|
Abstract:
|
We present a Bayesian hierarchical model for clustering density estimates, for data on the interval [0,1]. We use Bernstein polynomial priors to model the data and cluster curves based upon the parameters of the Bernstein polynomials. We use our approach to develop radiologist-specific mappings between categorical and continuous measures of breast density, in a setting where some radiologists have few continuous measures. Our approach allows us to borrow information across radiologists to obtain more stable density estimates. We evaluate our approach using simulation, considering cases where the number of clusters is assumed known and cases where the number of clusters is to be determined from the data. Finally, we describe how our approach might be extended to incorporate covariates, allowing us to identify radiologists whose mappings are similar based upon their characteristics.
|
