| Abstract : | This thesis consists of three main essays in Bayesian Econometrics. The first essayis about Bayesian Model Averaging with non conjugate priors with an application togrowth regressions. Bayesian Model Averaging is a novel technique for model selec-tion and model averaging with a lot of virtues in terms of predictiveness. Despite thevastness of the literature, most papers deal with the natural conjugate case for theparameters of the regressors, mainly because one can get analytical expression for theposterior moments and the marginal likelilihood. In this essay I extend the existingliterature taking into account non conjugate priors, by considering two limiting casesof a (non conjugate) multivariate student-t distribution. For the estimation of theposterior probalities, moments, marginal likelihood and predictive density, I proceedto Laplace Method, as it is proposed by Tierney and Kadane (1986) and Tierney, Kassand Kadane (1989) and a multivariate version of Theorem 3, of the latter paper, isderived. The approximations are nicely, incorporated to the MC(3) algorithm. Anapplication to growth regressions shows that the departure from the natural conjugatesetting results to some striking differences regarding the model selection procedure.The predictive results are strong and support the choice of the non conjugate setting.The second essay deals with the issue of objective Bayesian analysis in dynamic panelmodels with arbitrary cross sectional and intertemporal dependence. The objective-ness of the analysis refers to the estimation of Jreys prior of the model parameters.I extend the work of Phillips (1991) in a dynamic panel setting and I show, as Phillipsdid in a time series context, that at priors are not suitable in dynamic models toexpress ignorance. The arbitrariness of the structure of stochastic term deals with thefact that no functional assumptions are made for the error process. I allow for generalcovariance matrices and Ω and Σ , denoting the correlation among the time periods andcross sectional units respectively. I estimate the marginal posterior distribution ofthe autoregressive parameter for several models and I proceed to a comparison withthe flat prior case. My results are similar to those of Phillips (1991). In the the thirdessay, I propose a Bayesian most stringent test for serial correlation. I compare anumber of tests for autocorrelation that exist in the literature with the Bayesian one,in terms of power and stringency for several sample sizes and number of regressors.I show that the minimization of the Bayesian Shortcoming with respect to the priorparameter(s) is equivalent to the maximization of Power with respect to the sameparameter(s) and that under particular parametric settings, the resulting test is theMost Stringent. My findings show that for certain beta priors the Bayesian test at-tains very good overall performance in the alternative hypothesis parameter space.I also estimate the envelope power and under these certain prior distributions theproposed test is the most stringent.
|
|---|
Copyright © 2013 Library AUEB. All rights reserved.