State of the art adaptive dose finding designs
Research Objectives:
During development of a new drug, a key decision is the dose to bring forward. The choice of dose is supported by a dose-finding study, in which a number (often three) doses of the drug in development is examined in comparison to placebo (dose 0). Improving the design of such studies for better choice of dose could result in higher probability of hitting the best dose among those studied, or allow to evaluate more doses, which could eventually give a finer resolution among potential doses. One such improvement could be to allow for adaptation, that is, to modify the doses used in the study on the basis of results from within the study, on an ongoing basis. One set of existing methods (often used in cancer studies) is the continual reassessment type, which applies a logistic dose-response model. However, it is derived in a setting of cancer drug toxicity, which implicitly implies that a placebo group is not considered. Another set is based on the linear normal dynamic model (used in the ASTIN stroke study), which is a Bayesian model focusing on the change from one dose to the next. This approach may not be optimal for the problem considered. It would be most useful to have intermediate methods, which are more flexible than the logistic model and are more specific than the linear normal dynamic model.
Description of work:
The project should develop one or more such models and study the statistical properties of corresponding study designs. The project is suggested due to a specific issue in a drug to treat stroke, but a method that comes out of the project may make sense in other indications (diseases) as well. The first challenge in such modelling and conduct of adaptive designs is that many practical aspects have to be considered. For example, in stroke, the target evaluation is three months after the stroke, i.e., delaying data collection. Secondly, the optimal dose has to balance the beneficial and adverse effects of the drug, implying that the problem is essentially bivariate in nature. Third, the ongoing dose selection could focus on planning to use the dose that appears to be the best at that time, or the dose that is most informative for the model at hand, and these choices are not necessarily the same. This work will be a combination of theoretical work with the models and evaluation of the performance based on simulations.
Host Institution: H. Lundbeck A/S