New Statistical Methods for Latent Variable Models
This research project aims to develop new efficient statistical methods to better model linear and nonlinear relationships and accurately predict latent variables within LVMs. The hierarchical-likelihood, a generalization of Fisher's likelihood, will serve as a fundamental tool in this project. Within the h-likelihood framework, we propose estimation methods and inference procedures for various LVMs. To our best knowledge, h-likelihood has never been applied to the models in this project. H-likelihood will be thoroughly studied under different conditions. Further, it will be applied to the nonlinear model and ordinal data for the first time.
Theories in social and behavior sciences often incorporate theoretical concepts which cannot be measured directly and are often referred to as latent variables. The Latent variable model (LVM) is a broad family of models that are used to capture relationships of latent variables by means of multiple indicators.
LVMs are best known as factor analysis and structural equation models. The relationships between latent variables can be linear or non-linear; the indicator variables can be continuous and/or categorical. It is important to distinguish linear and nonlinear latent relations and identify the nature of indicators in establishing more meaningful and correct models for complicated situations. To achieve these goals well-fitting statistical models, efficient estimation procedures, and accurate prediction methods are highly needed.
Project period
2018-01-01 – 2021-12-31
Funding
The project is financed by the Swedish research council (VR)
Project members
Project members at the Department of Statistics:
- Professor Fan Yang-Wallentin, project leader
- Associate professor Shaobo Jin
- PhD candidate Johan Vegelius