Tristan posted on Saturday, March 03, 2012 - 6:13 am
I am new to SEM and MPLUS, and attempting to model a latent factor that influences all indicators in the structural model as per the recommendations of Podsakoff et al (2003). I am getting some weird results when I compare the path coefficients in my structural models (no CMV factor versus CMV factor present). The coefficients (and overall R2 for endogenous variable) differs significantly between CMV and no CMV models, so I want to check if I am modelling CMV correctly. My code is below:
MODEL: factor definition and path modelling omitted...
I'm working on the same common method test described above. And while my models are working, I was wondering how exactly I could use the output of such a CFA model to partition the variance, i.e., estimate the percentage of variance in responses due to trait, method, and random error components (Podsakoff et al., 2003; Williams, Cote, & Buckley, 1989)?
My understanding is that constraining all variables to load equally on the CMV factor allows one to compute the proportion of variance explained by the CMV factor. This is why I have constrained using (a).
However, I am also interested in testing changes in model fit between the 6-factor model and the 7 factor model. Constraining the factor loadings (a) gives a different chi-squared estimate compared to the unconstrained model. Can you please educate me on which of these is appropriate?