Identifiability and Statistical Inference in Latent Variable Modeling
MCML Authors
Nils Sturma
Dr.
* Former Member
Abstract
Nils Sturma
Dr.
* Former Member
Abstract
In this publication-based dissertation, we study latent variable models in parametric settings. Since latent variable models are families of marginal distributions, they generally feature a complicated geometry that may lead to identifiability issues and failures of standard inference methods. For example, the models often contain irregular points like algebraic singularities, where well-known methods such as the likelihood ratio test or Wald test are no longer valid. One contribution of this thesis is to develop a testing methodology that is valid even if the underlying model contains irregular points. The other focus of this thesis is the investigation of geometry and identifiability in certain types of linear structural equation models with latent variables.
BibTeXKey: Stu24