Institutes outside Greece
Advances in the bayesian analysis of statistical models with intractable normalising constants
Abstract
dc:descriptionBayesian statistics have had a profound impact on statistical learning, offering an approach to applying probability to statistical problems. The Bayesian inferential approach allows the researcher to update their beliefs about random events in light of observing new data about those events and to produce new posterior beliefs. Typically, the posterior probability distributions that are associated with the updated beliefs are intractable analytically and this has hindered inference of the posterior distribution. The discovery of Markov chain Monte Carlo (MCMC) algorithms (Metropolis et al., 1953; Hastings, 1970) provided a simulation-based approach to overcoming the intractability inherent in almost all posterior distributions. Despite being a general tool that has been used to routinely simulate observations from the posterior distribution, the implementation of MCMC faces computational difficulties in modern statistical problems: many popular statistical models for complex phenomena are intractable, as the likelihood function cannot easily be evaluated. This yields posterior distributions which are termed doubly-intractable. This thesis sought to develop novel statistical methods in this context, focusing on the problems of parameter inference and model selection. In the first part of the thesis we introduce a noisy MCMC algorithm which is no longer invariant for the target posterior distribution, for Bayesian model comparison and multimodel inference of doubly-intractable distributions. We provide convergence guarantees for the algorithm and illustrate the merits of this method with examples from social network analysis. In the second part, this thesis concentrates on the tractable pseudolikelihood approximation to the true likelihood, for social network analysis. Despite their tractability, pseudolikelihoods should be treated with caution because they can lead to an unreasonable inference. Here, our contributions are two-fold: (i) we present a viable approach to adjust the posterior distribution resulting from using the pseudolikelihood in an efficient manner, while providing the theoretical framework for this approach; (ii) we specify a method to adjust the pseudolikelihood itself in order to obtain a reasonable approximation to the true likelihood, which allows for implementation of widely used computational methods for statistical evidence estimation and pursuit of Bayesian model selection of doubly-intractable distributions.
Degree
thesis:*- Grantor dc:publisher
- Institutes outside Greece
- Year dc:date
- 2018
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Bouranis, Lampros
- Μπουράνης, Λάμπρος
Subjects
dc:subject × 8Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/57384