Multiple Imputation in the Presence of Derived Variables

• Activity Number: 614
• Type: Invited
• Date/Time: Thursday, August 8, 2013 : 8:30 AM to 10:20 AM
• Sponsor: ENAR
• Author(s): Manisha Desai*+ and Aya Mitani and Thomas Robinson
• Companies: Stanford University and Stanford University and Stanford University
• Keywords: multiple imputation ; longitudinal study ; derived variables ; passive imputation ; active imputation

Abstract

Multiple imputation (MI) in the presence of derived variables can be challenging. We face this issue in Stanford GOALS, a randomized clinical trial designed to assess an intervention to reduce obesity in children and adolescents, where the primary outcome is the rate of change in body mass index. All subjects enrolled in the study will have baseline measurements and up to 3 annual follow-up measurements. We anticipate that not all subjects, however, will have complete data. Of concern is that subjects with only baseline measures will be missing the outcome. Previous research evaluating MI of higher order terms suggests "active" approaches (in our case, directly imputing slope) may be superior to "passive" approaches (imputing underlying values before deriving slope). However, our group has demonstrated a more nuanced interpretation of the performance of passive and active approaches. We present a simulation study that characterizes the properties of estimates produced by MI under variations of passive and active approaches in the longitudinal setting when slopes are the outcome of interest.