Mediation with non-recursive models PreviousNext
Mplus Discussion > Structural Equation Modeling >
Message/Author
 Thomas Carpenter posted on Friday, May 11, 2012 - 2:08 pm
Hi there. I am trying some mediation path modelling. I have a series of predictors on my DV, some of which are mediated by others. However, two predictors have a bidirectional relationship. In other words

SelfForg on guilt apology recvforg morforg;
apology on guilt;
recvforg on apology;
morforg on apology;
guilt on apology;

There is more, but that's the important part. Once I introduce the non-recursive element to the model (guilt on apology;), I can no longer get the strength of each possible mediating path.

model indirect: SelfForg ind apology;

USED to generate a whole bunch of different mediating paths. Now it just tells me the sum total indirect effects and direct effects of apology on self-forgiveness (SelfForg). Does that make sense? What I want is to know how much of the indirect effect goes through each possible pathway (E.g., apology --> recvforg --> selfforg). Is there any way to do that? Or does adding the loop somehow derail our ability to know the strength of the individual mediating pathways (because you could go around in circles forever)?
 Thomas Carpenter posted on Friday, May 11, 2012 - 8:26 pm
Never mind, I was able to resolve the issue. Thank you.
 Paul W. Andrews posted on Monday, September 05, 2016 - 10:44 am
Hi I have a couple of questions about a model in which three continuous latent variables (X1, X2, X3) are related to each other in a cyclical fashion.

X2 on X1;
X3 on X2;
X1 on X3;

In each case, the latent variables are generated by three or more ordinal variables (each with 3 or more thresholds).

I am not sure if this is technically a non-recursive model, but I figured I'd post it here.

I apply this model to several different datasets (each with N>300).

My questions are:

1. I believe that this model is identified, because Mplus gives me estimates, but is there a way to be sure?

2. Assuming it is identified, I am uncertain about whether MLR or WLSMV is the best estimator. I get more stable estimates across the samples if I use MLR, but I’d like to be sure.

Your help would be much appreciated.
 Bengt O. Muthen posted on Monday, September 05, 2016 - 3:43 pm
1. I am not sure this model is identified, but perhaps it is just another way of representing their 3 covariances. You may want to ask on SEMNET. You should check the condition number Mplus prints - if the model is identified, it should be higher than at least 1 to the power of -10. You would also see very large SEs if the model is not identified.

2. When you say MLR I hope you don't mean not declaring the variables on the Categorical list, that is, ignoring their ordinality. WLSMV is probably not great here. Perhaps Bayes - which would be another check on identifiability - see the trace.
 Paul W. Andrews posted on Monday, September 05, 2016 - 4:58 pm
Hi Bengt,

Thanks for the feedback. I did declare the ordinal variables as categorical, and specifically requested "MLR" as the estimator.

In all the datasets where I applied the model, the Condition Number was on the order of 10E-2 to 10E-4. So given what you said, that seems like the model is identified.

I'd be interested to know why you think WLSMV is probably not so good here, if you have the time or can give me a reference.

Best,
Paul
 Bengt O. Muthen posted on Tuesday, September 06, 2016 - 3:15 pm
Note that with ML/R the predictor in each equation is treated as a continuously scored variable despite it being treated as an ordinal DV.

WLSMV doesn't handle missing data as well and you have to assume normality for the underlying latent response variables. WLSMV does not, however, have the problem of ML/R above. Bayes would be better in this regard.
Back to top
Add Your Message Here
Post:
Username: Posting Information:
This is a private posting area. Only registered users and moderators may post messages here.
Password:
Options: Enable HTML code in message
Automatically activate URLs in message
Action: