Mediation question using brms

Interfaces brms
#1
  • Operating System: Mac OSX
  • brms Version: 2.30

I’m attempting to do a multilevel mediation analysis. The dv is a likert scale response which I am modeling with a cumulative distribution (logit link). The mediator is categorical (Yes / No). When calculating the ab path, I am wondering if I can multiply the a and b coefficients directly (as below).

Knowledge_Judgment_mod <- bf(KnowledgeResponse ~ Condition + CounterfactualResponse + (1|Subject), family = “cumulative”)

Counterfactual_mod <- bf(CounterfactualResponse ~ Condition + (1|Subject), family = “bernoulli”)

model1 <- brm(Knowledge_Judgment_mod +
Counterfactual_mod,
set_rescor(FALSE),
data=data1,
iter=3000,
warmup=1500,
control = list(adapt_delta = .90),
prior = c(set_prior(“normal(0,1)”, class = “b”)))

a <- posterior_samples(model1, pars = “CounterfactualResponse_ConditionSpecific”)
b <- posterior_samples(model1, pars = “KnowledgeResponse_CounterfactualResponse”)
a_b <- a*b

summary(a_b)

CI_Model1_Mediation <- quantile(a_b$b_CounterfactualResponse_ConditionSpecific, probs = c(.025,.975))

CI_Model1_Mediation

1 Like
How to execute Two-Group Moderated Multilevel Mediation in brms (1-1-1 with level 1 moderator)?
#2

That approach seems to be right from what I see.

#3

Maybe this is of interest for you: I have drafted a function, mediation(), that prints a compact summary especially for mediation analysis (similar to the output from the mediation package).

You can see examples in a vignette here:
https://strengejacke.github.io/sjstats/articles/bayesian-statistics.html#summary-of-mediation-analysis

There’s a comparison of brms to the mediation package, and the results look good!

You need the current dev-version of sjstats to check out the mediation()-function.

3 Likes
Output interpretation: zero-inflated mediator and binary outcome
#4

I have a short follow-up question here -

If the mediator is normally distributed, and the outcome is negative binomial distributed, does taking the product of the coefficients still work for calculating the indirect effects? Thanks!

#5

Don’t think so because different links but correct me if I’m wrong

1 Like
#6

Ah, gotcha, so in your original example, it should work because the logit link applied to both. Makes sense. Any chance you know of a conversion for calculating the indirect effects in my case? Thanks again.

#7

I’d look up default link of neg binom in brms. That should help. I’m not at a computer and don’t remember off hand

#8

Hi - that link is broken now… do you still have this code available?

#9

Yes, the function is now re-implemented in the bayestestR-package, the related vignette is here.

1 Like
#10

Hi! Thanks a lot for providing this function! I have two questions related to it

  1. In my mediation model, the mediator is normally distributed and the dependent variable can be considered cumulative.
    bfy ~bf(y~x+m+(1+x+m|ID|Subject), family=“cumulative”)
    bfm ~bf(m~x+(1+x|ID|Subject), family=“gaussian”)
    As these families do not have the same link (cumulative having a logit-link) - can I still use the mediation()-function to calculate the effects (especially the indirect effect)? And if not, do you know a workaround for that?
  2. Is it possible to include covariates into the model als illustrated below? (Please note that I am not including the interaction terms with the covariate as random effect as well as this does not add much to the predictive capacity of the model)
    bfy ~bf(y~xcov1+mcov1+(1+x+m+cov1|ID|Subject), family=“cumulative”)
    bfm ~bf(m~x+(1+x|ID|Subject), family=“gaussian”)
    Thanks a lot for your help!