Message/Author 


Hello. I'm trying to run a multiple mediation model (1 IV, 2 mediators, 2 outcomes) with FIML, and with bootstrapped SEs and biascorrected CIs. A main goal is to examine specific indirect effects. Assuming I have read in the data correctly, etc., does the following segment of code look correct? analysis: type = general missing H1; estimator = ML; iterations = 50000; bootstrap = 10000; model: med1 on indvar; med2 on indvar; outcome1 on med1 med2 indvar; outcome2 on med1 med2 indvar; med1 with med2; outcome1 with outcome2; model indirect: outcome1 IND med1 indvar; outcome1 IND med2 indvar; outcome2 IND med1 indvar; outcome2 IND med2 indvar; output: sampstat standardized cinterval(bcbootstrap); Also, under the "Model results" section, I assume "Estimates" are unstandardized and "Std" are standardized, but I get equivalent "Estimates" and "Std" results. So I guess "StdYX" are standardized? Under the "total, total direct, specific indirect, and direct effects" section, again I assume "Estimates" are unstandardized and "StdYX" are standardized? Thanks very much for your help. 


It's hard to look at code and say if it looks okay. The best way to check is to run it and see if you get any errors, and if not, to check the results to see if you get what you expect. Estimates are not standardized. Both Std and StdYX are standardized. Std uses the variances of latent variables. Because you have none, they are therefore the same as the raw coefficients. Same for indirect effects. See the user's guide for more information. 


I didn't get any errors, but things are not always as they seem. I think I have done enough now, though, to be confident in that part of the code. And thank you for clarifying about the raw and standardized estimates. What I am trying to do now is a direct comparison of the two mediators in the model, to determine whether one is a stronger predictor. So, I would like to create a contrast of the two specific indirect effects (specific indirect effect for mediator 1 minus specific indirect effect for mediator 2) and to bootstrap a CI for it. Can you provide some guidance on how to code this in the program above? Thanks very much. 


You can use MODEL CONSTRAINT to create indirect effects and test their equality. Example 3.10 shows an example of how MODEL CONSTRAINT is used. See also the full description of MODEL CONSTRAINT in the user's guide. You can then use the CINTERVAL option to obtain bootstrapped confidence intervals. See the description of the CINTERVAL option in the user's guide. 


Hi Linda and Bengt, I have run a model with multiple mediators using FIML, have bootstrapped the model, have asked for indirect effects and confidence intervals. For the significant specific effects (i.e., +/ 1.96) the confidence intervals (95%) do not pass through zero, as I would expect. However, I have three effects that are not significant according to the est/S.E (zscore), yet the confidence intervals do not pass through zero. For example, for an a>b>c relationship I have a unstandardised bcoeff of .13, a S.E. of .08, and a zscore of 1.76. Yet the lower CI is .32, and the upper CI is .04. Do you know what is going on here, and what I can conclude? Is this a significant indirect effect or not? Many thanks for your help in advance. 


The zscore should be compared to a symmetric confidence interval. The fact that the bootstrapped confidence interval does not behave like the symmetric confidence interval suggests you should use the bootstrapped confidence interval. 


Dear Drs Muthen, I am a relatively inexperienced Mplus user and was wondering if FIML is used (by default) for path analysis? Specifically, does FIML handle missing data for (manifest) predictor and mediator variables (where some are categorical), as well as control variables that are included in the model via correlations with the predictor variables and pathways to the mediators. Thank you kindly, Ann 


FIML is used for all endogenous variables. It is not applied to exogenous variables. The theory does not cover this. If you want it to be applied to exogenous variables, you must bring all observed exogenous variables into the model by mentioning their variances in the MODEL command. When you do this, they are treated as endogenous variables and distributional assumptions are made about them. 


Hi Linda, Thank you for your response. Can this be done using the (example) script below? Also, would the descriptives (SAMPSTAT) produced by the model reflect FIML? Or would they have been computed with the missing data (and thus with case wise deletion)? MODEL: TOM ON LogOT2 (a); Remote ON TOM (b); Remote ON LogOT2; TOM ON parity (C1_a); Remote ON parity; LogOT2 WITH parity; TOM ON eduyears (C2_a); Remote ON eduyears; LogOT2 WITH eduyears; TOM LogOT2 Remote parity eduyears; ! Variances LogOT2 WITH TOM; ! Covariances TOM WITH parity; TOM with eduyears; Parity WITH eduyears; 


You need to mention the variances of only variables that appear only on the righthand side of ON. Remote does not fulfill that criterion. Check your output to be sure you get what you expect. The sample statistics are estimated using FIML. 


Great, thank you very much! One additional question: can categorical predictor and mediator variables (with more than 2 categories) be included in the model of a path analysis (by creating dummy variables) even if they cannot be MVN? 


Yes. For such mediators you put them on the categorical list. The WLSMV estimator is suitable here. 


Just to followup, is FIML used with the WLSMV estimator? Thanks again! 


No. WLSMV uses pairwise present. 

Back to top 