2004
Uppsala, Sweden
An adaptive design for dose-response using the Normal Dynamic Linear Model.
Michael K. Smith(1), Mark F. Morris(1), Ieuan Jones(1), Andy P. Grieve(1), Keith Tan(2)
(1) Biostatistics and Reporting, PGRD Sandwich; (2) Clinical Sciences, PGRD Sandwich.
Objectives: Adaptive designs have an intuitive appeal in the drug development environment where we are trying to balance allocation of patients to efficacious doses with maximising our learning about the dose-response relationship while doing so in as cost-efficient a manner as possible. We suggest a pragmatic approach to running an adaptive dose-response trial on a VAS numerical rating scale endpoint, dropping ineffective doses and terminating the study early if there is no clear clinical benefit.
Methods: A Normal Dynamic Linear Model has been used to describe the dose-response relationship. This is a very versatile and flexible model which allows for non-monotonic response functions. This could be an important consideration in the chosen endpoint if subjects drop out of the study early. The NDLM also provides, within the bayesian context, access to probabilistic statements about features of the dose-response relationship. These probabilistic statements form the basis of decision criteria for dropping doses and terminating the study outright if there is a low probability of showing sufficient efficacy. Simulations were performed to investigate the performance of these rules when the simulated data were analysed at different interim analyses. The timing and frequency of these interims could be optimised using the simulation results.
Results and conclusions: The results of the simulated trials show that it is possible to run a cost-effective adaptive trial based on a parallel group study with equal allocation to treatments. A simple study design makes the execution of the study much simpler and we get the benefits of an adaptive design (dropping ineffective doses) without the complexity of some other adaptive designs in the setup phase. The interim analyses show good power and low Type I error with significantly lower average sample sizes than conventional studies run to completion. The inferences from the NDLM correctly pick out doses with low probability of success from those that are efficaceous. The estimated cost savings due to stopping the study early and dropping individual doses due to lack of efficacy could be as much as $486,000.