|
Abstract:
|
Lack of adherence to study protocol and missing data are unavoidable in clinical trials. So the proper design of trials requires careful specification of the estimand, i.e. what is to be estimated and how post randomization events will be handled in the statistical analysis. Unfortunately many clinical trials were not designed with careful planning of the primary estimand and methods need to be justified post hoc. In this research, we use a real trial in Huntington's disease as an example. Due to data confidentiality, we will use it as a basis for simulation experiments. The simulations will first generate full endpoint data and then apply various missing data mechanisms that are consistent with the observed trial. Several plausible estimands suitable to both regulators and physicians will be defined and statistical methods will be chosen for each estimand. Lastly, the methods will be assessed by the statistical properties (eg bias, mean squared error, etc.). In addition we will assess the differences between the different estimates to evaluate whether an estimand consistent with a mixed model repeated measures analysis provides a reasonable base case with minimal assumptions.
|
