In mattblackwell/DirectEffects: Estimating Controlled Direct Effects for Explaining Causal Findings

DirectEffects

Overview

DirectEffects is an R package to estimate controlled direct effects (CDEs), which are the effect of a treatment fixing a set of downstream mediators to particular values. As of now, the package supports general doubly robust estimation of CDEs under both selection-on-observables and difference-in-differences assumptions. sequential g-estimation and a two-stage matching approach called telescope matching. For more information on how CDEs can be useful for applied research and a brief introduction to sequential g-estimation, see Acharya, Blackwell, and Sen (2016). For more on the telescope matching procedure, see Blackwell and Strezhnev (2022). For more on the difference-in-differences approach, see Blackwell, Glynn, Hilbig, and Phillips (2024).

Installation

You can install DirectEffects via CRAN for the current stable version or via GitHub for the development version.

# Installing from CRAN
install.packages("DirectEffects")

# Installing development version from Github:
# install.packages("devtools")
devtools::install_github("mattblackwell/DirectEffects", build_vignettes = TRUE)

Usage

DirectEffects uses a modular approach to specifying the models and estimators used to estimate the average controlled direct effect (ACDE). You can specify a propensity score and outcome regression model for each treatment variable (that is, each causal variable of interest), along with a set of parametric and machine learning estimators for these nuisance functions. The package has a workflow for models that assume the selection-on-observables and for models that utilize a difference-in-differences approach.

library(DirectEffects)

data(jobcorps)
my_aipw <- cde_aipw() |>
  set_treatment(treat, ~ female + age_cat) |>
  treat_model(engine = "logit") |>
  outreg_model(engine = "lm") |>
  set_treatment(work2year2q, ~ emplq4 + pemplq4) |>
  treat_model(engine = "logit") |>
  outreg_model(engine = "lm") |>
  estimate(exhealth30 ~ treat + work2year2q, data = jobcorps)

broom::tidy(my_aipw)

Specific estimator implementations

DirectEffects also implements a number of specific estimators outside the modular framework. These are often specific estimators proposed in the literature:

• sequential_g(): estimate controlled direct effects using two-stage linear models.
• telescope_match(): estimated controlled direct effects using a two-stage matching procedure with bias correction.

DirectEffects also provides diagnostics for these two approaches, including sensitivity analyses and balance checks.

mattblackwell/DirectEffects documentation built on Dec. 16, 2024, 6:14 p.m.