Synthetic Difference-in-Differences - Overview

Summary

Synthetic Difference-in-Differences (SDID), from Arkhangelsky, Athey, Hirshberg, Imbens & Wager (2021, AER), is a panel-data estimator that fuses DiD and synthetic control (SC). It fits the ordinary two-way fixed-effects (TWFE) DiD regression, but weighted by (i) SC-style unit weights ${\hat{\omega }}_i$ that make the control group’s pre-period trend parallel to the treated group’s, and (ii) time weights ${\hat{\lambda }}_t$ that pick out the pre-periods most predictive of the post-period. Like SC it weakens reliance on parallel trends; like DiD it is invariant to additive unit shifts and supports large-panel inference. This cluster also covers the other modern response to fragile parallel trends: event-study designs, Roth’s (2022) critique of pre-trend testing, and Rambachan & Roth’s (2023) Honest DiD sensitivity analysis.

Overview

Two literatures answer the same question — “what would the treated units have looked like without treatment?” — with different crutches.

• DiD assumes parallel trends: after removing additive unit and time effects, treated and control units would have moved together. It is typically used with many treated units.
• SC assumes a weighted average of control units can reproduce the treated unit’s pre-period path, and hence its counterfactual post-period path. It is typically used with one (or a few) treated units and a long pre-period.

Arkhangelsky et al. argue that “although the empirical settings where DID and SC methods are typically used differ, the fundamental assumptions that justify both methods are closely related” (§1, p. 2). Their estimator is a single weighted TWFE regression:

$$({\hat{\tau }}^{sdid},\hat{\mu },\hat{\alpha },\hat{\beta })=\arg \underset{\tau ,\mu ,\alpha ,\beta }{\min }\sum _{i=1}^N\sum _{t=1}^T{\left(Y_{it}-\mu -{\alpha }_i-{\beta }_t-W_{it}\tau \right)}^2{\hat{\omega }}_i^{sdid}{\hat{\lambda }}_t^{sdid}$$

DiD is the same regression with no weights (eq. 1.2); SC is the same regression with unit weights ${\hat{\omega }}^{sc}$ but no unit fixed effects ${\alpha }_i$ and no time weights (eq. 1.3). The weights make the regression “local”: it emphasises control units whose past resembles the treated units’ past, and pre-periods that resemble the post-period. The authors identify two benefits: robustness (bias removal when parallel trends fails in the raw data) and — less obviously — precision, because the weights “implicitly remov[e] systematic (predictable) parts of the outcome” (§1, p. 4).

The second half of this cluster concerns what to do when you stay inside the DiD framework and must defend parallel trends. The standard defence is the event-study plot with insignificant pre-treatment coefficients. Roth (2022) shows this defence is weak: pre-trend tests have low power, and conditioning on passing them can make bias worse. Rambachan & Roth (2023) replace the binary pre-test with a partial-identification sensitivity analysis that bounds post-treatment violations of parallel trends by what was observed pre-treatment. Arkhangelsky et al. themselves position SDID against this literature: “Our approach uses past data not only to check whether the trends are parallel, but also to construct the weights to make them parallel” (§6, p. 32).

Main Content

Setting and estimand ^def-sdid-setting

Balanced panel, $N$ units, $T$ periods, outcome $Y_{it}$, binary treatment $W_{it}$. Block assignment: the first $N_{co}$ units are never treated; the last $N_{tr}=N-N_{co}$ units are treated in periods $t>T_{pre}$, so $W_{it}=1\{i>N_{co},t>T_{pre}\}$ and $T_{post}=T-T_{pre}$. The target is the average treatment effect on the treated cells,

$$\tau =\frac{1}{N_{tr}{T}_{post}}\sum _{i=N_{co}+1}^N\sum _{t=T_{pre}+1}^T{\tau }_{it}$$

(eq. 4.2). Staggered adoption is handled by applying SDID once per adoption date and averaging (Appendix §8).

The three estimators as one regression family ^def-sdid-family

Estimator	Unit weights	Time weights	Unit FE ${\alpha }_i$	Time FE ${\beta }_t$
DiD (1.2)	uniform $1/N_{co}$	uniform $1/T_{pre}$	yes	yes
SC (1.3)	${\hat{\omega }}^{sc}$ (simplex, no intercept)	none (pre-periods unused in the final contrast)	no	yes
DIFP (SC with intercept)	$\hat{\omega }$ with intercept	uniform	yes	yes
SDID (1.1)	${\hat{\omega }}^{sdid}$ (simplex, intercept, ridge)	${\hat{\lambda }}^{sdid}$ (simplex, intercept)	yes	yes

DIFP is the Doudchenko–Imbens (2016) / Ferman–Pinto (2019) “demeaned SC”; the paper reports it to separate the gains from fixed effects from the gains from time weights.

Headline results of Arkhangelsky et al. (2021) ^thm-sdid-headline

1. Double robustness (§4.2): the bias term vanishes if either the unit weights balance the latent unit factors or the time weights balance the latent time factors — see the bias decomposition.
2. Asymptotic normality (Theorem 1): under a latent factor model $Y=L+W\circ \tau +E$ with large $N_{co}$, $T_{pre}$ and a comparatively small treated block, $({\hat{\tau }}^{sdid}-\tau )/V_{\tau }^{1/2}\Rightarrow \mathcal{N}(0,1)$ with $V_{\tau }$ of order $1/(N_{tr}{T}_{post})$ — see SDID Inference - Bootstrap, Jackknife and Placebo.
3. Empirical dominance (§3): in placebo simulations calibrated to CPS wage data and Penn World Table GDP, SDID has the lowest or near-lowest RMSE and bias among DiD, SC, DIFP and matrix completion in nearly every design.

Map of the cluster

Question	Note
How exactly are $\hat{\omega }$, $\hat{\lambda }$ and $\zeta$ computed?	SDID Estimator - Unit and Time Weights
Why does weighting + fixed effects beat DiD and SC? What do the simulations show?	SDID vs DiD vs Synthetic Control
How do I get a standard error — especially with one treated unit?	SDID Inference - Bootstrap, Jackknife and Placebo
How does this apply to geo tests and marketing panels?	SDID for Geo Experiments and Marketing Panels
How are dynamic effects ${\tau }_{\ell }$ estimated and plotted?	Event Study Designs and Dynamic Treatment Effects
Why is “no significant pre-trend” weak evidence?	Pre-Trend Testing and Its Pitfalls
What replaces the pre-test?	Honest DiD - Sensitivity to Parallel Trends Violations

Relevance to marketing measurement and applied work

Geo experiments and market-level campaign evaluations are exactly the data shape SDID was designed for: a modest number of geos ($N\approx 50$–$200$ DMAs), a long weekly pre-period, a small block of treated geo-weeks, and outcomes (sales, conversions) dominated by geo-level scale differences plus shared seasonality — i.e. strong additive fixed effects plus an interactive component. Practical implications:

• Against plain DiD / TWFE for geo lift: treated markets are rarely chosen at random (pilots go to large or strategic markets), so assignment correlates with latent factors and DiD is biased; SDID’s unit weights repair this, and its CPS placebo study shows it is more precise than DiD even under random assignment.
• Against pure SC / TBR-style level matching: geos differ by orders of magnitude in volume; SDID’s intercept + unit fixed effects mean the donor pool only has to match trends, not levels.
• Against CausalImpact/BSTS: SDID is a transparent, convex-weight, frequentist alternative with formal asymptotics; BSTS gives a full posterior and handles a single series more flexibly. Reporting both is a useful robustness pairing.
• For observational (non-randomised) marketing panels — staggered product launches, regional price changes, platform rollouts — the event-study / pre-trend / Honest DiD notes supply the defensible reporting standard: plot dynamic effects, report the power of the pre-trend test, and report the breakdown value $\bar{M}$ at which the lift conclusion fails.

Examples

California Proposition 99 (§2, Table 1). 39 states, 1970–2000, California treated from 1989: $T_{pre}=19$, $T_{post}=12$, $N_{co}=38$, $N_{tr}=1$. Estimated effect on per-capita cigarette sales (packs/year), with placebo standard errors:

	SDID	SC	DID	MC	DIFP
Estimate	$-15.6$	$-19.6$	$-27.3$	$-20.2$	$-11.1$
Std. error	$(8.4)$	$(9.9)$	$(17.7)$	$(11.5)$	$(9.5)$

DiD’s $-27.3$ “likely overstates the effect” because pre-trends are visibly non-parallel; SDID is both smaller in magnitude and has a standard error less than half of DiD’s despite being more flexible — “a result of the local fit of SDID (and SC) being improved by the weighting” (p. 8).

Minimal usage with the authors’ R package synthdid:

library(synthdid)
data("california_prop99")
setup   <- panel.matrices(california_prop99)      # Y (N x T), N0, T0
tau.hat <- synthdid_estimate(setup$Y, setup$N0, setup$T0)
se      <- sqrt(vcov(tau.hat, method = "placebo")) # N_tr = 1 -> placebo only
plot(tau.hat)                                      # parallel-trends picture + time weights

Connections

• Differences-in-Differences — the unweighted special case; SDID is “DiD on a reweighted panel.”
• Synthetic Control, Synthetic Control Bias Theory, Synthetic Control Requirements — SDID’s unit weights are ADH weights plus an intercept and ridge penalty; the factor-model bias analysis parallels Abadie’s.
• Generalized Synthetic Control Method — the explicit-factor-model alternative (Xu 2017). SDID balances the latent factors without estimating them and needs weaker assumptions (no known rank, no strong-factor condition).
• Synthetic Control Extensions — augmented SC (Ben-Michael, Feller & Rothstein) is nearly equivalent to SDID with a linear outcome model.
• Fixed-Effects Model — the TWFE regression being weighted.
• Difference-in-Differences with Multiple Time Periods - Overview — the Callaway–Sant’Anna cluster for staggered adoption; the event-study note links the two.
• Bayesian Difference in Differences — a Bayesian treatment of the same parallel-trends problem.
• Randomization Inference - Overview — placebo inference for SDID is a permutation-style procedure.

See Also

• Synthetic Control Inference and Diagnostics
• Standard Errors and Clustering
• Geo-Experiment Methodology - Overview
• Sensitivity Analysis in Observational Studies
• Group-Time Average Treatment Effects
• Aggregating Group-Time Effects