A1230
Title: Dependent Dirichlet processes via thinning
Authors: Laura D Angelo - Universita di Milano Bicocca (Italy)
Bernardo Nipoti - University of Milan Bicocca (Italy) [presenting]
Andrea Ongaro - University of Milano-Bicocca (Italy)
Abstract: An easy-to-implement strategy is proposed to define a flexible class of dependent Dirichlet processes using a thinning technique. Specifically, the well-known stick-breaking construction of the Dirichlet process is modified by introducing random breaks that follow a mixture distribution: a point mass at zero (indicating no break) and a Beta distribution (as in the Dirichlet process). This modification results in the definition of dependent processes that retain analytical tractability, allowing the thorough study of their properties and efficient algorithm development for posterior computation. The approach is illustrated through its application to a two-level clustering problem in the analysis of the Collaborative Perinatal Project data, showcasing its ability to cluster both patients and hospitals simultaneously.
