Abstract:
|
We consider optimal design of sampling schedules for binary sequence data. The motivating example is a clinical trial, where the measured responses are repeated measurements on a binary outcome. The decision is related to the trade-off between obtaining more information by more frequent sampling versus the incurred sampling cost. We propose an approach that allows to incorporate a variety of goals in the utility function. We include deterministic sampling cost, a term related to prediction, and one term related to inference in the underlying probability model. To avoid dependence on a specific parametric form, we use a non-parametric probability model, relying on minimal assumptions only. Assuming partial exchangeability for the binary sequence as in Quintana and Newton (1998), we use an implementation from Quintana and Mueller (2001), which assumes a DP prior for the mixture. Appropriate computational schemes are discussed and applied to a study on bladder cancer.
|