Statistical Methods for Dynamic Treatment Regimes with Survival Outcomes and Dose-Finding Designs

Improving patient outcomes in clinical research requires statistical methods tailored to distinct goals of therapeutic development. Early-phase trials must balance safety and efficacy when evaluating treatments, while chronic disease management requires strategies to guide individualized treatment decisions over time. This proposal highlights statistical methods in development to address these complementary challenges in modern biomedical research. The first part introduces a method for estimating optimal long-term treatment strategies for time-to-event outcomes using generalized survival random forests with a model re-fitting approach, accommodating varying patient histories while maximizing expected survival time. This is supported by theoretical results, simulation studies, and real-world application to pediatric Crohn’s disease data. The second part develops an improved methodology for Phase 1/2 trials by advancing utility-based dose selection to an alterative involving a monotone function of safety and efficacy. Simultaneously, a delayed-outcomes framework was also proposed to consider trial speed. This is supported by extensive simulation studies using a variety of random scenarios. The third part extends this work to combination therapy trials, and includes settings with ordinal treatment responses. Collectively, these methods advance the design and evaluation of clinical strategies across the health domain. By improving dose selection in early-phase trials and enabling robust estimation and validation of dynamic treatment regimes for chronic disease management, this work contributes practical tools to support safer, more effective, and personalized patient care.