User-written KHB command: How do I Interpret Indirect Effects and Determine Significance?

[Nicholas Smith]

Hello everyone,

I'm using the user-written KHB command in Stata (https://www.stata-journal.com/sjpdf....iclenum=st0236) to conduct mediation analyses using a binary DV, continuous IV, and three continuous mediator variables. The code works; however, I'm having a bit of trouble understanding the "components of difference" table. Here are my specific questions:

1. I understand that in the components of difference table, the total indirect effect is being decomposed, and, as you can see, one coefficient is positive, the others are negative. I'm assuming the negative coefficient and the negative "percent reduced" is indicating a suppressor effect of mastery_std (if this is wrong, please tell me)? If I am correct, what exactly does this mean?

2. How do I determine whether the mediating or suppressor effect for each individual mediator is statistically significant? Is there something that I can put into the code that will spit out a p-value, conf.interval, etc?

3. In the paper I'm writing, I discuss the contribution of each individual mediator for the indirect effect, which I can get from the components of difference table; however, because there seems to be a suppressor effect of mastery_std, reporting the "percent reduced" of the other two mediators (i.e., stress_std and esteem_std) would be incorrect because they add up to over 100%. So, for example, in the paper, it wouldn't make sense to write something like, "...When broken down by individual mediator, esteem_std and stress_std account for 18% and 86% of the association between X and Y, respectively." To determine the contribution of each individual mediator in this case, should I just examine each mediator individually and treat the other mediators as control variables?

Thank so much for your help!

My code:

khb logit anxiety discrimination_std || c.esteem_std mastery_std stress_std ///
[pw=WGTCENT] if caribbean==1, concomitant(i.female c.age_15 ///
c.income_3 i.nativity) summary disentangle or

[Phil Bromiley]

You'll increase your chances of a helpful answer if you follow the FAQ on asking questions - provide Stata code in code delimiters, readable Stata output, and sample data using dataex.

In Stata 15, the GSEM routines do mediation. With a user written procedures like khb, you may need to contact the author.

[daniel klein]

1. I'm assuming the negative coefficient and the negative "percent reduced" is indicating a suppressor effect of mastery_std (if this is wrong, please tell me)? If I am correct, what exactly does this mean?

You are correct. I do not understand the follow up question. What such a suppressor effect might mean, substantially, is a question that you should be able to answer best. We do not know your field of research, research question, your measures, etc.

2. How do I determine whether the mediating or suppressor effect for each individual mediator is statistically significant? Is there something that I can put into the code that will spit out a p-value, conf.interval, etc?

khb reports coefficients and standard errors. From this, it is easy to compute a test statistic and determine a p-value and confidence intervals. Although you will have to extract coefficients and standard errors from e(disentangle), not e(b) and e(V), Buis (2007) discusses this problem.

3. [...] however, because there seems to be a suppressor effect [...], reporting the "percent reduced" of the other two mediators [...] would be incorrect because they add up to over 100%.

I do not see anything incorrect about this; I agree that it is counter-intuitive. However, I do not necessarily think of this as a technical problem but a substantive one, which follows directly from your question in 1. My answer is, therefore, the same: step back from software issues and try to make sense of the suppressor effect, first. Once you have a clear idea about what is driving these results, you should be able to communicate these ideas to others.

Best
Daniel


Buis, M. L. 2007. Stata tip 53: Where did my p-values go? The Stata Journal, 7(4), pp. 584-586.

[Lydia Palumbo]

I have a follow up question on this old post, point 2. I am now following tip #53 by M.Buis.
I would like to check if f I understand correctly. The distribution that we need to consider is the Z distribution and the code would be similar to this:

local z=e(disentangle)[10,1]/e(disentangle)[10,2]
di 2*normal(-abs(`z'))

Is it correct? Thank you and best, Lydia