Abstract:
|
Phase I trials in oncology have advanced from studies of the safety of a single agent to studies of the simultaneous toxicity and efficacy of two agents. Furthermore, the new agents under study are often molecularly targeted agents (MTAs) rather than chemotherapeutic agents. The nature of MTAs brings into question the adage from the era of chemotherapy that "more is better", requiring deeper thought into how we model efficacy as a function of dose changes of either or both of the MTAs. We demonstrate that several published modeling choices implicitly place strong correlation constraints among the dose combinations that lead to over smoothing of the data. As an alternative, we propose the use of a conditional autoregressive (CAR) model, which allows us to model the correlation directly and leads to a direct control on the amount of smoothing that is used. We describe the general structure of CAR models and then present simulation results comparing the operating characteristics of a CAR model-based design with other existing designs. We then conclude with a discussion of several nuances of CAR models that require further study.
|