Why is output of R mediation package showing control/treat groups when the predictor is continuous?

Question

I am running a mediation model using the R mediation package, but I am not getting the correct output for my variable types. I have a continuous predictor, but the output is treating my predictor as a categorical variable.

In short:

Predictor = cognitive test score [Composite_Z] (continuous)

Mediator = self-awareness [Awareness] (dichotomous; variable type = numerical in order to run mediation)

Outcome = driving frequency [DRFRQ] (dichotomous)

10 Covariates = Age (continuous), Sex (dichotomous), Race (dichotomous), Education (dichotomous), Severity (continuous), Time (continuous), Seizures (dichotomous), Income (ordinal), Motor (continuous), UrbanRural (ordinal)

I have two models that are going into my mediation analysis (both of these run fine without issues)--

Model #1 : binary logistic regression examining the relationship between cognition (predictor) and self-awareness, while accounting for the covariates.

R code:

    fit.a.A3H1 <- glm(Awareness ~ Composite_Z + Age + Sex + Race +
 Education + Severity + Time + Seizures + Income + Motor 
+ UrbanRural, family=binomial(link="logit"), data=A3H1.df)

Model #2 : binary logistic regression examining the relationship between cognition (predictor) and driving frequency (outcome), while accounting for self-awareness (mediator) and the covariates.

R code:

    fit.total.A3H1 <- glm(DRFRQ ~ Composite_Z + Awareness + Age + Sex 
+ Race + Education + Severity + Time + Seizures + Income 
+ Motor + UrbanRural, family=binomial(link="logit"), data = A3H1.df)

Mediation model : Then, I put those models (outlined above) into my mediation analysis. I am using the R mediation package to run the analysis.

R code:

fitMed.A3H1 <- mediate(fit.a.A3H1, fit.total.A3H1, sims=1000, boot=FALSE, 
                       treat="Composite_Z", mediator="Awareness",
                       covariates = NULL, use_speed = FALSE)

Output:

Causal Mediation Analysis 

Quasi-Bayesian Confidence Intervals

                          Estimate 95% CI Lower 95% CI Upper p-value  
ACME (control)           -0.001007    -0.004191         0.00    0.27  
ACME (treated)           -0.000686    -0.002933         0.00    0.27  
ADE (control)             0.031882     0.003696         0.06    0.03 *
ADE (treated)             0.032203     0.003727         0.06    0.03 *
Total Effect              0.031196     0.004744         0.05    0.03 *
Prop. Mediated (control) -0.024010    -0.134304         0.08    0.25  
Prop. Mediated (treated) -0.016399    -0.113058         0.07    0.25  
ACME (average)           -0.000846    -0.003507         0.00    0.27  
ADE (average)             0.032042     0.003711         0.06    0.03 *
Prop. Mediated (average) -0.020204    -0.124782         0.08    0.25  

---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Sample Size Used: 350 

Simulations: 1000 

The output gives the results in groups, even though my predictor is a continuous variable. Does anyone know why I'm running into this problem and ways I could fix it? Any suggestions would be greatly appreciated!

What's also strange is that I do not have this problem when I run a separate mediation analysis with the same predictor/mediator/covariates (but a different outcome variable that is continuous instead of dichotomous). This separate mediation analysis is made up of a (1) binary logistic regression model and (2) linear regression model, and the output shows only the overall ACME/ADE/etc (ie, there are not groups).

Answer 1

It looks like the output from the mediate() function takes the form of whatever the final DV is: ie, if you have a binary DV, the outcome will be seperated into control and treated, whereas if you have a continuous DV, the output will not (as the output is on a numeric not a nominal scale, thus representing the β).

Using code from the mediate package vignette , please see the following example:

library(mediation)
data("framing", package = "mediation")
set.seed(2014)

#1: models exactly from mediation vignette: continuous m, binary DV 
med.fit <- lm(emo ~ treat + age + educ + gender + income, data = framing)
out.fit <- glm(cong_mesg ~ emo + treat + age + educ + gender + income, data = framing, family = binomial("probit"))
med.out <- mediate(med.fit, out.fit, treat = "treat", mediator = "emo", robustSE = TRUE, sims = 100)
summary(med.out)

#2: models swapped: binary DV, continuous m
med.fit <- glm(cong_mesg ~ treat + age + educ + gender + income, data = framing, family = binomial("probit"))
out.fit <- lm(emo ~ cong_mesg + treat + age + educ + gender + income, data = framing)
med.out <- mediate(med.fit, out.fit, treat = "treat", mediator = "emo", robustSE = TRUE, sims = 100)
summary(med.out)

As you can see #1 has a binary DV and continuous M, and output from mediate() is seperated into control and treated, whereas output from 2 has continuous DV and binary M, and output is not.

Answer 2

In case of a continuous treatment variable you will have to define the control group and the treatment group when calling mediate (here x and y ):

fitMed.A3H1 <- mediate(fit.a.A3H1, fit.total.A3H1, sims=1000, boot=FALSE, 
                   treat="Composite_Z", mediator="Awareness",
                   control.value = x, treat.value = y,
                   covariates = NULL, use_speed = FALSE)

I know it might not result with the effect you might be expecting but the package was built thinking of an experiment with a limited number of treatment options. (See section 3.3 in the vignette )