We estimated a model in which we modeled the intercept and slope for a construct and estimated the relationship between the I and S using a WITH statement. In that model the mean estimate for the slope was -.03 and was significant. In our next model we predicted the slope from the intercept using an ON statement and the slope mean became .68 and was still significant but the direction had changed. (The relation between slope and intercept was -.17 and was significant.) The changes in slope -- from negative to positive and from .03 to .68 -- seemed drastic. When I predict the slope from the intercept (or other time invariant covariates) should I no longer pay attention to the latent factor mean or is it changing that much? Note - model fit was good in each case.