Bayesian Nonparametric Estimation for Dynamic Treatment Regimes with Sequential Transition Times

• Activity Number: 591
• Type: Invited
• Date/Time: Wednesday, August 3, 2016 : 2:00 PM to 3:50 PM
• Sponsor: JASA, Applications and Case Studies
• Author(s): Peter F. Thall* and Yanxun Xu and Peter Mueller and Abdus S. Wahed
• Companies: MD Anderson Cancer Center and and The University of Texas at Austin and University of Pittsburgh
• Keywords: Dependent Dirichlet Process ; Gaussian Process ; G-Computation ; Inverse probability of treatment weighting; ; Markov chain Monte Carlo

Abstract

We analyze a dataset from a clinical trial involving multi-stage chemotherapy regimes for acute leukemia. The trial design was a 2x2 factorial for frontline therapies only. Since subsequent salvage treatments may affect survival time, we model therapy as a dynamic treatment regime, an alternating sequence of adaptive treatments and transition times between disease states. These sequences may vary between patients, depending on how the regime plays out. Overall survival time is a weighted average of all possible sums of successive transitions times. We assume a Bayesian nonparametric survival regression model for each transition time, with a dependent Dirichlet process prior and Gaussian process base measure (DDP-GP). We provide general guidelines for constructing a prior using empirical Bayes methods. To estimate mean overall survival time for each regime, the DDP-GP approach is compared by simulations with inverse probability of treatment weighting (IPTW) and doubly robust IPTW, for single-stage and multi-stage regimes with covariate-dependent treatment assignments. The simulations show that the DDP-GP approach can substantially improve inference compared to these IPTW methods.