I want to estimate a longitudinal SEM model that includes two latent variables (LVs: M, R) over two time points. The indicators for these LVs are categorical (4 resp. 5 categories). At the second time point the two LVs interact (MxR) to explain the examined outcome (D <- M + R + MxR) in my analysis.
Because the constructs are measured over time the estimation of residual correlations for each indicators of the 2 LVs is needed. Browsing the Mplus discussion WLSMV is preferred in this case over ML with numerical integration (MLRnum). Otherwise MLRnum has a nice ability to estimate LV interactions by the XWITH-command. It seems to be the preferred method when modeling latent interactions.
My question is now, which of both estimators (WLSMV or MLRnum) has to be selected in a longitudinal SEM with categorical indicators including a latent interaction.