Compare MI-GEE and SIMEX-GEE for Correlated Binary Data with MAR Missingness

• Activity Number: 244
• Type: Contributed
• Date/Time: Monday, July 30, 2012 : 2:00 PM to 3:50 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Yu-yi Hsu*+ and Yongming Qu and Alicia L. Carriquiry
• Companies: Iowa State University and Eli Lilly and Company and Iowa State University
• Keywords: MAR ; SIMEX ; GEE ; multiple imputation

Abstract

The method of generalized estimating equations (GEE), which estimates marginal effects for repeated measures, requires that missing data be missing complete at random for the estimates of the mean and the variance components to be consistent. In some special cases, the GEE estimators are still consistent even when the missing data are missing at random (MAR). When the data are MAR, one approach that has been suggested is to use multiple imputation (MI-GEE) which yeilds consistent estimators under MAR. We extend the simulation-extrapolation (SIMEX) method proposed by Cook and Stefanski (1994) for normal measurement error models to the case where we wish to analyze a dataset where some observations are missing at random. The SIMEX method relies on fewer assumptions about the distribution of the missing data but is more computer intensive than the MI method. We use simulation to compare the relative bias and the estimated variance of MI-GEE and SIMEX-GEE estimators.