Generalised latent variable models for location, scale, and shape parameters

Latent Variable Models (LVM) are widely used in social, behavioural, and educational sciences to uncover underlying associations in multivariate data using a smaller number of latent variables. However, the classical LVM framework has certain assumptions that can be restrictive in empirical applications. In particular, the distribution of the observed variables being from the exponential family and the latent variables influencing only the conditional mean of the observed variables. This thesis addresses these limitations and contributes to the current literature in two ways. First, we propose a novel class of models called Generalised Latent Variable Models for Location, Scale, and Shape parameters (GLVM-LSS). These models use linear functions of latent factors to model location, scale, and shape parameters of the items’ conditional distributions. By doing so, we model higher order moments such as variance, skewness, and kurtosis in terms of the latent variables, providing a more flexible framework compared to classical factor models. The model parameters are estimated using maximum likelihood estimation. Second, we address the challenge of interpreting the GLVM-LSS, which can be complex due to its increased number of parameters. We propose a penalised maximum likelihood estimation approach with automatic selection of tuning parameters. This extends previous work on penalised estimation in the LVM literature to cases without closed-form solutions. Our findings suggest that modelling the entire distribution of items, not just the conditional mean, leads to improved model fit and deeper insights into how the items reflect the latent constructs they are intended to measure. To assess the performance of the proposed methods, we conduct extensive simulation studies and apply it to real-world data from educational testing and public opinion research. The results highlight the efficacy of the GLVM-LSS framework in capturing complex relationships between observed variables and latent factors, providing valuable insights for researchers in various fields