Bayesian Nonparametric Estimation for Dynamic Treatment Regimes with Sequential Transition Times
Dynamic treatment regimes in oncology and other disease areas often can be characterized by an alternating sequence of treatments or other actions and transition times between disease states. The sequence of transition states may vary substantially from patient to patient, depending on how the regime plays out, and in practice there often are many possible counterfactual outcome sequences. For evaluating the regimes, the mean final overall time may be expressed as a weighted average of the means of all possible sums of successive transitions times. A common example arises in cancer therapies where the transition times between various sequences of treatments, disease remission, disease progression, and death characterize overall survival time. For the general setting, we propose estimating mean overall outcome time by assuming a Bayesian nonparametric regression model for the logarithm of each transition time. A dependent Dirichlet process prior with Gaussian process base measure (DDP-GP) is assumed, and a joint posterior is obtained by Markov chain Monte Carlo (MCMC) sampling. We provide general guidelines for constructing a prior using empirical Bayes methods. We compare the proposed approach with inverse probability of treatment weighting. These comparisons are done by simulation studies of both single-stage and multi-stage regimes, with treatment assignment depending on baseline covariates. The method is applied to analyze a dataset arising from a clinical trial involving multi-stage chemotherapy regimes for acute leukemia. An R program for implementing the DDP-GP-based Bayesian nonparametric analysis is freely available at https://www.ma.utexas.edu/users/yxu/.