Message/Author |
|
|
I thought I had posted this yesterday but it hasn't appeared, apologies for any duplication. I can fit a path analysis model with a negative binomial dependent variable using MLR and integration monte carlo. However wheni try to estimate an indirect effect ( potential mediator is binary) I receive an error message that the indirect and bootstrap commands are not available with the integration command. After much trial and error I haven’t found another way of estimating the model. How can I estimate indirect effects when the dep variable is negative binomial please? |
|
|
You can express the indirect effect using Model Constraint. With a binary mediator and ML, however, you need to consider a different type of indirect effect that just a product of slopes. See Muthén, B. (2011). Applications of causally defined direct and indirect effects in mediation analysis using SEM in Mplus. which is posted on our web site. |
|
|
Many thanks. |
|
|
Continuing the above. I understand that the product of the components of an indirect path only has a natural interpretation when both equations are linear. However, presumably if we use CONSTRAINT the p value associated with ab is some sort of test that ab is zero i.e. that that indirect path exists. Could you clarify please? How do we report the coefficient? ( I did look at the document you recommended but it is too technical for me.) Second I note that indirect appears to give exactly the same ab coefft and p value as using CONSTRAINT. In that case why isn’t INDIRECT available for neg binomial or have I misunderstood? Many thanks |
|
|
As you said MODEL INDIRECT and MODEL CONSTRAINT get the same results when a product is an appropriate expression of an indirect effect. They are both interpreted in the same way. If you have a count variable as a mediator or a final outcome, using a product for the indirect effect is not correct. I do not know what would be correct in this case. |
|
|
Thanks Linda. |
|
|
This is a follow-up question to the above discussion re testing indirect effects with a negative binomial DV. Am I understanding correctly that it is not possible to calculate indirect effects if the negative binomial DV is a count variable? I am using version 6.12 |
|
|
As I said above, I don't know what a correct indirect effect would be for a count variable. I'm not sure there is one. |
|
|
Dear Dr. Muthen, I have two variables (AxB) that have a positive correlation but when tested in a mediation analysis, variable A has a significant NEGATIVE INDIRECT effect on B. Is that normal? If so, how can it be explained? (note that the direct effect is positive and significant). Thank you in advance, Margarita (p.s. this blog has been really helpful - so thanks!) |
|
|
These things can happen. See the book by David MacKinnon that is referenced in the user's guide. |
|
|
I took a look at it and I will read more on inconsistent mediation models. Thank you for the reference. I just wanted to make clear, should a negative mediation effect be interpreted like a correlation? e.g. higher values on mediator -> lower values on DV (or vice versa) ? |
|
|
It should be interpreted as a regression coefficient. This depends on the scale of the mediator and the type of regression coefficient. But basically yes. Positive and negative have the same general meaning as for a correlation coefficient. |
|
|
As always, thank you for the prompt reply and all the help! |
|
Margarita posted on Thursday, October 03, 2013 - 3:49 pm
|
|
|
I have another question (sorry for bothering you again..) I was searching for an answer but I could not find one. Could I run a mediation analysis with 1 exogenous latent variable, and 2 endogenous observed variables (one of which acts as a mediator)? - all variables are continuous For example: [LATENT = L, MEDIATOR =M, OUTCOME =O] ANALYSIS: ESTIMATOR = ML; BOOTSTRAP = 1000; MODEL: L BY X1 X2 X3 X4; O ON L M; M ON L; MODEL INDIRECT: O IND L; Does this seem correct? Thank you for your time. |
|
|
This looks correct. |
|
Margarita posted on Friday, October 04, 2013 - 10:36 am
|
|
|
Thank you for your comments! |
|
Back to top |