Multiple Partial Imputation for Incomplete Longitudinal Data

• Activity Number: 260
• Type: Contributed
• Date/Time: Tuesday, August 13, 2002 : 2:00 PM to 3:50 PM
• Sponsor: Biometrics Section*
• Author(s): Steven Shoptaw*+ and Xiaowei Yang and Thomas Belin
• Affiliation(s): Friends Research Institute and University of California, Los Angeles and University of California, Los Angeles
• Keywords: Multiple Partial Imputation ; Missing Data ; Longitudinal Data Analysis

Abstract

In repeated-measures studies, especially in the fields of clinical trials, most datasets are usually incomplete and the missingness structure is complicated, including both intermittent missing and dropout missing. Both of them require special treatment. In this article, missing data problems in repeated-measures studies are fully discussed and a strategy named Multiple Partial Imputation (MPI) is proposed to handle both types of missing data. MPI offers a generic framework within which only intermittent missing data are imputed multiple times, and then these partially imputed datasets can be analyzed by most longitudinal modeling methods, after some modifications, to deal with only dropouts. In real life, the intermittent missing data usually can be assumed of missing at random or completely at random; thus, the data can be imputed by using Markov chain Monte Carlo techniques such as "data augmentation"; but dropouts cannot be assumed so and need to be treated specifically by using both likelihood or quasi-likelihood-based models. A clinical trial of smoking cessation interventions will be used as an illustration.