Structural Equation Modeling: Where Are Advantages? | /en/2008/01/structural-equation-modeling-where-are-advantages/

Guest *Dan* @ 2008-08-10 23:23:13 originally posted:

Yihui, you are exactly right. SEM is a useless statistical technique for the most part, and it has been abused since its outset. Many social sciences would do much better without statistics at all.

Guest *Alden* @ 2012-10-29 15:27:02 originally posted:

You have interesting perspectives. Some of the statements about SEM can be said about regression and other forms of statistics. Other statements are based on incomplete information - it's like if I were to tell you I do not like baseball because I do not like seeing red socks. So in reality, it's not baseball I don't like, it's the Red Sox.

Disadvantage 1: SEM considers only covariance structures
Comment: False. Some seminal kinds of SEM, like continuous-variable factor analysis, can be done off a covariance matrix. But you meeds means and categorical item thresholds in other types of SEM models (e.g., item response theory). In fact, you need the mean structure to do any sort of longitudinal growth modeling.

Disadvantage 2: Complexity of models.
Comment: This is more a generic excuse to not do anything rather than a real criticism of SEM. That something is complex does not mean we should not use it. I must acknowledge that it's too easy to complicate SEM models unnecessarily. One must be very careful: it's not as if SEM methods solve everything. They do not. They are a tool, like a wrench in a toolbox. They are not appropriate in all cases, just as regression is unnecessarily complex for some situations like sparse data.

Disadvantage 3: SEM is a process of hypothesis testing.
Comment: Incomplete. It can also be used for hypothesis generation. This comment is generic to any kind of statistics. Hypotheses can be rejected based on poor fit to the data. Data that fits well could be consistent with an alternative hypothesis, but never definitively proves it. It's our way of operationalizing the world we live in. Just like regression. In fact, SEM is a hyp'ed up regression. This reminds me of George Box's "All models are wrong. Some are useful." They key is not to find an SEM model that explains everything perfectly, but to structure your relationships in ways that conform to prevailing world-views. Psychometricians have concepts like construct, content, and criterion validity to examine how well our structures represent real-world processes.