it is high (80%)

Can you elaborate on your question, Nathan?

Agreement with what?

I mean, your question could be interpreted in several different ways. For example:

• SEM is a useful tool?
• Covariance-based SEM is preferable to Partial Least Squares?
• It is okay to create models that use both Formative indicators and Reflective indicators?
• We deal with Heywood cases the same way?

Well if you get a piece of research right it is good and if wrong it is bad. Good methodology depends on the research question, mathematics and many other things. As someone once said, We deal with nature as we find it not as we wish it was (though I admit some models wish VERY hard). However for those that think SEM is the answer to every question, I suggest think harder. David

Yes, my question was purposefully broad and vague, and not meant to pertain to a specific aspect of SEM. I'm more interested in seeing the discussion, and your overall impressions. To elaborate on my question further (@Hume F. Winzar), with something like SEM there seems to be only a moderate level of agreement among scholars on various aspects of its use. For example, some vary in their opinions on what represents "good" fit for different fit statistics, what constitutes "good fitting" models overall, and even what specific fit indices are more or less relevant than others depending on the type of analyses. As other examples: Some scholars disagree on what constitutes "adequate" sample size for different types of SEM, the appropriateness of some models (e.g., latent change score models), or the extent to which modification indices should be considered. Different analysis software packages can handle certain procedures differently.

To be clear, these are examples among many, and I am not looking for an answer to these or to any other specific aspect of SEM. I just interested in your impressions of overall agreement in an analysis domain that appears to be prone to different perspectives on what constitutes best practice.

My take: After having used SEM in many studies, reading many perspectives, and working with different methodologists, I suggest that the level of overall agreement among them is not much better than "moderate". Thoughts?

If you are talking about fit in SEM models, you should have the same scrutiny with GLM models. In Regression models explained variance is looked at, but that can be done in SEM Models as well and for comparable models, SEM explained variances are usually higher.

GLM models just stick to the variance of the explained variance and not to explaining the complete variance-covariance matrix as SEM models have to do.

SEM has more criteria it can fulfill, for most of these we can only guess what they should be in GLM models.

In SEM models one problem can be modeled quite easily: multi-collinearity. This can cause problems in GLM models. If you would use stepwise procedure in GLM models a backward and forward method give the same results when variable are orthogonal. (Though other problems might cause the problem as well).

Modelling error in SEM models is done easily, but can not or are hard to model in GLM models.

Usually I discard a lot of articles, because they promise me to provide values for certain parameters, but from the theoretical model and the statistical analysis it must follow, they can not do so. Sometimes one can not be able what has been given.

Most problems you mention do have a counterpart for GLM models, but it has not been discussed as detailed and hard as with SEM models.

In a way it now returns in Multi-level SEM models, but quite a lot of researcher are at a loss when asked to use such models. They are harder to run and one has to take more parameters into account.

At the moment I guess 40% of the statistical models we would require are available. And from the rest, a lot (30%) will not have solutions.