A Hybrid Model Incorporating Reasons for Dropout: A Simulation Study

• Activity Number: 244
• Type: Contributed
• Date/Time: Monday, July 30, 2012 : 2:00 PM to 3:50 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Jingjing Chen*+ and Fang Liu and Jeff Davidson
• Companies: and Octagon Research Solutions and Octagon Research Solutions
• Keywords: missing not at random ; missing at random ; multiple imputation ; pattern-mixture model

Abstract

Missing data are a part of almost all longitudinal data with dropouts. When data missing is caused by study design, the missingness can not be assumed as at random (MAR). It would typically be assumed to be missing not at random (MNAR). For example, in a randomized withdrawal design of active drug vs. placebo (typically used in pain studies), there is the intention that by withholding active treatment in one arm (placebo), subjects who are randomized to that arm will be unable to remain in the trial. We propose a hybrid model that incorporates single imputation for MNAR classified by dropout reasons and multiple imputation for MAR. A simulation study using the randomized withdrawal design is presented to illustrate how our method works in practice. Subjects are classified into random and nonrandom dropout groups by withdrawal reasons and missing data are subsequently imputed by multiple imputation and single imputation. We compare our proposed hybrid model with other missing data imputation methods without MAR assumption such as last observation carried forward (LOCF), baseline observation carried forward (BOCF), and pattern-mixture model (PMM).