The NRC report on missing data was highly critical of past practices in drug trials and recommended changes in design and conduct, sophisticated methods of analysis, and careful attention to the definition of causal estimands. In protocols and discussions with sponsors since the report appeared, we have seen relatively minor changes in design and conduct, somewhat simplified versions of the sophisticated methods of analysis, and indications of some confusion about what was meant by causal estimands.
We will explain what we think causal estimands are and what estimands may be appropriate for regulatory trials. Some estimands indeed require a substantial change in design and conduct, namely the continuing collection of primary outcome data on most subjects notwithstanding discontinuation of treatment, and we will discuss methods for achieving this continuing collection. Other estimands consider the discontinuation itself to be the outcome and therefore do not require further information on subjects after discontinuation. Still other estimands depend only on data before discontinuation; these estimands may have some usefulness in certain settings, but they cannot be assumed to be sufficient in all cases.
We will talk about the general classes of methods that may be used for testing and estimation of various estimands, but we will not broadly recommend particular methods. Our message (and NRC’s, we think) is that there is no general solution for all missing data problems. Rather, methods need to be chosen in light of careful thinking about what the specific problem is.
At the completion of this tutorial, participants should be able to:
- Discuss the Office of Biostatistic’s current thinking on handling dropouts from clinical trials
- Identify general classes of methods used in testing and estimation of various estimands