|
Abstract:
|
Since the issuing of the ICH Addendum on estimands in 2021, many clinical trials are required to use a treatment policy estimand. Under this estimand, subjects who discontinued treatment may contribute to the estimate of the treatment effect. Models for missing data of these subjects are sometimes based on subjects who discontinued treatment but who stayed in the study. Such subjects may be few in number, so other “inclusive” approaches try to make use of all subjects. In alleviating the sparsity problem in this way, we may encounter bias problems. This presentation looks at bias. If patterns of outcomes of subjects who discontinue study treatment early differ from those who stay on treatment, it may be difficult in “inclusive” approaches to avoid bias, despite the sophisticated models that have been proposed. In addition, sparsity may force compromises in the model for the missing data; sparsity can also lead to instability in the planned model, sometimes leading to systematic bias in the estimate of the treatment effect. This presentation assesses the sources and significance of such bias using simulations based closely on real repeated-measures clinical outcomes.
|
