Approximating Priors by Mixtures of Natural Conjugate Priors

Abstract

The purpose of this paper is twofold: First, to show that a natural enlargement of the class of natural conjugate priors-namely, mixtures of natural conjugate priors-also leads to mathematically tractable solutions. Secondly, to show that this enlargement is "adequate" in that any prior may be arbitrarily closely approximated by a suitable member of this class.Specifically, a general method for approximating an arbitrary prior density is first given with approximation (convergence) defined pointwise or in the L 1 or total variation senses. Corresponding posterior densities are shown to provide approximations in corresponding senses, whether or not the approximating priors are those given herein. Then formulas are given for posterior densities when the prior is a mixture of natural conjugate priors, and it is shown that an arbitrary prior density and its posterior density may be approximated by such a mixture and its posterior. In these senses, Bayesians may find it satisfactory to confine attention to mixtures of natural conjugate priors. The results are illustrated by several exponential family examples.

A semiparametric Bayesian approach to extreme value estimation

“…Based on theoretical results by De Vore e Lorenz (1993) and Asmussen (1987), Wiper et al (2001) used mixtures of Gamma densities to approximate any density defined over [0, ∞). See also Dalal and Hall (1983) and Dey et al (1995) for related work.…”

A Bayesian Nonparametric Approach to Test Equating

“….}. Her work extended from the results of Dalal and Hall (1983) and Diaconis and Ylvisaker (1985) who proved that for sufficiently large p, mixtures of the form given in (2) can approximate any c.d.f. on [0, 1] to any arbitrary degree of accuracy.…”

Exact Bayesian modeling for bivariate Poisson data and extensions

“…Appropriate choice of the mixing probabilities and the gamma density parameters can provide arbitrarily close approximation (Dalal and Hall 1983) to a wide range of possible scenarios regarding the dependence among θ 's. In general the literature regarding multivariate mixtures of gamma densities is sparse (we are aware of the work of Gaver 1970 in the general setting).…”

“…Dependent prior distributions can be quite flexible since they offer the opportunity to represent special patterns. We define and base our inference on a prior which is a finite mixture of conditionally independent gamma densities and extend the idea of mixtures of conjugate priors (Dalal and Hall 1983). Both the joint and marginal posterior distributions can be obtained as a finite mixture of conditionally independent gamma distributions, i.e.…”