I am using MPlus 6.12 to conduct a path analysis with 1 binary outcome variable (DV), a continuous mediating variable, and 5 continuous IVs. I have a couple of questions regarding this:
(1) I expect from past research that my 5 IVs will be correlated. Should I therefore specify this in the input syntax? If so, should I use WITH (i.e. v3 WITH v2), or something else? I read that Mplus assumes that exogenous variables are correlated but I have noticed that I am given different model fit results depending on whether or not I specify these correlations.
(2) At the moment I am using WLS estimator for this model, but I have noticed from previous posts that others have used WLSMV. Can you tell me which is best in my particular case?
(3) Lastly, I have a broad question regarding improper solutions for SEM and/or path analysis. Is it incorrect to have standardized estimates that are negative (less than 0)? Or does this just indicate a negative relationship between the variables?
1. You should not bring the covariates into the model. The model is estimated conditioned on the covariates. Their means, variances, and covariances are not model parameters. If you want to see their correlations, get their descriptive statistics. With WLSMV doing this changes the model.
3. Factor loadings can be positive or negative. They are regression coefficients.
Thank you Linda. I have a follow-up question. I'm wondering what the advantage of using an ESTIMATOR=ML approach with my model may be? I do not have any missing data, so if I did employ ML would I need to use numerical integration? Lastly, can I obtain estimates of the indirect effects if I use ML? I know that MODEL INDIRECT cannot be used with ML but I'm wondering whether these can be obtained another way?
You would need numerical integration for maximum likelihood and categorical outcomes. MODEL INDIRECT is not available in this case. You would need to use MODEL CONSTRAINT to define the indirect effects.
lopisok posted on Wednesday, March 11, 2015 - 4:43 am
I'm doing a path analysis with a binary outcome, 2 continuous mediators and 6 continuous IV's.
I followed the instructions from EXAMPLE 3.17: PATH ANALYSIS WITH A CATEGORICAL DEPENDENT VARIABLE AND A CONTINUOUS MEDIATING VARIABLE WITH MISSING DATA Using MLR and montecarlo integration.
Everything works fine but I was wondering why I do not get any fit statistics like TLI, CFI, RMSEA, SRMR? Can I only compare models through the BIC AIC?
When I ask MODINDICES I get the message " MODINDICES option is not available for ALGORITHM=INTEGRATION. Request for MODINDICES is ignored."
I read in other posts that when using listwise deletion the montecarlo integration command can be dropped. I did this but I still get the same message. Is it just impossible or senseless to ask for cfi, tli for some reason I don't understand at the moment?
That example uses maximum likelihood estimation. ML and categorical outcomes requires numerical integration. In this situation chi-square and related fit statistics are not available. You can use the default WLSMV instead. This will give probit rather than logistic regression.
lopisok posted on Friday, March 13, 2015 - 4:04 am
Thank you very much for your answer. Is ML recommended or WLSMV? Is there an advantage in using one estimator over the other? Does the advantage mainly lie in WLSMV providing fit statistics?
Does it remain correct if I use ML and report the standardized coefficients?
You can use either estimator. WLSMV is better with more factors because ML with categorical requires one dimension of integration for each factor with categorical indicators. ML has better missing data handling. And if fit statistics are important to you, you can get those with WLSMV.
It is correct to use ML and the CATEGORICAL option.
MSP posted on Tuesday, December 08, 2015 - 12:00 pm
Hi. I am trying to do a path analysis with model:
Y1 on X1-X10; Y2 on X1-X10; Y3 on X1-X10; Z on Y1-Y3; !X and Y are continuous (Ys are factor scores of Xs) !Z is dichotomous outcome
When I don't use an estimator (default WLSMV), it outputs this: SERIOUS COMPUTATIONAL PROBLEMS OCCURRED IN THE UNIVARIATE ESTIMATION OF THE THRESHOLDS/MEANS, VARIANCES AND/OR SLOPES FOR VARIABLE Y1/Y2/Y3
When I put estimator = MLR, it outputs normally and the fit indices are just the Loglikehood (H0) and the Information Criteria (AIC/BIC).
My questions would be: 1. Is MLR estimator an acceptable choice for this path model? 2. Are two fit indices acceptable in this case? No chi-square available?
It seems strange that you have both factor scores Ys and their indicators Xs in the model.
It is also strange that WLSMV would have problems when ML doesn't. Perhaps you Xs have very large variances.
We need to see the two outputs to say more - send to Support along with your license number.
MSP posted on Thursday, December 10, 2015 - 10:51 am
Thanks for the response. It is strange.
I was advised to compare a model where the factors are fitted (model with formative factors), and a model where the factor parts are 'constrained' (i.e. factor scores). I was trying to do the second model with this query.
I will send my outputs to the support email. I am using a school computer lab desktop though and not sure how to get license number. Please advise.
Dear professors, I run path model with binary outcome (dependent variable and mediator are continuous). My data was collected through a two-stage stratified cluster sampling procedure and I also took into account sample weights. The model looks like X-M-Y. I use command Model indirect to estimate total and indirect effect. I have a couple of questions regarding this model.
1. What measure is used to estimate relationship between X and M? 2. It seems like total, direct and indirect effect are given as a probit coefficient, but coefficient that shows effect from X to M is given in other units of measurement. I multiplied coefficients ( coefficient from X to M )* (coefficient from M to Y ) by myself to get total effect and I got the same total as mplus gave me. So does mplus multiply coefficients inspite of their different units of measure? 3. Is it possible to apply KHB method in mplus?
1. I think you meant above independent and mediator variables are continuous so this is a linear regression. 2. Direct effect is a probit coefficient with WLSMV which I assume you are using. Indirect effect is a product. 3. What is KHB?