A Semi-Parametric Bayesian Generalized Least Squares Estimator

TL;DR

This paper introduces a semi-parametric Bayesian estimator for generalized least squares that adaptively models heterogeneous error distributions using a Dirichlet process prior, improving estimation accuracy in complex data settings.

Contribution

It presents a novel semi-parametric Bayesian approach with a data-driven number of normal components for error distribution modeling in GLS, applicable to systems and panel data.

Findings

Smaller posterior standard deviations compared to parametric methods

Reduced mean squared errors in simulation experiments

Effective application to empirical data models

Abstract

In this paper we propose a semi-parametric Bayesian Generalized Least Squares estimator. In a generic setting where each error is a vector, the parametric Generalized Least Square estimator maintains the assumption that each error vector has the same distributional parameters. In reality, however, errors are likely to be heterogeneous regarding their distributions. To cope with such heterogeneity, a Dirichlet process prior is introduced for the distributional parameters of the errors, leading to the error distribution being a mixture of a variable number of normal distributions. Our method let the number of normal components be data driven. Semi-parametric Bayesian estimators for two specific cases are then presented: the Seemingly Unrelated Regression for equation systems and the Random Effects Model for panel data. We design a series of simulation experiments to explore the performance of our estimators. The results demonstrate that our estimators obtain smaller posterior standard deviations and mean squared errors than the Bayesian estimators using a parametric mixture of normal distributions or a normal distribution. We then apply our semi-parametric Bayesian estimators for equation systems and panel data models to empirical data.