Marginal likelihood estimation via power posteriors

摘要: Model choice plays an increasingly important role in statistics. From a Bayesian perspective crucial goal is to compute the marginal likelihood of data for given model. However, this typically difficult task since it amounts integrating over all model parameters. The aim paper illustrate how may be achieved by using ideas from thermodynamic integration or path sampling. We show can computed via Markov chain Monte Carlo methods on modified posterior distributions each This then allows Bayes factors probabilities calculated. that approach requires very little tuning and straightforward implement. new method illustrated variety challenging statistical settings.