Joint Modeling of Incomplete Data with Mixed Variable Types Using Latent-Variable Models

• Activity Number: 397
• Type: Invited
• Date/Time: Tuesday, August 6, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Survey Research Methods Section
• Author(s): Thomas R. Belin*+
• Companies: UCLA Department of Biostatistics
• Keywords: multiple imputation ; correlated binary data ; factor analysis ; parameter-extended ; Metropolis-Hastings ; multivariate normal model

Abstract

It is common in research to have many variables measured on a modest number of cases and to have a variety of data types (e.g., continuous, binary, ordinal, nominal). Longitudinal data and other clustered data structures are also common. This talk will present methods that have emerged in an effort to develop flexible model-based imputation methods for high-dimensional data sets. Key ideas include handling missing continuously-scaled items using a factor-analysis strategy to reduce the number of covariance parameters in a multivariate normal model, using growth-curve models and factor-analysis ideas together for longitudinal continuously-scaled variables, using a parameter-extended Metropolis-Hastings algorithm to sample the correlation matrix in a multivariate probit model in a way that can be extended to several ordinal variables, and applying the parameter-extended Metropolis-Hastings idea to a multinomial probit model in a manner that can be extended to several nominal categorical variables. Examples are offered to illustrate the methods, and simulation studies are used to explore statistical properties of the procedures and compare them with alternative approaches.