Message/Author 

Jetty posted on Wednesday, January 08, 2014  8:57 am



I am running a path analysis that features a combination of continuous and ordered categorical endogenous variables. The mediators are continuous or binary. N is 1200. All variables are observed. Two questions: First, Do I have to worry about the distribution of any of the variables? Does bootstrapping take care of potential nonnormality and other issues? Some of covariates are highly skewed. Second, I plan to use the diagram to display the overall results. Given that I have a combination of ordered categorical and continuous outcomes, which is the best estimate to use (parameter estimates, STDYX, STD, or a combination? Thank you! 


Use WLSMV and put the categorical DVs on the Categorical list. You can look at the STDYX values  unless the effect is from a binary covariate in which case you use STDY. You don't have to worry about distributions and with n=1200 there is likely no need for bootstrapping. 

Jetty posted on Wednesday, January 08, 2014  11:55 am



Thanks, Bengt! Is there a way to get MPlus to produce STDY? 


This is forthcoming in the cases not currently given, but in the meanwhile you can simply use STDYX and get to STDY for a binary covariate by dividing it by the sample SD of the covariate. 

Jetty posted on Monday, February 10, 2014  9:39 am



Bengt, I am trying to decide whether using MLR might be better in my case since I would like seeing logits and ORs for interpretability purposes. I have a path analysis, categorical ordered DV, all mediators are continuous (skewed but nevertheless continuous). I ran my models using both MLR and WLSMV and nothing changes substantially. My main loss is the ability to estimate mediated effects (IND) as they don't seem to be available with MLR. Is MLR an acceptable choice in my case? Thank you! 


I would recommend MLR above WLSMV if you have a lot of missing data. Otherwise, the choice is yours. You can use MODEL CONSTRAINT for indirect effects when MODEL INDIRECT is not available. 

Jetty posted on Tuesday, February 11, 2014  7:29 am



Linda, is it ok to use MODEL CONSTRAINT with different scale coefficients (combination of linear regression coefficients and logits)? What is the best way to assess the statistical significance of the indirect effects? Thank you 


How the coefficients should be defined in MODEL CONSTRAINT is discussed in the following paper available on the website: uthén, B. (2011). Applications of causally defined direct and indirect effects in mediation analysis using SEM in Mplus. There will be an option for the effects in the next version of Mplus. 

Jetty posted on Wednesday, February 26, 2014  9:05 am



Hi Bengt, In your very first response you mentioned that I shouldn't worry about distributions in response to my concern that some of my continuous mediators have high skew and kurt. I decided to run the analysis in two ways: using the nontransformed variables and then using a logtransformed variables with normal skew and kurt. The results are different for the 2 paths going from IV to the transformed mediators. I am not sure which results to trust. Thank you! 


That indicates that the untransformed mediators don't have linear regressions on the IVs. You may want to very that by looking at plots. In such a case you may want to make a transformation to make the relationship more closely linear. 

Back to top 