0
$\begingroup$

I am trying to hand calculate the standard error for the a path in a moderated mediation model where the moderator modifies the a path. For example:

Mediator = X + Moderator + X*M

I gather that the a path would be calculated like so (with coefficients from the above model):

X + (X*M * Moderator)

How would I calculate the standard error in this case? I am not having any luck finding the answer so far.

Update

In response to the request for more information, I am updating the question.

Say I think that the coping mediates the relationship between gender and depression but that this mediation effect depends on age. I run 2 regressions:

Outcome = Coping

                  Estimate Std. Error t value Pr(>|t|)   
(Intercept)       4.2063     0.1529  27.513  < 2e-16 ***
Age               0.6621     0.2739   2.417  0.01714 *  
Gender            0.6853     0.2189   3.131  0.00218 ** 
Age*Gender       -0.8737     0.3726  -2.345  0.02065 *  

Outcome = Depression, with Coping

                  Estimate Std. Error t value Pr(>|t|)    
(Intercept)      5.19472    0.31975  16.246  < 2e-16 ***
Coping          -0.59130    0.07054  -8.382  1.1e-13 ***
Age             -0.37615    0.21850  -1.721 0.087717 .  
Gender          -0.66654    0.17728  -3.760 0.000263 ***
Age*Gender       0.20956    0.29682   0.706 0.481535 

I believe that I would need to calculate the indirect and total effects (conditional on Age = 45) with the following:

indirect = 0.6853*-0.59130 + 45*-0.8737*-0.59130
total = indirect + -0.66654 + 45*0.20956

How would I calculate the SE for the (a) path to Coping through Age*Gender and/or how would I calculate the SE for these conditional indirect and total effects?

$\endgroup$
  • $\begingroup$ This is unclear. There isn't enough information here for this to be answerable. $\endgroup$ – gung May 23 '17 at 16:23
1
$\begingroup$

You do not calculate a standard error in this case. When you multiply coefficients together (i.e., an indirect effect), the product is not normally distributed. What we generally do instead is get a bootstrapped 95% confidence interval around the product.

I have R code, sample R data, and a .pdf guide here that implements moderation, mediation, and moderated mediation. This code implements the equations discussed in Hayes's (2015) paper proposing the index of moderated mediation.

$\endgroup$
  • $\begingroup$ That is essentially what I need---some way of testing the conditional indirect effect. I have used Hayes' macros before, and it is great that you have adapted them for R. However, I am running robust regressions (lmrob). I have not found any packages for moderated mediation that work with lmrob objects. Any ideas? $\endgroup$ – user162454 May 23 '17 at 21:05
  • $\begingroup$ This will be computationally intensive, but you could do the bootstrapping manually. You basically sample a random list of rows n from your data set (with replacement), do the analyses, extract the index of moderated mediation and/or indirect effects, and save them into an object. Do this, say, 5000 times in a for loop. Then get the mean and 1.96*SD of your indirect effects. And you have a bootstrap confidence interval. I have done this before, but don't have the code on me currently. $\endgroup$ – Mark White May 23 '17 at 21:17
  • $\begingroup$ @user162454 actually, there is a much easier way than I just said above! The code I linked to uses the lavaan package. The lavaan package has options for various robust standard errors: first order derivatives, Huber-White, Satorra-Bentler, etc. You can simply add to the code in the sem() function a specific se or estimator argument. Search this documentation for "robust": cran.r-project.org/web/packages/lavaan/lavaan.pdf $\endgroup$ – Mark White May 24 '17 at 0:23

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.