22  R Packages Used in This Book

This appendix lists the R packages used in the book. The point is practical: for each method, which package does the computation, and which chapters use it?

Most packages live on CRAN. graph, RBGL and Rgraphviz, which pcalg needs, come from Bioconductor and require BiocManager. Eight are GitHub-only: npcausal, medoutcon, synthdid, bartik.weight, ShiftShareSE, ssaggregate, and two of my own, didint and tvhte. MASS, boot, survival and parallel ship with R and need no installation.

Five of those GitHub packages appear only in chunks marked eval: false — bartik.weight, ShiftShareSE and ssaggregate in the shift-share chapter, didint and tvhte in the DiD chapter. The book renders without them. Install them only to run those examples yourself.

22.1 Working With The Environment

A minimal reproducible workflow is:

cd ~/projects/books/causal_econometrics_guide

# 1. Bioconductor dependencies of pcalg (once)
Rscript -e 'install.packages("BiocManager"); BiocManager::install(c("graph","RBGL","Rgraphviz"))'

# 2. Every CRAN package the chapters load
Rscript -e 'install.packages(c(
  "remotes", "BART", "bcf", "broom", "causaldata", "causaldrf",
  "causaleffect", "cobalt", "dagitty", "data.table", "did", "DoubleML",
  "etwfe", "EValue", "fixest", "ggdag", "gmm", "grf", "haven", "hdm",
  "igraph", "ipw", "ivreg", "knitr", "lavaan", "lmtest", "lmtp", "ltmle",
  "marginaleffects", "MatchIt", "mlr3", "mlr3learners", "pcalg",
  "policytree", "ranger", "rdrobust", "riskRegression", "sandwich", "sem",
  "sensemakr", "SuperLearner", "survey", "survminer", "tidyverse", "tmle",
  "WeightIt"))'

# 3. GitHub packages the book needs in order to render
Rscript -e 'remotes::install_github(c("ehkennedy/npcausal", "nhejazi/medoutcon",
  "synth-inference/synthdid"))'

# 4. Optional --- only for the `eval: false` examples
Rscript -e 'remotes::install_github(c("paulgp/bartik-weight", "kolesarm/ShiftShareSE",
  "borusyak/shift-share", "xiangao/didint", "xiangao/tvhte"))'

quarto render

The book uses Quarto’s execute: freeze: auto setting (configured in _quarto.yml), so each chapter’s computed output is cached under _freeze/. Re-rendering only re-executes chunks whose source has changed. When upgrading a package whose API may have moved, delete the relevant _freeze/<chapter>/ directory to force a re-run.

22.2 Package Map

Package Used For Main Book Chapters
dagitty (Textor et al. 2016) DAG and ADMG construction, adjustment-set search, d-separation Identification, DAG workflow, Smoking cessation
ggdag Visual rendering of DAGs/ADMGs built with dagitty Identification, DAG workflow, IV/RDD
causaleffect (Tikka, Hyttinen, and Karvanen 2019) Pearl–Shpitser ID algorithm for general ADMGs DAG workflow
tidyverse (dplyr, tidyr, purrr, ggplot2) Data wrangling and plotting throughout Every chapter
haven Reading Stata .dta files used as examples Estimation, DiD, Shift-share IV
causaldata Built-in NHEFS and other causal-inference example datasets Sensitivity analysis, Smoking cessation, IV/RDD
hdm Pension (pension) dataset and high-dimensional inference Estimation
SuperLearner Stacked machine-learning nuisance estimation Nonparametric
npcausal Influence-function-based ATE/ATT estimators (Edward Kennedy) Nonparametric
tmle Targeted maximum likelihood estimation Nonparametric
DoubleML, mlr3, mlr3learners Double/debiased ML for partially linear and interactive models Nonparametric
did Callaway & Sant’Anna estimators; mpdta example dataset DiD
etwfe Wooldridge’s extended TWFE for staggered DiD DiD
fixest High-dimensional fixed-effects OLS, Poisson, IV; clustered SEs DiD, IV/RDD, Poisson-IV
synthdid Standard DiD, synthetic control, synthetic DiD; Prop 99 example DiD
sem Classical two-stage least squares via tsls() IV/RDD
MASS Multivariate normal sampling for IV simulations IV/RDD
rdrobust Local-polynomial RDD, sharp and fuzzy designs IV/RDD
gmm Generalized method of moments for nonlinear IV Poisson-IV
lavaan SEM and CFA syntax for classical mediation Mediation
medoutcon Causal mediation: natural and interventional (in)direct effects Mediation
pcalg PC, GES, FCI, RFCI causal-discovery algorithms Causal Discovery (both chapters)
igraph Graph plotting for CPDAGs/PAGs; graph objects for causaleffect Causal Discovery, DAG workflow
Rgraphviz, graph Install-time dependencies of pcalg (graph classes; pcalg’s own plot methods) Causal Discovery (indirect)
MatchIt Nearest-neighbour, optimal and full matching Matching
WeightIt Weighting for ATE/ATT, including CBPS and entropy balancing Matching, Continuous treatments
cobalt Covariate-balance tables and love plots Matching
marginaleffects Contrasts and marginal effects after weighting Matching
grf Generalized random forests; causal forests Heterogeneous effects
policytree Optimal policy trees from causal-forest scores Heterogeneous effects
ranger Fast random forests, used as a SuperLearner library Nonparametric
boot Bootstrap standard errors (ships with R) Nonparametric
lmtp Longitudinal modified treatment policies Continuous treatments
causaldrf Ready-made GPS and dose-response estimators Continuous treatments
ltmle Longitudinal TMLE for time-varying treatments G-methods
ipw Inverse-probability weights for time-varying treatments G-methods
survey svydesign/svyglm for the weighted marginal structural model G-methods
survival Cox and Kaplan-Meier fitting (ships with R) Survival
survminer Survival-curve plots Survival
riskRegression Cause-specific and absolute-risk regression Survival
BART Bayesian additive regression trees for the outcome model Bayesian
bcf Bayesian causal forests for heterogeneous effects Bayesian
sensemakr Omitted-variable-bias sensitivity contours Sensitivity analysis
EValue E-values for unmeasured confounding Sensitivity analysis
ivreg Two-stage least squares with diagnostic tests IV/RDD
lmtest, sandwich Robust and clustered covariance matrices, and the tests that use them Estimation, IV/RDD
broom Tidying model objects into data frames for tables and plots Most estimation chapters
data.table Builds the medoutcon example dataset (as.data.table, setnames) Mediation
knitr Table rendering with kable Ten chapters, throughout
parallel parLapply cluster bootstrap of the second-stage SEs (ships with R) Poisson-IV
bartik.weight Rotemberg weights for shift-share designs (GitHub; eval: false) Shift-share IV
ShiftShareSE Adao-Kolesar-Morales standard errors (GitHub; eval: false) Shift-share IV
ssaggregate BHJ shock-level aggregation and inference (GitHub; eval: false) Shift-share IV
didint Difference-in-differences with spatial interference (mine; eval: false) DiD
tvhte Time-varying heterogeneous event-study effects (mine; eval: false) DiD

22.3 dagitty and ggdag

The dagitty package provides a small DSL for graphs and the standard graph-theoretic identification queries:

library(dagitty)

g <- dagitty("dag {
  X -> A
  X -> Y
  A -> Y
}")

adjustmentSets(g, exposure = "A", outcome = "Y")

adjustmentSets returns all minimal sufficient sets for the backdoor criterion. Bidirected edges encoded as A <-> Y represent unobserved common causes (an ADMG), and dagitty correctly returns no adjustment set in that case.

ggdag consumes dagitty objects and renders them through ggplot2:

library(ggdag)
ggdag(g) + theme_dag_blank()

22.4 causaleffect

When dagitty::adjustmentSets returns an empty list, the effect may still be identified through a non-backdoor route (front-door, more general ID-algorithm patterns). causaleffect implements Tikka & Karvanen’s R port of the Pearl–Shpitser ID algorithm:

library(igraph)
library(causaleffect)

g <- graph_from_literal(A -+ M, M -+ Y, A -+ Y, Y -+ A)

# Select the A<->Y pair by endpoint name. igraph's edge ordering is not part of
# its documented contract, so a literal index like c(2, 4) can break silently.
edge_ends  <- ends(g, E(g))
confounded <- (edge_ends[, 1] == "A" & edge_ends[, 2] == "Y") |
              (edge_ends[, 1] == "Y" & edge_ends[, 2] == "A")
g <- set_edge_attr(g, "description", index = which(confounded), value = "U")

causal.effect(y = "Y", x = "A", G = g, simp = TRUE)
# → \sum_{M} P(M|A)\left(\sum_{A} P(Y|A,M) P(A)\right)

The bidirected edge convention is two reciprocal directed edges with description = "U". The function either returns a symbolic identification expression or raises an error indicating the effect is not identifiable.

22.5 etwfe and fixest

fixest handles the regression models with high-dimensional fixed effects. Its feols, fepois, and feglm use a formula syntax that keeps DiD and IV specifications readable:

library(fixest)

feols(y ~ x | id + year, data = df, vcov = ~id)        # TWFE, clustered SE
fepois(y ~ x + offset(log(pop)) | id + year, data = df)  # Poisson with FE
feols(y ~ x | id + year | x_endo ~ z, data = df)         # IV/2SLS

etwfe wraps fixest for Wooldridge’s extended two-way fixed-effects DiD:

library(etwfe)

mod <- etwfe(fml = lemp ~ lpop, tvar = year, gvar = first.treat,
             data = mpdta, vcov = ~countyreal)
emfx(mod, type = "event")

emfx() aggregates the cohort × time interaction coefficients into an overall ATT, event-time effects, or calendar-time effects.

22.6 pcalg

pcalg provides PC, GES, FCI, RFCI, GIES, and LINGAM under one consistent S4 interface. It uses the graph package for graph objects and (for its own plot methods) Rgraphviz; the discovery chapters in this book plot with igraph instead. The Bioconductor dependencies must be installed via BiocManager before pcalg can be installed from CRAN.

library(pcalg)

# Constraint-based: PC algorithm
pc_fit <- pc(suffStat = list(C = cor(data), n = nrow(data)),
             indepTest = gaussCItest, labels = colnames(data),
             alpha = 0.01)

# Score-based: Greedy Equivalence Search
ges_fit <- ges(new("GaussL0penObsScore", data))

# Latent-variable case: FCI / RFCI
fci_fit  <- fci(suffStat = list(C = cor(data), n = nrow(data)),
                indepTest = gaussCItest, labels = colnames(data), alpha = 0.01)
rfci_fit <- rfci(suffStat = list(C = cor(data), n = nrow(data)),
                 indepTest = gaussCItest, labels = colnames(data), alpha = 0.01)

Both observed (PC, GES) and latent (FCI, RFCI) chapters use these as the primary algorithms.

22.7 lavaan and medoutcon

lavaan ports the SEM model-string syntax familiar from EQS, Mplus, and LISREL into R. It is used in the mediation chapter for classical SEM mediation:

library(lavaan)

model <- "
  m ~ a * x
  y ~ b * m + c * x
  indirect := a * b
  total    := c + indirect
"
fit <- sem(model, data = df)
parameterEstimates(fit)

For causal mediation under the potential-outcomes framework, the book uses medoutcon (Hejazi & van der Laan), which estimates natural and interventional (in)direct effects with cross-fitted nuisance estimators. The interventional effects are the ones to reach for when a treatment-induced confounder of the mediator is present, since the natural effects are not identified in that case.

22.8 Practical Advice

The packages above cover the examples in this book. A few useful packages outside the main text are:

  • MatchIt and WeightIt for matching and weighting estimators of the ATE/ATT
  • grf (Generalized Random Forests) for heterogeneous treatment effects and instrumental-forest IV
  • bnlearn for an alternative causal-discovery toolkit focused on Bayesian networks
  • lavaan.survey and blavaan for survey-weighted and Bayesian SEM
  • mediation for the classical Imai/Keele/Tingley mediation framework

When upgrading any of these packages, re-render affected chapters after deleting the relevant _freeze/<chapter>/ directory so that Quarto does not reuse stale cached results.