Prior Sensitivity Analysis without Model Re-fit

TL;DR

This paper introduces a new method for prior sensitivity analysis in Bayesian inference that avoids re-fitting models by computing distances between posteriors using Monte Carlo integration, saving computational resources.

Contribution

The authors develop a novel approach to quantify prior sensitivity without re-fitting models, applicable to complex latent variable models, using integral expressions of posterior distances.

Findings

Efficient computation of posterior distances using Monte Carlo integration.

Applicable to various models including hierarchical and Gaussian process models.

Demonstrated effectiveness through multiple example models.

Abstract

Prior sensitivity analysis is a fundamental method to check the effects of prior distributions on the posterior distribution in Bayesian inference. Exploring the posteriors under several alternative priors can be computationally intensive, particularly for complex latent variable models. To address this issue, we propose a novel method for quantifying the prior sensitivity that does not require model re-fit. Specifically, we present a method to compute the Hellinger and Kullback-Leibler distances between two posterior distributions with base and alternative priors, using Monte Carlo integration based only on the base posterior distribution, through novel integral expressions of the two distances. We also extend the above approach for assessing the influence of hyperpriors in general latent variable models. We demonstrate the proposed method through examples of a simple normal distribution model, hierarchical binomial-beta model, and Gaussian process regression model.