I am working on a number of CFAs for one measure, determining which factor structure proposed in the literature is most appropriate. The measure contains 8 reliable subscales. One proposed model consists of 3 factors: 1 latent variable comprised of 4 subscales, 1 latent variable comprised of 3 subscales, and then the 8th subscale serving as a 3rd factor. I have been advised that one way to run this model in MPLUS is to make the 3rd factor a latent variable with one loading (the single subscale comprising the 3rd factor) and then to fix that loading at the reliabilty coefficient for that subscale. Is this the proper way to proceed? If so, can anyone recommend references for this method? If it is not the proper way, I would appreciate advice on how to proceed.
With a single indicator, you can correct for reliability or simply use the observed variable in the analysis. If you correct for reliability, you should be sure the reliability estimate is correct. See slide 44 of the Topic 1 course handout to see how to do this in Mplus. I would imagine this topic is covered in most SEM textbooks like the Bollen book.