Multiple Imputation of Missing Categorical Data: A Latent Class Analysis Approach

• Activity Number: 419
• Type: Contributed
• Date/Time: Wednesday, August 5, 2009 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Mulugeta Gebregziabher*+ and Stacia M. DeSantis
• Companies: Medical University of South Carolina and Medical University of South Carolina
• Keywords: case-control ; latent class ; multiple imputation

Abstract

We propose a multiple imputation approach that is based on a latent class imputation model for analyzing case-control data with missing values of a categorical covariate. We study its statistical properties via an extensive simulation study and make comparisons with existing methods under seven missing data mechanism scenarios that include missing completely at random, missing at random, and not missing at random. Comparisons with complete case analysis, multiple imputation, loglinear imputation, and EM algorithm are made with respect to bias, asymptotic standard error (ASE), and 95% coverage probabilities of parameter estimates. We show the proposed method is superior under many missingness scenarios, resulting in smaller bias, ASEs and better coverage probabilities. A data example from a case-control study of multiple myeloma and polymorphisms of IL6 genes are considered.