THE ROLE FOR STRUCTURAL EQUATION MODELING IN THE ANALYSIS OF NETWORKS

Structural equation modeling (SEM) is a scientific framework for using quantitative methods to study causal networks. SEM can be implemented either using classical covariance-likelihood methods or using Bayesian methods, thus it allows for considerable flexibility in specification and estimation. SEM is generally useful for three purposes, (1) modeling building/network discovery, (2) model comparison, or (3) hypothesis testing. In this presentation, I consider the characteristics and requirements of SEM and how it may relate to other modeling traditions that fall within the broad category of “graphical models.” There is particular interest in comparing SEM to what are called “Bayesian Networks” (BNs) or sometimes “probabilistic networks,” the latter being addressed more fully in a companion presentation to this one. The talk will begin with a description of the history of graphical modeling, including both the evolution over time within the SEM tradition, as well as the development of some other kinds of network modeling approaches. The usage of SEM will be illustrated by example and the data requirements needed for SEM will be featured. Emphasis will be placed on the inductive and hypothesis testing capabilities of SEM, which can be contrasted with other network methods that are explicitly designed for deductive usage, such as forecasting conditional probabilities or linking to structured decision processes (strengths of BNs). Ultimately, SEM constitutes one of a complement of approaches to the analysis of networks. It will be argued that the broader enterprise of network analysis is potentially very useful for the practical application of scientific findings.