I'm studying about artificial neural networks (ANN) for the first time and I am struck by how the concepts of neural networks appear to be similar to structural equation modeling (SEM). For example,

  • input nodes in ANN remind me of manifest variables in SEM
  • Hidden nodes in ANN remind of latent variables in SEM
  • Every feature in ANN gets an input node as every observed variable gets a manifest variable in SEM
  • ANN can have several output nodes just as SEM can have several final dependent variables
  • Both can be use explanatory and predictive purposes (I think)

So please explain to me the differences between these two forms of statistical analysis


Short answer: With SEM, the goal is generally to understand the relationships between the variables. With the type of ANNs you have been studying, the nodes are a way of transforming the data so that the predictor variables can better explain the outcomes. Ultimately the similarity is pretty superficial: while the diagrams look similar, you will struggle to get good predictions from an SEM and you will also struggle to interpret the relatships between variables in an ANN.

Pedantic answer: there are lots of different types of SEMs and ANNs. Many do not look so similar. E.g., a kohonen network looks little like an SEM, and is not great for prediction. When SEM is used to address endogeneity, it can be good for prediction, but such SEMs usually don't get drawn as pretty network diagrams.

  • $\begingroup$ While ANN's have their own criticisms, like poor explainability, I'm equally concerned by a SEM model that is supposed to explain something but doesn't make good predictions. $\endgroup$ – Galen Apr 29 '20 at 17:15

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.