[Mediation-information] moderated mediation analysis using 'mediation' package

[Liviu Andronic]

Hello


On Thu, Mar 22, 2012 at 7:47 PM, Teppei Yamamoto <teppei at mit.edu> wrote:
> One thing you could do is to set the moderator to different values along its
> range (running a separate mediate call for each of those values) and plot
> the estimated mediation effects against those moderator values. That way,
> you will see how the indirect effect (or direct effect if that's what you
> are more interested in) varies as a function of the moderator, which I think
> is exactly the goal of moderated mediation analysis.
>
Thank you again for all the explanations.

I managed to estimate the mediation effect over the range of the
mediator (although it's quite CPU-consuming)
#estimate mediation effects
med_c <- lapply(quantile(modmed$ENGAGESP), function(x){
    mediate(mod.m2, mod.y2, dropobs=F, treat='EVPUBLIC', mediator='TRUST',
    covariates=list(ENGAGESP=x),  boot=T, sims=50)
})


and then plotting the resulting objects
#plot mediation effects
par(mfrow=c(2,3))
lapply(1:length(med_c), function(x) plot(med_c[[x]], main=names(med_c[x])))
par(mfrow=c(1,1))


and subsequently plot the estimated mediation effects against the
moderator values.
#plot the estimated mediation effects against the moderator values
(xa <- ldply(1:length(med_c), function(x)
data.frame(med_c[[x]][c('d0', 'z0')])))
plot(quantile(modmed$ENGAGESP), xa[,'d0'])
plot(quantile(modmed$ENGAGESP), xa[,'z0'])


This wasn't immediately straightforward, however, and required some
vectorizing magic. Could 'mediation' provide some functions or methods
that would help automating some of the procedures above?

Regards
Liviu