I have run a very basic path analytic model using the complex procedure using the primary code below. Using the code below, however, I do not receive the correlations among the exogenous variables (covariates) in the output file as traditionally would occur using the With statement in other MPlus procedures. Is this not a default option in the Complex procedure?
CLUSTER IS agencyid; MISSING IS BLANK; USEVARIABLES private agysup9 ebpasm ed white q47 yrsagy male Q7;
In most parts of Mplus, the model is estimated conditioned on the covariates as in regular regression so the means, variances, and covariances of the covariates are not estimated as model parameters. If you want these values, you can obtain them using TYPE=BASIC;
Thanks Linda...let me pursue this a little further. When I constrain some of the correlations between covariates to 0 (e.g., agysup9 with private) not only do the degrees of freedom increase (as if MPlus is estimating this correlation), but it gives me the correlations among the other covariates in the output.
I was hoping you could help me understand the benefit of running a path model with observed variables set up as latent variables with one indicator versus simply running the model with the observed variable as a true observed variable. I was advised in the past to run a path analysis using latent variables with one indicator and am getting highly inadequate model fits. When I re-run these models using observed variables, the model fits are adequate. I am not quite sure what could be driving this difference.
There should be no difference between using observed variables versus putting a latent variable behind each observed variable if the residual variance of the observed variable is fixed to zero, for example,
I run a LPA with 2 classes as best solution and included the latent class variable as binary outcome in a regular path analysis. Is there a way to link both analyses and do you have any examples for that? Or is this an appropriate procedure?
I'm conducting a multiple-group path analysis (a MANOVA-type of an analysis), but I'm a bit confused what would be the preferred way of modeling the IVs: should I include e.g. the covariances between the IVs into the model (which seems to treat the IVs as DVs?) or not (as in regular regression approach)? This troubles me, because including these covariances (even by setting them zero) seems to influence the sample size used in the analysis. I'm using MLR estimator.
The default in Mplus is to use all available information and TYPE=MISSING. Missing data theory applies to only the dependent variables. If you include the covariates in the model instead of estimating the model conditional on the covariates, you make distributional assumptions about them. It is your choice whether it is worth doing this.