Gang Fang's Blog

There is a section devoted to Structural Equation Models (SEMs) in Peter Spirtes’ paper, but he didn’t seem to explain what Structural Equation Modeling actually are.

What is Structural Equation Modeling?

My answer: Unlike Causal Bayesian Networks, which can be defined by their parts (i.e., G, P and the local directed Markov condition), Structural Equation Modeling isn’t a single statistical technique that can be summed up by multiple parts/aspects but a general modeling framework that integrates different multivariate techniques. The main purposes of this general framework are to analyze and test the relationships of observed and latent variables.

However, we can derive insights from the name itself: structural equation models often contain postulated causal connections among some latent variables and causal connections linking latent variables and observed variables. These connections are represented using equations and the causal structures can also be presented using graphs.

Latent variables are variables thought to exist but which can’t be observed directly, like an attitude, intelligence or mental illness.

Common methods in the SEM framework include confirmatory factor analysis, path analysis, multi-group analysis and so on.

Critique: this answer is a synthesis of multiple readings. Can be assumed accurate.