

Effect of a dichotomous mediator 

Message/Author 


Hi, 1.For modeling SEM with Mplus, is it necessary to take out the nonsignificant paths as often done with LISREL/AMOS? 2.When using a dichotomous mediator (M1d) and continuous predictors (X1, X2) and outcome (Y) do I need to specify the use of WLSMV estimator or any other estimator? I also have a second continuous mediator (M2) and two covariates (cv1 and cv2) in my model. M2 and Y are latent with 4 and 3 indicators respectfully, while the rest are manifest variables. I am trying to determine whether the path to Y is X1>M1d>Y or X2>M2>Y. X1 and X2 are correlated. I don’t need to model this correlation as default in Mplus is correlating independent variables, correct? 3.I do get X2 > M2d and M2d >Y are significant when cv1, cv2, X2 and M1 are in the model too. Using bootstrapping to estimate the confidence interval for the indirect mediated effect (if that is okay), do I model "Y ind M2d X2" or "Y ind M1 M2d X1 X2 cv1 cv2"? 4.Is the confidence interval for the effect of the mediation that I want use the first one given, the STDYX Standardization, or the STD standardization one? How can I get more decimal points? Is there a way to get the 90% confidence interval instead? Thanks for the help (this is a repost to correct for not limiting my post to one window before, so sorry if there ends up being duplication). 


1. I do not believe this type of model trimming is good practice. 2. Use WLSMV. 3. Y ind M2d X2 is correct. 4. Use StdYX. 


Thanks! Regarding: #2 for setting estimator = WLSMV, that is in addition to setting parameterization = theta correct? Is that all I need to do to address the dichotomous mediator? #3 Then is Y ind M2d X2 taking into account the covariates? Theoretically, I think for my hypothesis it is important for X1 and X2's correlation to be taken into account to estimate the mediation effect of M2d for the path from X2 to Y. Is it necessary to explicitly state something like X1 WITH X2 anywhere? My understanding was that the default for all independent predicting variables is to be correlated. #4 Is there a way to get the 90% confidence interval for bootstrapping? I saw in the user guide that it can be given but I don't see it anywhere in my output nor anywhere in the list of options to select. Also, why does the pvalue from some paths change after selecting the bootstrapping option? 


I would only use the Theta parameterization if I got a message saying the model cannot be run without it. Yes, the indirect effects uses partial regression coefficients. See the CINTERVAL option in the user's guide. The pvalues change because the standard errors change. 


Sorry, I still can't find anything in the user guide to specify getting the lower 5% and upper 5% confidence intervals. It seems like it is supposed to just be generated when cinterval(bootstrap) is specified in the output and bootstrap = 5000 is specified in the analysis. Is the lower and upper 5% only available on newer versions of Mplus? I'm working with Mplus5. 


They should come out automatically. This was also the case in Version 5. 

Back to top 

