Bayesian Analysis Large Sample Methods推荐.pdfVIP

Large Sampl e Method s In order to make Bayesian inference about th e parameter 0^ given a model f{x\6)^ one needs to choose an appropriat e prior distribution for 6. Given th e data x , th e prior distribution is used t o find th e posterior distribution and var- ious posterior summary measures , depending on th e problem . Thu s exact or approximate computation of th e posterior is a major problem for a Bayesian . Under certain regularity conditions , th e posterior can be approximated by a normal distribution with th e maximum likelihood estimate (MLE) as th e mean and inverse of th e observed Fisher information matrix as th e dispersion matrix , if th e sample size is large . If more accuracy is needed, one may use th e Kass-Kadane-Tierney or Edgeworth typ e refinements . Alternatively, one may sample from th e approximate posterior and take resort t o importance sam- pling . Posterior normality has an important philosophical implication , which we discuss below. How th e posterior inference is influenced by a particular prior depend s on the relative magnitude of th e amount of information in th e data , which for i.i.d. observations may be measured by th e sample size n or nl{6) or observed Fisher information 7^ (defined in Section 4.1.2), and th e amount of informa- tion in th e prior, which is discussed in Chapter 5. As th e sample size grows, th e influence of th e prior distribution diminishes . Thu s for large samples , a precise mathematical specification of prior distribution is not necessary . In most cases of