Academic

Abstract

Empirical likelihood-based inference incorporates the advantages of both parametric and nonparametric modeling by allowing inference on parameters of the data distribution, while avoiding the need for strong distributional assumptions that characterizes parametric inference. The Bayesian approach to empirical likelihood complements these advantages by offering the flexibility for fitting complex data and allowing for a wide range of priors to be incorporated to achieve modeling objectives. In this talk, we will cover the key computational challenges in Bayesian empirical likelihood (BayesEL) inference that inhibit its widespread usage. We will examine how the new variational inference approach to BayesEL computations addresses these challenges and yields superior computation-accuracy time trade-off over existing methods. Asymptotic properties of the proposed variational BayesEL posterior will also be discussed.