|
Abstract:
|
In studies of disease progression and the effect of interventions, a statistical approach is to specify a model that aims to characterize the main elements of the data generating process. Some challenges in this approach are in specifying a reasonable model that incorporates the context of the scientific setting, performing valid estimation and inference and converting model estimates into quantities of substantive interest. I will describe how we have used joint models of longitudinal prostate-specific-antigen measurements and event-time outcomes in prostate cancer studies to understand the natural history of the disease, assess the effect of interventions, and produce dynamic predictions. In the setting of head and neck cancer, I will describe work developing multistate models of event-time outcomes in which subjects transition between multiple outcome states over time, with each transition requiring its own model specification. A question is whether these context-driven parametric statistical models outperform more agnostic machine learning methods for outcome prognostication. I present results comparing random forests to our context-driven Bayesian multistate models.
|
