ABSTRACT
Missing values or incomplete data are commonly encountered in clinical trials. One of the primary causes of missing data is the dropout. Reasons for dropout include, but are limited to, refusal to continue in the study (e.g., withdrawal of informed consent), perceived lack of efŒcacy, relocation, adverse events, unpleasant study procedures, worsening of disease, unrelated disease, noncompliance with the study, need to use prohibited medication, and death (DeSouza et al., 2009). Following the idea of Little and Rubin (1987, 2002), DeSouza et al. (2009) provided an overview of three types of missingness mechanisms for dropouts. These three types of missingness mechanisms include (1) missing completely at random (MCAR), (2) missing at random (MAR), and (3) missing not at random (MNAR). MCAR refers to the dropout process that is independent of the observed data and the missing data. MAR indicates that the dropout process is dependent on the observed data but is independent of the missing data. For MNAR, the dropout process is dependent on the missing data and possibly the observed data. Depending upon the missingness mechanisms, appropriate missing data analysis strategies can then be considered based on existing analysis methods in the literature. For example, commonly considered methods under MAR include (1) discard incomplete cases and analyze complete cases only, (2) impute or Œll in missing values and then analyze the Œlled-in data, (3) analyze the incomplete data by a method such as likelihood-based method (e.g., maximum likelihood, restricted maximum likelihood, and Bayesian approach), moment-based method (e.g., generalized estimating equations [GEEs] and their variants), and survival analysis method (e.g., Cox proportional hazards model) that does not require a complete data set. On the other hand, under MNAR, commonly considered methods are derived under pattern mixture models (Little, 1994) which can be divided into two types: parametric (see Diggle and Kenward, 1994) and semi-parametric (Rotnitzky et al., 1998).
