinal-Categorical Data in Large-Scale Surveys

• Activity Number: 478
• Type: Topic Contributed
• Date/Time: Wednesday, August 12, 2015 : 8:30 AM to 10:20 AM
• Sponsor: Survey Research Methods Section
• Author(s): Maria De Yoreo* and Jerry Reiter
• Companies: Duke University and Duke University
• Keywords: surveys ; missing data ; bayesian nonparametrics ; Dirichlet process ; ordinal ; multiple imputation

Abstract

Survey data often consist of a large number of categorical variables, some of which are ordinal. Item nonresponse is common in these settings, and can be dealt with using multiple imputation. We distinguish between variables which are fully observed, those which are missing but to a small degree, and those which have a large proportion of observations missing. It is these latter variables we are most concerned with modeling flexibly, in order to provide imputations which preserve the complex relationships in the data. We specify a joint model for the set of variables which are not completely observed, using a flexible Dirichlet process mixture specification for the unordered categorical variables and the latent continuous variables which drive the ordinal responses. This approach allows for capturing complex distributional features and dependencies, leading to a powerful and practical imputation engine.