Synthetic Control Methods

The synthetic control method enables researchers to estimate causal effects by constructing a synthetic version of a treatment unit through a weighted combination of control units. It is implemented as both an R package and a Stata routine, based on the methods developed in the following papers Abadie, Diamond, and Hainmueller (2010), Abadie, Diamond, and Hainmueller (2015), and Abadie, Diamond, and Hainmueller (2011). This work was awarded the Gosnell Prize for Excellence in Political Methodology.

→ Read the explainer — a self-contained tutorial on the synthetic control method for R and Stata users, with the canonical Proposition 99 example worked end-to-end.

Source on GitHub: j-hai/Synth (R package) · j-hai/synth-stata (Stata routine).

What’s new in Synth 1.2-0 (May 2026)

The R package was updated in May 2026 with a substantial set of user-facing additions:

• Built-in inference. synth_inference() returns split-conformal (Chernozhukov–Wuthrich–Zhu 2021) or parametric prediction intervals around the synthetic counterfactual. generate_placebos() plus mspe_test() give the canonical Abadie–Diamond–Hainmueller (2010) placebo p-value in one call.
• Ergonomic data prep. synth_data() is a one-line wrapper around dataprep() for the common case (panel + treated unit + treatment date + auto-controls).
• Alternative QP backends. Optional quadopt = "cvxr" (CVXR + ECOS) and quadopt = "torch" (Frank-Wolfe simplex LS via the torch package, with CPU/CUDA/MPS support). Both live in Suggests: — no required dependency.
• ggplot2 support. autoplot() methods on the inference and placebo objects produce publication-quality figures.
• Cross-platform parallel placebos. parallel = TRUE does the right thing on Windows (PSOCK cluster) and unix-likes (forks).
• Two vignettes. vignette("synth-quickstart") for a 5-minute intro and vignette("inference") for a deep dive on the canonical Basque example.

Worked example: California’s Proposition 99

The 1988 California cigarette-tax measure is a textbook case for the synthetic control method. With the new API the entire workflow is six function calls:

library(Synth); library(haven); library(ggplot2)
data(smoking)  # or read_dta("smoking.dta")

dp <- synth_data(
  panel              = smoking,
  outcome            = "cigsale",
  unit_col           = "state_id",
  time_col           = "year",
  treated            = "California",
  treatment_time     = 1989,
  predictors         = c("lnincome", "age15to24", "retprice", "beer"),
  special_predictors = list(
    list("cigsale", 1988, "mean"),
    list("cigsale", 1980, "mean"),
    list("cigsale", 1975, "mean")),
  unit_names_col     = "state_name"
)

fit  <- synth(dp)
inf  <- synth_inference(fit, dp, method = "conformal", alpha = 0.10)
pl   <- generate_placebos(fit, dp)
test <- mspe_test(pl)        # one-sided p-value = 0.026

autoplot(inf)                # 90% conformal band
autoplot(pl, mspe_threshold = 5)  # placebo overlay

The synthetic California puts about 84% of weight on Utah, Nevada, Montana, and Connecticut (matching the published Synth paper). Post / pre MSPE ratio is 128 and the placebo p-value is 0.026 — the effect is unusually large relative to other states.

The previous release (1.1-10, April 2026) addressed quadopt = "LowRankQP" fail-fast, the missing-data check in dataprep(), quieter defaults, and a path.plot() y-axis fix for negative-valued series.

The Stata routine was likewise updated in April 2026 to version 0.0.8 with native Apple Silicon support (the optimizer plugin now ships an arm64 Mach-O slice; previously it failed to load on M-series Macs running Stata 17+ natively), portable C source that compiles cleanly on macOS / Linux / Windows, and typo / version-declaration cleanup.

Synth for R — also on GitHub

Synth for Stata — also on GitHub