sdidevent

Event-Study extension of Synthetic Difference-in-Differences for R.

This is an R port of the Stata sdid_event package by Diego Ciccia, Damian Clarke, and Daniel Pailañir, which implements the event-study decomposition of Ciccia, Clarke, Pailañir (2024). The underlying SDID estimator is Arkhangelsky, Athey, Hirshberg, Imbens, Wager (2021, AER); this package delegates the estimator to the synthdid R package and implements only the event-study decomposition on top.

Install

# install.packages("remotes")
remotes::install_github("awhobbs/sdidevent")

The synthdid dependency is on GitHub (not CRAN):

remotes::install_github("synth-inference/synthdid")

Quick example

library(sdidevent)
library(synthdid)

data(california_prop99)
setup <- panel.matrices(california_prop99)

est <- sdid_event(setup$Y, setup$N0, setup$T0,
                  vce = "bootstrap", breps = 200, seed = 1)

print(est)
plot(est)

Output (truncated):

Event-Study Synthetic Difference-in-Differences
  N0 = 38 controls, N1 = 1 treated
  T0 = 19 pre periods, T1 = 12 post periods
  Inference: bootstrap (B = 200, ci_method = normal)
  Aggregate ATT: -15.6038  (SE = ~9.9)

Inference options

vce	Default?	What it does
"bootstrap"	Yes	Cluster bootstrap (matches Stata default).
"placebo"		Random treatment permutation among never-treated clusters.
"jackknife"		Leave-one-out. Unit-level by default; cluster-level if cluster is supplied.
"off"		No inference; only point estimates.

For state-level treatment assignment, pass cluster = state_vec (a length-N vector aligned with the rows of Y). This activates cluster bootstrap or cluster jackknife.

Staggered adoption

sdid_event_panel() auto-detects staggered adoption (multiple distinct first-treated periods) from the long-form treatment indicator:

library(sdidevent)
# `df` has columns unit, year, y, treat where `treat` switches on at
# different years for different units.
est <- sdid_event_panel(df, outcome = "y", unit = "unit", time = "year",
                        treatment = "treat",
                        vce = "bootstrap", breps = 500, seed = 1)
print(est)
print(est$cohort_summary)   # per-cohort N_tr, N_post, cohort_att

For matrix entry, pass a length-N integer vector for T0 giving each unit's adoption period (NA for never-treated controls):

# 12 never-treated controls, 3 units adopting at t = 6, 3 at t = 10.
cohort_of <- c(rep(NA, 12), rep(6, 3), rep(10, 3))
est <- sdid_event(Y, N0 = 12, T0 = cohort_of,
                  vce = "bootstrap", breps = 500, seed = 1)

The per-cohort decomposition follows Ciccia, Clarke, Pailañir (2024) equations 3, 4, 6, 7 and matches Stata sdid_event to ~1e-6 on point ATTs; see inst/extdata/staggered_stata_golden.csv for the reference output.

What's added vs Stata

• Cluster-jackknife with the Cameron-Gelbach-Miller (2008) small-cluster SE and $t(G-1)$ reference distribution. Not in the Stata package.
• Standard R conventions: coef(), vcov(), plot(), summary() S3 methods. ggplot2-based event-study chart.
• keep_draws = TRUE exposes the raw bootstrap matrix for downstream inference (e.g., joint CIs).

Citation

If you use this package, please cite both the methodology and the port:

@article{ciccia2024short,
  title   = {A Short Note on Event-Study Synthetic Difference-in-Differences Estimators},
  author  = {Ciccia, Diego and Clarke, Damian and Pailañir, Daniel},
  journal = {arXiv preprint arXiv:2407.09565},
  year    = {2024}
}

@article{arkhangelsky2021synthetic,
  title   = {Synthetic Difference-in-Differences},
  author  = {Arkhangelsky, Dmitry and Athey, Susan and Hirshberg, David A and
             Imbens, Guido W and Wager, Stefan},
  journal = {American Economic Review},
  volume  = {111},
  number  = {12},
  pages   = {4088--4118},
  year    = {2021}
}

@manual{hobbs2026sdidevent,
  title  = {sdidevent: Event-Study Synthetic Difference-in-Differences},
  author = {Hobbs, Andrew},
  year   = {2026},
  note   = {R package version 0.2.0}
}

License

GPL-3. See LICENSE.md.