Causal Mediation Analysis

Example Question

This article uses data_diabetes_mediation, a teaching dataset based on a diabetes risk profile. The example question is:

Does plasma glucose explain part of the association between obesity and diabetes?

library(gtregression)
library(dplyr)

data("data_diabetes_mediation", package = "gtregression")

dissect(data_diabetes_mediation)

Logistic Outcome

For a binary outcome, use outcome_approach = logit. The effects are reported as predicted probability differences, which are often easier to explain than odds ratios in a mediation table.

For final analyses, use a larger number of bootstrap simulations such as sims = 500 or sims = 1000. The article uses a smaller value to keep the example quick to run.

diabetes_med <- mediation_analysis(
  data = data_diabetes_mediation,
  exposure = obesity,
  mediator = glucose,
  outcome = diabetes,
  covariates = c(age, blood_pressure, pregnancies, diabetes_pedigree),
  outcome_approach = logit,
  sims = 100,
  seed = 123
)

diabetes_med

## a flextable object.
## col_keys: `Effect`, `Estimate`, `95% CI`, `p-value`, `Interpretation` 
## header has 1 row(s) 
## body has 4 row(s) 
## original dataset sample: 
## 'data.frame':    4 obs. of  5 variables:
##  $ Effect        : chr  "Total effect" "Direct effect" "Indirect effect" "Proportion mediated"
##  $ Estimate      : chr  "0.226" "0.176" "0.051" "0.224"
##  $ 95% CI        : chr  "0.163 to 0.291" "0.114 to 0.235" "0.019 to 0.087" "0.087 to 0.382"
##  $ p-value       : chr  "<0.001" "<0.001" "<0.001" "<0.001"
##  $ Interpretation: chr  "Overall exposure-outcome association" "Association not through the mediator" "Association through the mediator" "Share of total effect through the mediator"
##  - attr(*, "variable_labels")= Named chr [1:7] "Obesity" "Plasma glucose" "Diabetes" "Age" ...
##   ..- attr(*, "names")= chr [1:7] "obesity" "glucose" "diabetes" "age" ...

The returned object keeps the table body, fitted models, bootstrap draws, and exposure comparison values available for checking.

diabetes_med$table_body

##                effect              Effect   estimate   conf.low conf.high
## total           total        Total effect 0.22624186 0.16258768 0.2907013
## direct         direct       Direct effect 0.17561583 0.11427535 0.2347698
## indirect     indirect     Indirect effect 0.05062603 0.01899140 0.0870712
## proportion proportion Proportion mediated 0.22376950 0.08741395 0.3821232
##            p.value                             Interpretation
## total            0       Overall exposure-outcome association
## direct           0       Association not through the mediator
## indirect         0           Association through the mediator
## proportion       0 Share of total effect through the mediator

diabetes_med$values

## $reference_value
## [1] "No"
## 
## $exposure_value
## [1] "Yes"

diabetes_med$models$mediator

## 
## Call:
## stats::lm(formula = .mediation_formula(mediator, c(exposure, 
##     covariates)), data = df)
## 
## Coefficients:
##       (Intercept)         obesityYes                age     blood_pressure  
##          98.23467            6.66535            0.39847            0.06945  
##       pregnancies  diabetes_pedigree  
##           0.40874            5.56161

diabetes_med$models$outcome

## 
## Call:  stats::glm(formula = f, family = stats::binomial(), data = df)
## 
## Coefficients:
##       (Intercept)         obesityYes            glucose                age  
##         -6.848643           0.952467           0.034610           0.012419  
##    blood_pressure        pregnancies  diabetes_pedigree  
##          0.003185           0.052420           0.590009  
## 
## Degrees of Freedom: 726 Total (i.e. Null);  720 Residual
## Null Deviance:       953.8 
## Residual Deviance: 731.1     AIC: 745.1

head(diabetes_med$boot)

##       total    direct   indirect proportion
## 1 0.2299255 0.1705385 0.05938700  0.2582880
## 2 0.2550856 0.1917216 0.06336401  0.2484029
## 3 0.2126707 0.1747508 0.03791990  0.1783034
## 4 0.2907202 0.2280457 0.06267444  0.2155834
## 5 0.2906804 0.2387094 0.05197098  0.1787908
## 6 0.2486133 0.1939385 0.05467481  0.2199191

Path Diagram

plot_mediation() draws the exposure, mediator, outcome, and the direct and indirect paths.

plot_mediation(diabetes_med)

If the figure is being used only to explain the causal structure, hide the estimates.

plot_mediation(diabetes_med, show_estimates = FALSE)

Quoted Names

Quoted column names and stored character vectors work too. This is useful inside scripts, functions, and Shiny-style workflows.

exposure_var <- "obesity"
mediator_var <- "glucose"
outcome_var <- "diabetes"
covariate_vars <- c(
  "age", "blood_pressure", "pregnancies", "diabetes_pedigree"
)

med_quoted <- mediation_analysis(
  data = data_diabetes_mediation,
  exposure = exposure_var,
  mediator = mediator_var,
  outcome = outcome_var,
  covariates = covariate_vars,
  outcome_approach = "logit",
  sims = 100,
  seed = 456
)

med_quoted

## a flextable object.
## col_keys: `Effect`, `Estimate`, `95% CI`, `p-value`, `Interpretation` 
## header has 1 row(s) 
## body has 4 row(s) 
## original dataset sample: 
## 'data.frame':    4 obs. of  5 variables:
##  $ Effect        : chr  "Total effect" "Direct effect" "Indirect effect" "Proportion mediated"
##  $ Estimate      : chr  "0.226" "0.176" "0.051" "0.224"
##  $ 95% CI        : chr  "0.155 to 0.305" "0.107 to 0.241" "0.019 to 0.094" "0.108 to 0.376"
##  $ p-value       : chr  "<0.001" "<0.001" "<0.001" "<0.001"
##  $ Interpretation: chr  "Overall exposure-outcome association" "Association not through the mediator" "Association through the mediator" "Share of total effect through the mediator"
##  - attr(*, "variable_labels")= Named chr [1:7] "Obesity" "Plasma glucose" "Diabetes" "Age" ...
##   ..- attr(*, "names")= chr [1:7] "obesity" "glucose" "diabetes" "age" ...

Linear Outcome

For a continuous outcome, use outcome_approach = linear. In this example, the outcome is body mass index, so the effects are reported as mean differences.

med_linear <- mediation_analysis(
  data = data_diabetes_mediation,
  exposure = obesity,
  mediator = glucose,
  outcome = bmi,
  covariates = c(age, blood_pressure, pregnancies, diabetes_pedigree),
  outcome_approach = linear,
  sims = 100,
  seed = 789
)

med_linear

## a flextable object.
## col_keys: `Effect`, `Estimate`, `95% CI`, `p-value`, `Interpretation` 
## header has 1 row(s) 
## body has 4 row(s) 
## original dataset sample: 
## 'data.frame':    4 obs. of  5 variables:
##  $ Effect        : chr  "Total effect" "Direct effect" "Indirect effect" "Proportion mediated"
##  $ Estimate      : chr  "10.642" "10.487" "0.155" "0.015"
##  $ 95% CI        : chr  "10.107 to 11.197" "9.943 to 11.078" "0.046 to 0.300" "0.004 to 0.028"
##  $ p-value       : chr  "<0.001" "<0.001" "<0.001" "<0.001"
##  $ Interpretation: chr  "Overall exposure-outcome association" "Association not through the mediator" "Association through the mediator" "Share of total effect through the mediator"
##  - attr(*, "variable_labels")= Named chr [1:7] "Obesity" "Plasma glucose" "Body mass index" "Age" ...
##   ..- attr(*, "names")= chr [1:7] "obesity" "glucose" "bmi" "age" ...

plot_mediation(med_linear)

How To Report

A compact reporting sentence might look like this:

In this teaching analysis, plasma glucose explained part of the model-based obesity-diabetes association. Effects were estimated on the predicted probability difference scale using logistic outcome models and bootstrap confidence intervals.

The table footnote records the exposure comparison, mediator, outcome, bootstrap replicates, and adjustment variables so readers can see what was estimated.

What Not To Claim

Mediation estimates should not be treated as automatic causal proof. A cautious analysis should consider:

• whether the exposure clearly precedes the mediator;
• whether the mediator clearly precedes the outcome;
• whether exposure-mediator, mediator-outcome, and exposure-outcome confounding have been handled;
• whether post-exposure confounders are present;
• whether the model forms are plausible;
• whether a DAG or subject-matter argument supports the causal interpretation.

Use the table and plot to support interpretation after that thinking has been done.