Bayesian Finite Mixture Models for Regression with Cluster-Specific Variable Selection
Heterogeneous data are ubiquitous in scientific studies. In regression problems, the response in different subpopulations may be influenced by different subsets of covariates. We propose using mixtures, especially mixtures of finite mixtures (MFM), to model the joint distribution of the response and the covariates. In particular, we adopt a parameterization that explicitly involves vectors of regression coefficients within each subpopulation, each assigned spike-and-slab priors to achieve subpopulation-specific variable selection. MCMC algorithms are used for computing, leading to versatile posterior inferences, such as clustering, individual profiling, and predictions.
Dr. Tan is interested in Bayesian modeling, computing and their applications. She works in MCMC, Bayesian variable selection, mixture models and online learning. She has served as an Associate Editor for the Journal of Computational and Graphical Statistics since 2017.