Indirect effect with count dependent ...
Message/Author
 Jason Jones posted on Wednesday, March 26, 2014 - 8:17 pm
Drs. Muthen,

I'm trying to test a mediation model with a count dependent variable (input code below). My understanding is that I cannot use the "model indirect" command with a count outcome. From reading through posts it seems I need to use the "model constraint" command, but after thorough searching and reading through "Applications of causally defined direct and indirect effects in mediation analysis using SEM in Mplus", I'm still just not sure what my input code should look like to get the indirect effects. Can you please advise? Thank you so much for your help.

Usevariables ARE avoM_P1 avoM_P2 avoM_P3
anxM_P1 anxM_P2 anxM_P3 Cmon4 Dmon4 Mmon4
Risk4;
count = Risk4;

MISSING ARE ALL (999);

Model:

avoM by avoM_P1* avoM_P2 avoM_P3;
avoM @ 1;
anxM by anxM_P1* anxM_P2 anxM_P3;
anxM @ 1;
Know by Mmon4* Dmon4 Cmon4;
Know @ 1;

Risk4 ON Know avoM anxM;
Know ON avoM anxM;

Model indirect:
Risk4 IND Know avoM;
Risk4 IND KNow anxM;

Output: standardized ;
 Linda K. Muthen posted on Thursday, March 27, 2014 - 1:50 pm
You can use MODEL CONSTRAINT to create the product of the risk4 ON know and the know ON avoM coefficients. This is the indirect effect with respect to the log rate. The paper shows how to do this for the mean. It is not simply a matter of exponentiating the indirect effect. Version 7.2 will have an option to do this.
 Jason Jones posted on Monday, March 31, 2014 - 6:23 pm
Dr. Muthen,

Thank you so much for the helpful information. The input below seems to work, but can you confirm that it looks correct to you? Y is a count DV, the Xs and mediator are continuous.

Model:
M on X1 (a1);
M on X2 (a2);
Y on M (b);
Y on X1 X2;

Model Constraint:
New (med1 med2);
med1 = a1*b;
med2 = a2*b;

If I understand correctly, the effects on the count outcome (including the indirect effects) are interpreted in terms of the log rate. But can the effects of the Xs on the mediator be interpreted as standard regression coefficients? Is there anything else I need to do to report the indirect effect results? Or just report the numbers in the output? Thank you so much again for your help.
 Bengt O. Muthen posted on Monday, March 31, 2014 - 6:33 pm
Yes, the effects on the mediator are regular linear regression coefficients.

For what to report with mediation, see MacKinnon's book and/or ask on SEMNET.