Structural Equations with Latent Variables: Computational Solutions for Modern Data Problems

Abstract

Structural Equation Modeling (SEM) is a flexible and popular method for data analysis in the social and behavioral sciences. SEM is particularly suitable for research situations in which concepts cannot be measured directly, or where the instruments used for measurement are error-prone. Examples are "trust" or "well-being" -- concepts which are indirectly measured with questionnaires. But the modern data landscape is changing, and SEM is reaching its limits: Classic survey and experimental research is being supplemented (and sometimes even supplanted) by research using measurements from register data, wearable sensors, images, internet databases, genetic sequencing, advanced brain imaging techniques, and more. The SEM method is not always available for these, but the problems of fallible measurement do not disappear in this modern data landscape, and many research questions using this data still involve causal relations between latent constructs. Analyses using SEM are therefore of great value for research with such new measuring instruments. Thus, the goal of this dissertation is to make SEM analyses available to a wider range of these modern datasets. To this end, I develop several solutions to problems encountered in the application of SEM to such data.