Message/Author |
|
|
I am evaluating a mediation model with an independent variable, two mediating latent variables, and a dependent variable. I have three control variables (age, SES, and ethnicity)for which I want to account in the model based on significant bivariate correlations with indicator variables. Where is it best to include them of the following options? 1. Regress the indicator variables with which these are correlated directly on the control variables? 2. Regress the latent variable with which at least one indicator variable is correlated with a control variable on that control variable? 3. Some other method? Is it correct to have these control variables in multiple places in the model (i.e., regressed on both latent variables if appropriate)? Thank you! |
|
|
You should use the approach in number 2. |
|
|
Dear Professor(s) Muthen, I would like to follow up on this thread. I would be interested in the answer to the second question: Is it correct to have control variables in multiple places in the model (i.e., regressed on both latent variables if appropriate)? My example is the following: What is the difference between the following two syntaxes: PRmeri BY eqOPP rewEFF IQ; PRmeri on male edu age age2 empl; rJP on PRmeri male edu age age2 empl; VS. PRmeri BY eqOPP rewEFF IQ; PRmeri on male edu age age2 empl; rJP on PRmeri; Are the controls for my second equation needed? Or had Mplus already "done the job"? Does Mplus work with already the partial correlations? Thank you in advance for your help, Zsofia Ignacz |
|
|
In you second option, you leave out male edu age age2 empl. Mplus will not control for these variables if they are left out. I suggest using option 1. |
|
|
Dear Professor Muthen! Thank you very much for your quick answer! Kind regards to you, Zsofia Ignacz |
|
|
Dear Dr. Muthen, Please I regress my DVs and mediators on control variables, my fit decreases substantially. Is there a way to fix this? Thanks! |
|
|
Unless you also have moderation involving the mediator, you could look at the Modification indices. |
|
|
I am having difficulty with model fit when adding covariates. I am conducting a mediation model with four independent variables, one mediator, and two outcome variables (all observed variables). I have four control variables (Age, SO_Gay, SO_Lesb, SO_Bi_Ma) When I run the syntax as follows: PTSD ON LGB_Trau Personal Vicarios Trauma RS_like Age SO_Gay SO_Lesb SO_Bi_Ma; Depress ON LGB_Trau Trauma Personal Vicarios RS_like Age SO_Gay SO_Lesb SO_Bi_Ma; RS_like ON LGB_Trau Trauma Personal Vicarios; I get great fit of the model. I read that I am supposed to include covariates on every line of "ON" syntax. However, when I add the covariates to the mediation line (RS_like) the model fit indices change drastically with Chi-Square, RMSEA and SRMS equal to 0 and the CFI/TLI equal to 1. Am I not adding the covariates properly? Why is this happening? |
|
|
It looks like your model becomes just-identified when you add the covariates to the mediator. Just-identified means that the model imposes no restrictions - there are no left out paths. This means you have zero degrees of freedom and the model cannot be tested. This is not a problem per se. |
|
|
Thank you for the quick response. Does this then mean that I should ignore/not report model fit indices? I ask because this is a basic model with which I plan to add more mediators but anticipate the same problem occurring (i.e., that by adding covariates to the mediators the model will become just identified). What would be the best method of handling this? |
|
Back to top |