Sampling-Based Approaches to Bayesian Inference in Econometrics

本文研究的重点是计量经济学中的贝叶斯多重积分问题的解，使用一类被广泛称为Gibbs采样器的数值积分程序。该项目的目标是使贝叶斯推理与计量经济学中的经典方法一样常规和可靠，并使用贝叶斯或经典方法为迄今进展甚微的突出计量经济学问题提供解决方案。该项目涉及Gibbs采样器分析的扩展，以开发适用于系统地评估数值精度的中心极限定理，评估对初始条件的敏感性，并将这些方法应用于未被识别的模型。该项目还将利用吉布斯抽样的连续条件原理和相关的数据增强技术，开发大多数计量经济学模型的模块化处理。这些模块包括回归中的半信息先验、删失回归、异方差、看似不相关的回归、多项概率模型、降阶回归、潜在因素模型、多变量(移动平均)自回归和状态空间模型。

The focus of this research is on solutions of the Bayesian multiple integration problem in econometrics, using a class of numerical integration procedures widely known as Gibbs samplers. The objectives of the project are to make Bayesian inference as routine and reliable as classical methods in econometrics, and to provide solutions of outstanding econometric problems on which little progress has been made heretofore using either Bayesian or classical approaches. The project involves extensions of the analysis of Gibbs samplers to develop central limit theorems appropriate to the systematic assessment of numerical accuracy, evaluate senstivity to initial conditions, and apply these methods ot underidentified models. The project will also exploit the successive conditioning fundamental to Gibbs sampling and the related technique of data augmentation, to develop a modular treatment of most econometric models. These modules include semi-informative priors in regressions, censored regression, heteroskedasticity, seemingly unrelated regressions, the multinominal probit model, reduced rank regression, the latent factor model, multivariatre (moving average) autoregression, and state space models.