Missing data frequently occur in clinical trials as well as observational studies. An important source for missing data are patients who leave the study prematurely, so-called dropouts. Alternatively, intermittent missing data might occur as well.
When patients are evaluated only once under treatment, then the presence of dropouts makes it hard to comply with the intention-to-treat (ITT) principle. However, when repeated measurements are taken then one can make use of the observed portion of the data to retrieve information on dropouts. Generally, commonly used methods to analyse incomplete data include complete-case (CC) analysis and, in longitudinal studies, an analysis using the last observation carried forward (LOCF). However, these methods rest on strong and unverifiable assumptions about the missing mechanism. Over the last decades, a number of analysis methods have been suggested, providing a valid estimate for, e.g., the treatment effect under less restrictive assumptions.
The assumptions regarding the dropout mechanism have been classified by Rubin and co-workers as: missing completely at random (MCAR), missing at random (MAR) and missing not at random (MNAR).
In the first part of the course we will review various repeated measurements models and indicate under which missing data mechanism they will provide valid estimates of the treatment effect. Finally, since it is impossible to verify that the dropout mechanism is MAR we argue that, to evaluate the robustness of the conclusion, a sensitivity analysis thereby varying the assumption on the dropout mechanism should become a standard procedure when analyzing the results of a clinical trial.
The second part of the course focuses on multiple imputation (MI), specifically the fully conditional specification (FCS, MICE), which is often considered the gold standard to handle missing data. We will discuss in detail what MI(CE) does, which assumptions need to be met in order for it to perform well, and alternative imputation approaches for settings where MICE is not optimal. The theoretic considerations will be accompanied by demonstrations and short practical sessions in R, and a workflow for doing MI using the R package mice will be proposed.
