Skip to contents

dma

pkgdown

Distributional mediation analysis using energy regression.

Overview

dma implements causal mediation analysis with engression for nuisance estimation. The point of using a distributional regression method is that the counterfactual outcome distribution can change in ways that are not summarized by a conditional mean. The package supports four effect decompositions:

  • Natural (N): Direct and indirect effects
  • Organic (O): Organic direct and indirect effects
  • Randomized Interventional (RI): RIDE and RIIE
  • Recanting Twins (RT): Four-way path decomposition

The estimands follow Liu, Williams, Rudolph, and Diaz (2024). Here engression is used for both the outcome regressions and the Riesz-representer density-ratio steps.

Installation 

# Install from GitHub
devtools::install_github("xiangao/dma")

Quick Start 

library(dmaR)

result <- dma(
  data = df,
  trt = "A",
  outcome = "Y",
  mediators = "M",
  covar = "W",
  effect = "N",
  d0 = \(data, trt) 0,
  d1 = \(data, trt) 1
)

print(result)
tidy(result)
plot(result)

# Counterfactual density plots
plot_counterfactual_density(result)                        # marginal P(Y|do(A=a))
plot_counterfactual_density(result, use_weights = TRUE)    # all mediation regimes

What is included

  • Distributional outcome regression via engression (learns full P(Y|X), not just E[Y|X])
  • Neural network Riesz representer estimation for density ratios
  • Cross-fitting with parallel fold processing via future.apply
  • Observation weights propagated through all nuisance estimation stages
  • Coefficient plots and counterfactual density visualization
  • Weighted counterfactual density estimation for the mediation regimes

Vignettes

Full documentation: https://xiangao.github.io/dma/

Vignette Description
Distributional mediation analysis Comparison with crumble on natural and organic effects, oracle vs estimated distributions, Monte Carlo study
Distributional mediation Non-linear DGP demonstrating distributional mediation with weighted counterfactual densities
Effect types All four effect decompositions (N, O, RI, RT) with oracle potential outcome distributions
Advantages of engression When distributional regression outperforms conditional mean methods