Comparing Multiple Imputation Methods for Correlated Data

• Activity Number: 461
• Type: Contributed
• Date/Time: Wednesday, August 6, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Mental Health Statistics Section
• Author(s): David Kline*+ and Eloise Kaizar and Rebecca R. Andridge
• Companies: Ohio State University and Ohio State University and Ohio State University
• Keywords: multiple imputation ; missing data ; mixed models ; longitudinal data

Abstract

Many applications generate correlated data that can be analyzed with hierarchical methods, such as analyses that involve multiple studies or sites and longitudinal data. Thus, missing data methods for these applications must properly adjust for the clustered structure of the data. Imputation methods have been proposed for these situations, including one-step multivariate methods (PAN) and methods that rely on a sequence of conditional distributions. Motivated by a longitudinal study of behavioral effects of pediatric traumatic brain injury, this simulation study compares the performance of the conditional approach as implemented in MICE to a multiple imputation procedure based on joint hierarchical Bayesian model specification that extends PAN to allow for missing cluster level data. Specifically, interest lies in the performance of the methods for imputation of cluster level missing data. Our analyses indicate that MICE may attenuate estimates of regression coefficients and, perhaps more importantly, standard errors when there are moderate clustering effects.