MIIV (http://fgme2017.de/kenneth-bollen/) looks like a very promising approach. Are there any experiences from research practice? Application-oriented publications? When is it appropriate to use this approach?
I'm not familiar with any application oriented application. I'm sorry. Clearly, one needs to be written. This approach is always appropriate to use. Essentially, the approach is more robust to model mispecification than standard estimation procedures in SEM. You may a) find similar results than say using ML, or b) substantively different results. If a), it gives you confidence that your ML results are the 'correct' ones. If b), the IV approach is giving you the correct answer.