

Model Fit indices: dichotomous & cont... 

Message/Author 

Gavin Elder posted on Thursday, November 04, 2010  11:34 am



Hello there, I am testing a model in which three mediators account for the relationship between one IV and one DV (3 separate indirect effects). One of the mediators is a dichotomous variable (70% group 1, 30% group 2), while the rest of the variables in the model are continuous (all are "observed" variables, not latent). I have tested the model in MPlus using both ML and WLSMV, separately. Each estimator produces the same pattern of results, but of course, WLSMV has only two indicators of model fit: RMSEA and WRMR. Model fit for both estimation procedures is solid across indices. I read a paper in which the authors reported using a combination of the two approaches, and subsequently reported the greater number of model fit indices (e.g., CFI/TFI, AIC/BIC, Chisquare). The problem is twofold: (1) I cannot find any suggestions online about which model fit indices need to be reported when you have the situation of dichotmous AND continuous outcome variables, and (2) I'm not sure how to approach this in MPlus (i.e., which estimator to specify to obtain more indices of model fit). Any help you could provide would be greatly appreciated! Thanks! PS: I'm using analysis: bootstrap=2000, and output: standardized cinterval(bcbootstrap) 


Because you are using BOOTSTRAP, you obtain only WRMR which is an experimental fit statistic. If an RMSEA is given, it is not correct and should not be interpreted. If you are interested in indirect effects, you need to use WLSMV. In Mplus, indirect effects with maximum likelihood estimation when there is a binary mediator are not defined. 

Gavin Elder posted on Friday, November 05, 2010  3:55 am



Thank you for your reply Dr Muthen. If I understand correctly: using a model with one categorical mediator and two continuous mediators (total=3 indirect effects), explaining the relationship between one IV and one DV (both continuous)  I will only report WRMR to indicate model fit? Should I be concerned that reviewers might have reservations? 


I would run the model without the BOOTSTRAP option to obtain fit statistics that have not been bootstrapped and report them and the bootstrapped standard errors. 

Back to top 

