At first sight, MIMIC looks very interesting, but if one analyses the model in more detail, one must arrive to the conclusion, that is is a failure. In most cases, there are both indirect (vai the unobserved variable) and direct influences of the causal variables on the indicators, and it is impossible to separate the two. Additionally, the functional relations may be different. If there wre a clear linear relationship between causes and unobserved variable (U=sum(ai*Xi)) and also clear linear functions between the unobserved variable and the indicators (Ij=bi*U), one has a multicollinearity between the indicators. Although you would have very good (OLS)-estimates for any of the indicators, this will likely cause troubles (like nonconvergence) for the combined estimation approach. You could try that by simulations with a simple model with assumed functions for U and I.