B1955
Title: Bayesian analysis and inference for semiparametric generalized linear models with discrete or continuous data
Authors: Paul Rathouz - University of Texas at Austin (United States) [presenting]
Entejar Alam - U of Texas at Austin (United States)
Peter Mueller - UT Austin (United States)
Abstract: A class of dependent nonparametric Bayesian priors is introduced, building on previous semiparametric generalized linear models. Assuming a Dirichlet process prior on the centering function (reference distribution) of an exponential family, we characterize the resulting exponential family of probability models as inhomogeneous normalized completely random measures. We discuss the corresponding Levy intensity and introduce posterior simulation algorithms. The latter is implemented as a variation of MCMC algorithms for normalized random measures. We discuss the special case of ordinal outcomes. Finally, we connect with a fast-expanding literature on non-parametric Bayesian models for dependent random measures, including especially the widely used dependent Dirichlet process models. The proposed model can be characterized as an inhomogenous dependent Dirichlet process model with varying weights and atoms.