Inverse stable prior for exponential models

TL;DR

This paper introduces a new class of inverse stable priors for exponential models that are proper, flexible, and allow closed-form posteriors, improving Bayesian inference for various data distributions.

Contribution

The paper proposes a novel inverse stable prior family with closed-form posteriors, flexible distributional shapes, and practical algorithms for Bayesian analysis in exponential models.

Findings

Inverse stable priors outperform inverted beta in some shrinkage scenarios.

Closed-form expressions for moments and MGF facilitate analysis.

Effective algorithms enable practical Bayesian inference.

Abstract

We consider a class of non-conjugate priors as a mixing family of distributions for a parameter (e.g., Poisson or gamma rate, inverse scale or precision of an inverse-gamma, inverse variance of a normal distribution) of an exponential subclass of discrete and continuous data distributions. The prior class is proper, nonzero at the origin (unlike the gamma and inverted beta priors with shape parameter less than one and Jeffreys prior for a Poisson rate), and is easy to generate random numbers from. The prior class also provides flexibility in capturing a wide array of prior beliefs (right-skewed and left-skewed) as modulated by a bounded parameter $\alpha \in (0, 1).$ The resulting posterior family in the single-parameter case can be expressed in closed-form and is proper, making calibration unnecessary. The mixing induced by the inverse stable family results to a marginal prior distribution in the form of a generalized Mittag-Leffler function, which covers a broad array of distributional shapes. We derive closed-form expressions of some properties like the moment generating function and moments. We propose algorithms to generate samples from the posterior distribution and calculate the Bayes estimators for real data analysis. We formulate the predictive prior and posterior distributions. We test the proposed Bayes estimators using Monte Carlo simulations. The extension to hierarchical modeling and inverse variance components models is straightforward. We can find $\alpha$ (which acts like a smoothing parameter) values for which the inverse stable can provide better shrinkage than the inverted beta prior in many cases. We illustrate the methodology using a real data set, introduce a hyperprior density for the hyperparameters, and extend the model to a heavy-tailed distribution.