A2037
Title: Bayesian semiparametric hierarchical copula density estimation for grouped data
Authors: Jichan Park - Korea University (Korea, South) [presenting]
Taeryon Choi - Korea University (Korea, South)
Abstract: Copula models provide a flexible framework for modeling dependence in multivariate data, but copula density estimation for grouped bivariate data remains challenging when dependence structures vary across related groups and exhibit latent heterogeneity within groups. A Bayesian hierarchical copula density estimator for grouped bivariate data is proposed. The model uses a parametric copula baseline and accounts for family uncertainty through reversible jump selection. It further introduces a hierarchical Dirichlet process mixture of latent copula atoms together with a group level smooth nonparametric adjustment. The hierarchical Dirichlet process shares copula atoms across groups while allowing multiple latent dependence components within each group. Residual departures from the parametric baseline are represented by a two dimensional hierarchical spectral expansion with tensor product basis functions and group level smoothing priors. Posterior inference is carried out by Markov chain Monte Carlo with reversible jump updates, blocked Gibbs sampling for the hierarchical Dirichlet process component, and updates for the smooth adjustment. Simulation studies and an empirical application illustrate the proposed framework in terms of copula family selection, dependence estimation, cross group sharing, and within group heterogeneity.
