Mediation with imputed data PreviousNext
Mplus Discussion > Structural Equation Modeling >
Message/Author
 Maria Niarchou posted on Thursday, May 24, 2012 - 2:20 pm
Hi Linda and Bengt,

I am exploring mediation and need to estimate the indirect effect.

I used multiple imputation to deal with missing data and I am now running the analyses on the imputed data. Hence, I specify type=imputation.

Following earlier posts, to obtain the indirect effect I am using the MODEL CONSTRAINT subcommand to obtain the indirect effect in this way:

Model:
y on m (p1);
y on y;
m on x (m);

MODEL CONSTRAINT:
new(ind1);
ind1 = p1*m;

In the output line I request standardised estimates by specifying std (output: std).

However, in the output 'ind1' is estimated for the unstandardised but not for the standardised results.

I could manually multiply the corresponding standardised estimates for 'm' and 'p1' but I would like to obtain the standard error and p values as well.

Is there any way I could specify my model so that it calculates the standardised indirect effects?

Many thanks
 Linda K. Muthen posted on Thursday, May 24, 2012 - 4:26 pm
You would need to standardize it in MODEL CONSTRAINT to obtain the standard error.
 Maria Niarchou posted on Thursday, May 24, 2012 - 4:36 pm
Thanks Linda,

I am afraid I am not very familiar with this - I would I achieve it?
 Linda K. Muthen posted on Thursday, May 24, 2012 - 6:57 pm
See Example 5.20.
 Maria Niarchou posted on Friday, May 25, 2012 - 10:23 am
Hi Linda,

I had a go at following example 5.20 to obtain the standard errors for the standardised estimates.

However, I must have set up the model incorrectly as I received an error message which suggested I should specify

PARAMETERIZATION=THETA

and when I specified it (to see what would happen) the model could not even be identified.

Could you recommend a reference related to example 5.20?
 Linda K. Muthen posted on Friday, May 25, 2012 - 11:57 am
Please send your output and license number to support@statmodel.com.
 Maria Niarchou posted on Tuesday, July 10, 2012 - 11:46 am
Dear Linda,

To calculate the direct/indirect effects on imputed data, I used the commands

Model:
y on m (p1);
y on x;
m on x (m);

MODEL CONSTRAINT:
new(ind1);
ind1 = p1*m;

I then went on to adjust the model to account for potential confounders (cov1 - cov5).

Model:
y on m (p1);
y on x cov1 - cov5;
m on x cov1 - cov5 (m);

MODEL CONSTRAINT:
new(ind1);
ind1 = p1*m;

The results I obtain are much more different than the ones on the complete cases where I am able to use the model indirect command:

Model:
y on m ;
y on x cov1 - cov5;
m on x cov1 - cov5;

Model indirect:
Y on X;

However, if I aply the model constrain command to the complete cases the results are in line with those of the imputed data.

It would be great if you could help me to identify where I have gone wrong...

Many thanks
 Linda K. Muthen posted on Tuesday, July 10, 2012 - 4:16 pm
The variable y have five more predictors in the second model. The regression coefficients are partial regression coefficients controlling for all of the predictors. This is why the indirect effect changes.
 Maria Niarchou posted on Tuesday, July 10, 2012 - 6:03 pm
Thanks Linda!
 Maria Niarchou posted on Wednesday, July 11, 2012 - 10:04 am
Hi Linda,

I just realised that my email was not 100% clear.

I used the following line on complete cases to adjust for 5 confounders:

Model:
y on m ;
y on x cov1 - cov5;
m on x cov1 - cov5;

Model indirect:
Y IND X;

And I now want to do the same for imputed data so I need to reflect the process in the lines above when writing up the model constrained. So I use:

Model:
y ON cov1-cov5;
y on m (p1);
y on x cov1 - cov5;
m on x cov1 - cov5 (m);

MODEL CONSTRAINT:
new(ind1);
ind1 = p1*m;

This produces results equivalent to the ones in with the model indirect command but I wanted to check with you that this is sound.
 Linda K. Muthen posted on Wednesday, July 11, 2012 - 5:08 pm
The following line holds all of the regression coefficient for m equal:

m on x cov1 - cov5 (m);

I think you want:

m on x (m)
cov1 - cov5;
 Maria Niarchou posted on Wednesday, July 11, 2012 - 8:36 pm
The line above does not reproduce the results I had obtained using model indirect in the complete cases.

However, if I rewrite it as below I do replicate the results:

m on cov1-cov5;
m on x (m);

Is it OK to regress the covariates on the mediator to reproduce the adjustment I originally made on the IV-> m path (
m on x cov1 - cov5;)?
 Bengt O. Muthen posted on Thursday, July 12, 2012 - 2:21 am
The statement

m on x (m)
cov1 - cov5;

is the same as

m on cov1-cov5;
m on x (m);

If you don't get the same results, send those 2 outputs to Support.

It is certainly ok to add covariates in the m regression on x.
 Maria Niarchou posted on Thursday, July 12, 2012 - 7:08 am
Thank you Bengt,

You are right the two lines produce the same results - not sure what happened earlier.

Many thanks
 Minyoung Kwak posted on Monday, April 15, 2013 - 5:45 pm
Dear Dr. Muthen,

I would like to calculate the standardized indirect coefficient using the WLSMV estimator. I have used "Model Constraint" command because one of the dependent variables was binary, and sampling weights and multiple imputation were used. I would like to calculate the standardized indirect coefficient for the path from hp(binary)--> na -->dp.

DATA:
FILE IS "data.dat" ;
Type=Imputation;
VARIABLE:
Names are hp dp sa1-sa5 cont1 -cont5 ;
Categorical= hp;
Missing are all (-9999);
Weight= RWTR;
ANALYSIS:
Estimator= WLSMV;
MODEL:
na by sa1 - sa5 ;
dp on hp
na (a1)
cont1 -cont5 ;
na on hp (b1)
cont1 - cont5 ;
hp on cont1 - cont5;
dp (v2);
hp (v1);
Model Constraint:
New(a1b1 stdab );
a1b1=a1* b1;
stdab = a1b1*SQRT(v1)/SQRT(a1b1**2*v1+v2);

When I ran this syntax, I received an error message saying that " PARAMETERIZATION=THETA" was needed. However, when I added theta parameterization to the analysis command, all of the standard errors and p-values for coefficients were zero. I was wondering if my specification for the standardization of indirect path is incorrect and how I could fix this problem.

Thank you very much.
 Linda K. Muthen posted on Tuesday, April 16, 2013 - 2:56 pm
Please send the output and your license number to support@statmodel.com.
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: