Loading...
Please wait, while we are loading the content...
Generalized Linear Latent Variable Models for time dependent data
| Content Provider | Semantic Scholar |
|---|---|
| Copyright Year | 2013 |
| Abstract | Latent variable models are a fundamental tool for the analysis of multivariate data. The importance of such models is due to the crucial role that latent variables play in many fields, e.g. psychological and educational, socioeconomic, biometric, where often constructs are not directly observable. In these contexts, the different nature of the observable variables often causes theoretical and practical problems. In this respect, one can refer to the general theoretical framework of Generalized Linear Latent Variable Models (GLLVM) that represents an extension of the generalized linear models when covariates are not observable. This approach allows to model mixed data type at the same time and includes a variety of methods like the classical factor analysis if the observed variables are continuous and the latent trait models in the case of categorical data. Recently, the interest devoted to these kind of models has increased noticeably but there are still unsolved problems from both the methodological and the practical point of view. The aim of this research project is to study the problems of parameter estimation and goodness of fit in the framework of GLLVM for time dependent data. The analysis of longitudinal data exploits models that include latent variables and random effects, where the random effects that are incorporated into the model reflect the unobserved heterogeneity of individual behaviours. Such models can be interpreted also as dynamic factor models because they allow exactly to analyze the dynamics of latent variables over time through autoregressive processes. Remarkably, the literature of categorical time series usually does not deal with latent variables, especially in the multivariate framework. The project aims at filling this gap by extending and implementing generalized linear latent variables models in the context of dynamic processes by considering various specifications. In particular, the focus is on estimation problems: those related to the computational burden of the maximization of the log-likelihood and those related to the analytical intractability of integrals for which a numerical approximation is required. A further aspect that will be considered in the project is the goodness of fit problem of these models. Indeed, in presence of categorical manifest variables with many categories, the sparseness problem invalidates the classical goodness of fit tests. The aim of the project is to propose new solutions to the case of sparse data. |
| File Format | PDF HTM / HTML |
| Alternate Webpage(s) | https://www.aricweb.unibo.it/AssegniRicerca_Richieste/_doc/2013/ID%5B11498%5Dprogetto.pdf |
| Language | English |
| Access Restriction | Open |
| Content Type | Text |
| Resource Type | Article |
