Missing Data

Missing data are commonly seen in randomized clinical trials. When missingness is not completely random, a complete-case analysis that ignores the missing data process often leads to biased estimates of the average treatment effect. This chapter defines different missing data mechanisms, discusses their impact on inference, and presents statistical methods that address missing data, including likelihood-based analysis, inverse probability weighting, and imputation. Each of these methods either models the missingness process or the observed outcome distribution. A more robust approach that combines the virtue of each of these modeling approaches is also introduced. This approach is doubly robust such that it yields a consistent estimate of the average treatment effect if either one of the missingness model or the outcome model is correctly specified, but not necessarily both. The chapter concludes with a brief discussion of sensitivity analyses used to assess the impact of unmeasured factors that affect both the missingness and outcomes. Throughout, statistical and practical considerations are discussed in the context of randomized clinical trials where the primary analysis is to compare two treatments and to estimate the average comparative effect among the enrolled population.