A Comparison of Methods for Analysis of Longitudinal Categorical Data with Dropouts

• Activity Number: 194
• Type: Contributed
• Date/Time: Monday, August 5, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Takayuki Abe*+ and Yuji Sato and Manabu Iwasaki
• Companies: Keio University School of Medicine and Keio Univeristy School of Medicine and Seikei University
• Keywords: longitudinal data ; categorical response ; missing data ; multiple imputation ; GEE ; clinical studies

Abstract

In clinical studies, patients sometimes discontinue the study for various reasons (e.g. lack of efficacy, adverse experiences), and part of the data on the primary variable becomes missing. From the Intention-to-Treat principle, statistical methods should include all randomized patients in analysis, and some methods for missing data have been proposed. In our research, we focused on longitudinal categorical data (specifically ordinal response), and the comparison between two treatment groups (e.g. new drug and placebo), because analysis methods for categorical incomplete data have not been sufficiently evaluated. Some methods; complete-case analysis, last observation carried forward (LOCF), multiple imputation (MI) with auxiliary variables, weighted generalized estimating equations (WGEE), generalized linear mixed-effects models (GLMM), and a combination of these (e.g. MI + GEE) were evaluated. Actual data from a clinical study on antidepressants was used in these evaluations. Simulation studies were performed under some missing mechanisms, and sample sizes. MI + GEE method showed higher performance (i.e. coverage probability and bias), specifically in small sample clinical trials.