Abstract:
|
Our limited understanding of human biology creates a hurdle for development of personalized medicines. To overcome this limitation, we need to take better advantage of the available clinical trial data and statistical methods we currently have in hand. Traditional statistical methods may not be suited to help us meet this need. To identify patient characteristics which could predict differences, we borrow ideas from recently developed subgroup identification methods and propose the VG method, which combines (a) the idea of individual treatment effect (ITE) from the Virtual Twins method (Foster et al 2013, Stat Med) and (b) the GUIDE method (Loh 2002, Stat Sinica). Simulation results show that the VG method has less variable selection bias than the Virtual Twins method and it also has higher statistical power than the Gi method (Loh et al 2004, arXiv.org), when there are prognostic variables with strong prognostic effects. The type I error and predictive performance of the Virtual Twins, Gi and VG methods are also compared utilizing data from a randomized clinical trial for Alzheimer's disease.
|
ASA Meetings Department
732 North Washington Street, Alexandria, VA 22314
(703) 684-1221 • meetings@amstat.org
Copyright © American Statistical Association.