Athens University Economics and Business (AUEB)
Using and expanding hybrid latent variable models for modelling missing data and for discovering influential points
Abstract
dc:descriptionThe term Hybrid is used in general to describe the mixed character of something having two kinds of components. Hybrid latent variable models allow the specification of two or more classes and withinclass structures that can range from local independence to complex structural relations between latent continuous variables. Chapter 1 discusses the literature in the area of Hybrid latent variable models, with details concerning the estimation procedure based on EM of a Hybrid 2LC/2PL-Hom model for binary data. By the notation 2LC/2PL-Hom we denote a model in which the response to any item is governed by a discrete latent variable which subdivides the population into two latent groups (2LC) and additionally that a response to any item in a particular one of these groups is governed by another continuous latent variable. The response probability for this group of subjects is described by the two-parameter item response (2PL) model. The other latent class is assumed homogeneous (Hom) without any latent structure. Model _t is the object of discussion and investigation of the next two chapters. Overall and Limited-Information tests are presented and their performance is explored through simulated examples of 2LC/2PL-Hom models in order to assess them in terms of power and Type I error rates. Emphasis is placed upon sparse data which is usually the case in areas of applications of latent variable models. The percentage of expected frequencies less than or equal to 5 is taken as a measure of sparseness. In Chapter 4, use of the Hybrid 2LC/2PL-2PL model is proposed to deal with Missing Not At Random (MNAR) binary data in social surveys. This model assumes that the missing (unobserved) data depend on something other than the observed data and, under necessary assumptions and constraints, permits the missing data to be incorporated into the analysis. In particular, under the proposed model, the response propensity of individuals depends on the specific latent groups into which the population is separated by the discrete latent variable which is measured by the set of binary items in an attitude survey. Influence analysis is studied in Chapter 5 for a 2PL IRT model for binary data for the detection of influential points with the main objective of alerting us when the results of fitting such a model are very misleading and also enabling us to identify the source of this influence.
Degree
thesis:*- Grantor dc:publisher
- Athens University Economics and Business (AUEB)
- Year dc:date
- 2014
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Tsompanaki, Evgenia
- Τσομπανάκη, Ευγενία
Subjects
dc:subject × 10Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/38744