We don't have such an example. Just BY statements to define the latent variables and ON statements to relate the observed and latent variables according to you path model. Nothing special needs to be done in this case. Path model examples are in Chapter 3. CFA examples are in Chapter 5.
Hello, I have a similar question/concern. I am conducting an SEM and unsure how to explain my model using the traditional "measurement" and "structural" model differentiation/fit comparison.
I have two latent variables theorized to be indicated by 5 observed (manifest) variables, and I want to examine whether these latent variables predict 6 observed outcome variables. When I entered all the variables into the model simultaneously and specified the ON statements, I was able to modify the latent variable (measurement?) portion of the model and improve the fit.
Is this considered an SEM or path analysis? Do I need to model the latent variables separately (do not include the observed outcome variables in the usevar command), and then run a second model with the observed outcome variables included?
Thank you. To confirm, I should only test the latent portion of the model and include those variables in the “usevariable” command (measurement model). Once this fits, I can then test the full structural model by adding the observed outcome variables to the "usevariable" command and add the ON syntax from the latent variables to these outcome variables.
In terms of the structural model, is it acceptable to make additional modifications to the latent variables once the outcome variables and ON statements are added? I had strong fit for the measurement model but poor fit for the structural model. I was able to improve fit for the structural model by making additional modifications to the (measurement) latent variables (i.e., additional indicators were freed to load on latent variables).
What would it mean that additional modifications to the measurement (latent variables) portion of the model were needed at the structural level? How would one interpret/explain this to a reader? Thanks!
I am new to using MPlus and am running a partially latent path analysis. For partially latent models in LISREL, I was taught that if I assume the measurement error variance for the observed variables (the single indicator latent variable") to be 0, then I have to specify that in the model. Is this also the case in MPlus? Since I could not figure out how to do so, I ran a model with the syntax below (I list my variables first).
Variables in my model: Predictors: religiosity (latent), contact ("cp2," observed), and gender (observed). Outcome: transphobia (latent) All predictors have direct effects on the outcome variable, but religiosity also has an indirect effect on the outcome.
DATA: FILE = MPlusTrimmed_n=399.dat; VARIABLE: NAMES = Gender CP1 CP2 TP1 TP2 TP3 RFP1 RFP2 RFP3; USEVARIABLES ARE Gender CP2 TP1 TP2 TP3 RFP1 RFP2 RFP3; ANALYSIS: ESTIMATOR IS ML; MODEL: TransPhob by TP1 TP2 TP3; Relig by RFP1 RFP2 RFP3; TransPhob on Gender Relig CP2; CP2 on Relig; MODEL INDIRECT: TransPhob ind Relig; Output: Stdyx;
Is the syntax correct? I am confused by the output as it shows no estimated error variance for gender, but it does for contact. I would appreciate help understanding this difference.
There is no need to put a factor behind a single indicator with zero measurement error - this is done automatically by Mplus.
Mplus treats covariates ("x variables", exogenous observed variables) as not having the parameters of their marginal distribution being part of the model - just like in regression. Your gender variable is an "x variable" but your CP2 variable is not - so you see mean and variance parameters related to CP2 but not related to x.
Just to make sure I understand, the variance estimated for cp2 is the equivalent of a 'disturbance' in SEM language because it's locally exogenous?
I also wanted to ask if the ML estimation I used is an appropriate estimation method for a model like mine where one of the predictor variables (i.e., gender) is a dichotomous nominal variable? In other words, a model that includes both nominal and continuous predictor variables? From what I read, it seems so, but I wanted to confirm.
Lastly, I wanted to confirm that I do not need to specify gender as “Nominal” under “Variable” because it’s not a dependent/endogenous variable?