Athens University Economics and Business (AUEB)
Efficient bayesian marginal likelihood estimation for generalized latent trait models
Abstract
dc:descriptionThe term latent variable model (LVM) refers to a broad family of models which are used to capture abstract concepts (unobserved/latent variables or factors) by means of multiple indicators (observed variables or items). The key idea is that all dependencies among p observed variables are attributed to k unobserved ones, where k << p. That is, the LVM methodology is a multivariate analysis technique which aims to reduce the dimensionality, with as little loss of information as possible. Most importantly, the LVMs account for constructs that are not directly measurable, as for instance individuals’ emotions, traits, attitudes and perceptions. In the current thesis, the LVMs are studied within the Bayesian paradigm, where model evaluation is conducted on the basis of posterior model probabilities. A key role in this comparison is played by the models’ marginal likelihood, which is often a high dimensional integral, not available in closed form. The properties of the LVMs are implemented here in order to efficiently approximate the marginal likelihood. In particular, Chapter 1 presents the origins and the basic ideas of the different types of latent variables models. The key aspects of the Bayesian analysis were outlined and the recent literature with respect to the LVMs is reviewed. Chapter 2 focuses on LVMs with binary data and describes the steps required in the Bayesian model assessment (prior specification, sampling from the posterior and Bayesian model comparison). In Chapter3 two different formulations of the LVM marginal likelihood are presented. The variance components associated with each approach are specified and the factors that influence the corresponding errors are outlined. Additionally, the effect of the sample covariation on the estimators is explained and an index of the sample’s divergence from independence is introduced, as a multivariate extension of covariance. Chapter 4 illustrates how the properties of the LVMs can be used to simplify often used estimators in the literature and to reduce the computational time required. Chapter 5 draws a link between Bayesian statistics and ideas initially presented in thermodynamics. It is shown that probability distribution divergences can be estimated via thermodynamic integration, while new thermodynamic marginal likelihood estimators are introduced. In Chapter 6, the methods discussed throughout in this thesis are illustrated in simulated and real life datasets. The thesis closes with a brief discussion on the current findings which prompt future research.
Degree
thesis:*- Grantor dc:publisher
- Athens University Economics and Business (AUEB)
- Year dc:date
- 2013
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Vitoratou, Vasiliki
- Βιτωράτου, Βασιλική
Subjects
dc:subject × 13Rights
dc:rights- Statement dc:rights
-
- BY
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/37291