Nonparametric Bayes local hierarchical models for biomedical data
Statistical Science, Duke University
(February 26, 2009 2:30 PM - 3:30 PM)
In modern biomedical research, it is increasingly common to encounter high-dimensional and complex data, such as gene expression profiles over time, longitudinal trajectories in biomarkers and images. Increasing the complexity is the common interest in combining information across data of different sources. Bayesian hierarchical models provide a useful paradigm for addressing these problems, but parametric assumptions and the curse of dimensionality present difficulties. To address these challenges, this talk presents a general class of local partition mixture models, which facilitate sparse modeling of high-dimensional random effects distributions. These models provide a generalization of commonly-used latent class models, finite mixture models and Dirichlet process mixture models, with some clear advantages in terms of favoring a simultaneous reduction in dimensionality and improvement in fit. The methods are illustrated through applications to modeling of reproductive hormone curves, gene expression data and joint modeling of images and captions.