Robustness of a Multivariate Normal Approximation for Imputation of Incomplete Categorical Data

• Activity Number: 256
• Type: Contributed
• Date/Time: Tuesday, August 13, 2002 : 2:00 PM to 3:50 PM
• Sponsor: Section on Survey Research Methods*
• Author(s): Coen Bernaards*+ and Thomas Belin and Joseph Schafer
• Affiliation(s): University of California, Los Angeles and University of California, Los Angeles and Pennsylvania State University
• Keywords: general location model ; missing data ; simulation evaluation

Abstract

Questionnaire data often contains categorical data with both binary outcomes and multiple-ordered answer categories. In practice, questionnaires are often hampered by nonresponse problems. In order to proceed with standard analyses, multivariate normal multiple-imputation implemented in NORM, SAS 8.2 and Splus 6 is a common procedure to handle nonresponse. However, little is known about the robustness of this imputation method when the data are not normally distributed.

A study was carried out where data were simulated under a known logistic regression model with two binary and three continuous predictors. Next, nonresponse was simulated using several methods. The mean and covariance matrix were estimated using both the general location model, the correct model for this problem, and the multivariate normal imputation model. The logistic regression model was estimated from both imputed data models and compared to the known logistic regression model parameters. The talk will report on precision of the multivariate normal imputation model and compare to the general location model.