Causal mediation with binary variables PreviousNext
Mplus Discussion > Categorical Data Modeling >
 Katherine Keenan posted on Tuesday, November 11, 2014 - 4:54 am
I am performing a causal mediation analysis with binary exposure x, binary mediator m and binary outcome y, but am getting unexpected estimates of the direct and indirect effect.
Model input:
m on x;
y on x;
y on m;
model indirect:
y ind m x ;

The model coefficients show a strong independent positive association between x and m, m and x, and x and y, so I would expect some direct and some indirect effects. But the estimate of natural DE and IE show there is no indirect effect at all-

I find this strange given that there is a strong association between X and M, and M and Y. I am estimating the model incorrectly?

Many thanks,
 Bengt O. Muthen posted on Wednesday, November 12, 2014 - 5:47 pm
We just responded to you on Support where the issue seems to be that the counterfactual causal effects are conditioned on covariates = 0. The value 0 may not be the substantively relevant value to condition on. When these covariates were centered so that their means are 0 you got larger effects.
 Selahadin Ibrahim posted on Tuesday, April 14, 2015 - 5:38 am
Hi Bengt,
Thanks for the 2014 paper “Causal Effects in Mediation Modeling: An Introduction With Applications to Latent Variables” and in bringing latent variables into the counterfactual models.
I have a question regarding Nominal mediators. I have a binary outcome and a nominal mediator with three categories. Do I have to use the known class method as is done in your previous working paper? Or is there an easier way of doing it in mplus version 7.3? your guidance on this is appreciated.


 Bengt O. Muthen posted on Tuesday, April 14, 2015 - 8:26 am
No easier way in 7.3; our new language doesn't cover that case. So, yes, take the Knownclass approach.

And if you get a nice application written up, please send to me.
 Selahadin Ibrahim posted on Tuesday, April 21, 2015 - 5:34 am
Thank you Bengt.

 Kathleen Kennedy-Turner posted on Friday, September 08, 2017 - 2:24 pm
Would it be possible to direct me to a source/paper that would explain why the PNDE is not significant in my model but the path coefficients are?

Or can I explain it as the PNDE is the pure effect from IV to DV and the path coefficients are also taking into account other variables in the model, meaning they are eating up the variance potentially allowing for that path from IV to DV to come through?

Thank you,
 Bengt O. Muthen posted on Friday, September 08, 2017 - 4:35 pm
Are both M and Y continuous?

Are you using ML? ML with bootstrapping? Or Bayes?
 Kathleen Kennedy-Turner posted on Friday, September 08, 2017 - 4:50 pm

M is continuous but Y is binary zero-inflated. I'm using WLSMV.

 Bengt O. Muthen posted on Sunday, September 10, 2017 - 1:14 pm
Binary, zero-inflated sounds strange; I am not sure that makes sense - just do binary.

If this doesn't help; please send your output to Support along with your license number.
 Kathleen Kennedy-Turner posted on Sunday, September 10, 2017 - 1:54 pm
Hi again,

Thank you. I have chosen WLSMV based on past discussions, as my dataset as way more zeros than 1s. But could that be affecting the PNDEs? I didn't find it appropriate to use ML given the zero-inflation.

I was just thinking that the path coefficients are derived by accounting for everything in the model, while the PNDE is only the effect from IV to DV, if I understand correctly. So I was thinking the path coefficients are significant because there is variance accounted for by other variables which is not the case for the PNDEs?

 Bengt O. Muthen posted on Sunday, September 10, 2017 - 5:14 pm
I would need to see the output to say what is going on.
 Amelia Rock posted on Wednesday, August 07, 2019 - 1:02 pm
Hello, I am conducting mediation analysis with continuous Xs, continuous and ordinal Ms, and a binary Y. I am trying to understand whether I can use the counterfactually-estimated indirect effects I got when I ran the model without specifying the two exposure values to be compared as one must do when the exposure is continuous. The user guide says the default is one for the first value and zero for the second value, but I do not know whether this makes sense for my exposures (observed means of scales measuring experiences of stigma). Is there guidance for choosing the two values to be compared (understanding of course that subject area knowledge should be applied as well)?
 Bengt O. Muthen posted on Thursday, August 08, 2019 - 10:10 am
The scale of X is important. Standardize X so that 0 is the mean and 1 is one SD above the mean. Otherwise your results are meaningless.
 Amelia Rock posted on Friday, August 09, 2019 - 11:14 am
Makes perfect sense, many thanks.
Back to top
Add Your Message Here
Username: Posting Information:
This is a private posting area. Only registered users and moderators may post messages here.
Options: Enable HTML code in message
Automatically activate URLs in message