Synthetic Difference-in-Differences (SDID)

• Synthetic Difference-in-Differences is a hybrid estimator that combines synthetic control weighting with difference-in-differences to address non-parallel trends in panel data.
• It uses data-driven unit and time weights to improve pre-treatment alignment between treated and control groups, enhancing causal inference.
• The method offers robustness through fixed effects and double robustness properties, making it suitable for staggered adoption and event-study analyses.

Synthetic Difference-in-Differences (SDID) is a panel-data estimator for causal effects that combines synthetic-control-style reweighting with difference-in-differences (DiD) differencing. In the formulation of Arkhangelsky et al., SDID chooses unit weights so that control units resemble treated units in the pre-treatment period, chooses time weights so that pre-treatment periods resemble post-treatment periods, and then estimates the treatment effect in a weighted two-way fixed-effects regression. The method is designed for settings in which conventional DiD is vulnerable because untreated outcomes need not satisfy parallel trends in raw data, while standard synthetic control is too restrictive because it emphasizes pre-treatment fit without the same invariance to additive unit and time shifts (Arkhangelsky et al., 2018).

1. Conceptual position in the DiD and synthetic control literature

SDID was introduced to resolve a familiar tension in panel causal inference. Standard DiD is attractive because it removes common time shocks and accommodates additive unit and time heterogeneity, but it relies on parallel trends. Synthetic control instead constructs a weighted average of donor units that closely matches the treated unit’s pre-treatment path, but standard formulations are less naturally suited to multiple treated units, staggered adoption, and persistent treated-control level differences. SDID combines these logics by using synthetic-control-style weighting and then applying a DiD-style contrast on the reweighted panel (Arkhangelsky et al., 2018).

A useful way to situate the estimator is as a constrained comparison of three designs:

Method	Core weighting structure	Main identifying intuition
DiD	Equal weights on units and periods	Parallel trends after additive differencing
Synthetic control	Data-driven unit weights	Match pre-treatment outcome path
SDID	Data-driven unit weights and time weights	Reweighted DiD after improving pre-treatment comparability

This hybrid structure was anticipated conceptually by work that treated DiD, synthetic control, regression adjustment, and matching as variants of a common linear counterfactual imputation problem. In particular, Doudchenko and Imbens proposed a framework with a nonzero intercept, flexible weights, and elastic-net regularization, thereby articulating the synthetic-DID-style idea of combining synthetic-control weighting with DID’s ability to absorb a permanent additive difference between treated and control units (Doudchenko et al., 2016).

Subsequent expositions emphasize that SDID differs from both parent methods in two specific ways. Relative to DiD, it does not require treated and control units to have parallel average pre-treatment trends in aggregate. Relative to standard synthetic control, it does not require the treated unit or treated group to lie in the convex hull of controls, and it allows level differences via fixed effects (Clarke et al., 2023). This suggests that SDID is best understood not as a compromise in an informal sense, but as a reweighted DiD estimator whose weighting stage is designed to make the DiD comparison more credible.

2. Core estimator and weighting structure

The canonical SDID setup is the block-treatment design

$W_{it}=1\{i>N_0,\ t>T_0\},$

with $N_0$ never-treated controls, $N_1=N-N_0$ treated units, $T_0$ pre-treatment periods, and $T_1=T-T_0$ post-treatment periods. The target estimand is the average treatment effect on the treated block,

$$\tau = \frac{1}{N_1T_1}\sum_{i=N_0+1}^{N}\sum_{t=T_0+1}^{T}\tau_{it}.$$

The estimator is a weighted two-way fixed-effects regression: $$\hat\tau^{sdid}, \hat\mu, \hat\alpha, \hat\beta = \arg\min_{\tau,\mu,\alpha,\beta} \sum_{i=1}^{N}\sum_{t=1}^{T} \omega_i\lambda_t \bigl(Y_{it}-\mu-\alpha_i-\beta_t-W_{it}\tau\bigr)^2.$$ Here $\omega_i$ are unit weights and $\lambda_t$ are time weights. In the special case of uniform weights, this collapses to standard DiD; with unit weights but without the unit fixed effects, it moves toward synthetic control (Arkhangelsky et al., 2018).

The unit weights are chosen to make the weighted control-group pre-treatment outcomes resemble the treated-group pre-treatment outcomes, up to an intercept: $$(\hat\omega_0,\hat\omega) = \arg\min_{\omega_0\in\mathbb{R},\,\omega\in\Omega} \sum_{t=1}^{T_0} \left( \omega_0+\sum_{i=1}^{N_0}\omega_iY_{it} -\frac{1}{N_1}\sum_{i=N_0+1}^{N}Y_{it} \right)^2 +\zeta^2 T_0 \|\omega\|_2^2.$$ The time weights are chosen analogously: $N_0$0 In the original construction, the control weights are nonnegative and sum to one, the intercept terms $N_0$1 and $N_0$2 allow level adjustment, and the unit-weight objective includes an $N_0$3 penalty to stabilize the solution. The recommended regularization is

$N_0$4

with $N_0$5 estimated from pre-treatment first differences (Arkhangelsky et al., 2018).

A central identity is that SDID can be written as a weighted double-difference: $N_0$6 This representation makes clear that the method is still a DiD estimator in structure; the novelty is that both the cross-sectional comparison and the temporal comparison are data-adaptively localized before differencing. In expository treatments, this is often summarized as

$N_0$7

with the weights chosen so that pre-treatment treated and control trajectories align as closely as possible (Ciccia, 2024).

3. Identification logic, robustness properties, and asymptotic theory

The theoretical analysis of SDID is built around the outcome model

$N_0$8

where the untreated systematic component is allowed to have latent interactive structure,

$N_0$9

This is more general than the additive model $N_1=N-N_0$0 underlying conventional DiD. If untreated outcomes were exactly additive, standard DiD would be sufficient; SDID is designed for the empirically important case in which there is an additive component plus a low-rank interactive remainder (Arkhangelsky et al., 2018).

Two robustness properties are central. First, because the estimating regression includes unit and time fixed effects and the weighting problems include intercepts, SDID is invariant to additive row and column shifts,

$N_1=N-N_0$1

This is the DID-style robustness to level differences that standard synthetic control lacks. Second, the theory establishes a double robustness-type property: if the unit weights balance the latent unit factors well, SDID remains accurate even with imperfect time weights, and conversely if the time weights balance latent time factors well, SDID remains accurate even with imperfect unit weights (Arkhangelsky et al., 2018).

The asymptotic analysis assumes Gaussian, independent-across-units errors with bounded covariance eigenvalues, a large-panel regime in which treated units and treated periods both grow while treatment remains a small fraction of the sample, and regularization satisfying

$N_1=N-N_0$2

Under these conditions, the estimator is asymptotically linear and asymptotically normal: $N_1=N-N_0$3 and

$N_1=N-N_0$4

The variance expression is explicit in the original analysis and is interpreted as the optimal variance obtainable when the systematic component and error structure are known (Arkhangelsky et al., 2018).

Applied treatments of the method add operational identifying conditions. In particular, implementations typically require a strongly balanced panel, no always-treated units in the estimation sample, at least two pre-treatment periods, and some pure controls that are never treated. For event-study SDID, the validity conditions include no anticipation, comparability of treated cohorts to never-treated controls via synthetic reweighting, stable treatment timing within cohorts, and no contamination across groups (Clarke et al., 2023).

4. Staggered adoption and event-study decomposition

Although the original exposition emphasized a common adoption date, SDID has been extended to staggered adoption by defining treatment groups according to adoption date and then aggregating adoption-specific estimates. In the Stata implementation of Clarke et al., treatment groups are defined by adoption year $N_1=N-N_0$5, SDID is applied separately within each adoption group, and the overall average treatment effect is formed as a weighted average of the adoption-specific estimates (Clarke et al., 2023).

A more explicit dynamic treatment-effect formulation appears in the event-study extension of SDID. For a treated cohort $N_1=N-N_0$6, the lag-$N_1=N-N_0$7 effect is defined as the treated-control outcome gap at event time $N_1=N-N_0$8 relative to the weighted pre-treatment average gap: $N_1=N-N_0$9 The cohort-specific SDID estimate is then exactly the average of these dynamic effects: $T_0$0 Thus the cohort-level ATT is not separate from the event-study coefficients; it is literally their mean (Ciccia, 2024).

In staggered adoption, lag-specific estimates are aggregated across all cohorts for which that lag is observed. If $T_0$1 is the set of cohorts with an observable $T_0$2-th post-treatment effect and $T_0$3, then

$T_0$4

The overall ATT can also be decomposed into a weighted average of lag-specific event-study effects: $T_0$5 This decomposition is important for interpretation. Each $T_0$6 compares the lag-$T_0$7 treated-versus-synthetic-control differential to its own weighted pre-treatment baseline, and each aggregated $T_0$8 is a cohort-size-weighted average across eligible cohorts. The event-study plot should therefore be read as a dynamic path relative to a learned pre-treatment benchmark, not as raw differences (Ciccia, 2024).

5. Inference, software implementations, and empirical workflow

Inference for SDID is typically based on asymptotic normality together with resampling or placebo procedures. The Stata implementation sdid supports three main variance estimators: block bootstrap, jackknife, and placebo / permutation inference. Bootstrap resamples units as blocks and recomputes the estimator, jackknife leaves out one unit at a time with original weights fixed, and placebo inference repeatedly assigns placebo treatment among controls. The methods have different domains of applicability: bootstrap requires more than one treated unit, jackknife is not defined when a treatment group has only one treated unit, and placebo inference requires more controls than treated units and homoskedasticity across units (Clarke et al., 2023).

The same implementation organizes the practical workflow around several diagnostics. Data are expected to be balanced and free of always-treated observations; users typically inspect the estimated ATT, unit weights, time weights, adoption-specific treatment effects, and the treated-versus-synthetic trajectory plots. The command stores outputs such as e(ATT), e(se), e(tau), e(lambda), and e(omega), and supports explicit comparison across method(sdid), method(did), and method(sc) (Clarke et al., 2023).

Dynamic SDID estimation is operationalized in the sdid_event Stata package. Its default output reports the overall ATT from sdid together with $T_0$9 for $T_1=T-T_0$0. With the disag option, it also reports cohort-specific $T_1=T-T_0$1 and cohort-by-event-time effects $T_1=T-T_0$2. The cohort-specific $T_1=T-T_0$3 can also be retrieved from the e(tau) matrix in sdid (Ciccia, 2024).

A recurring practical point is that SDID is informative not only through the point estimate but through the estimated weighting structure. Sparse or highly concentrated unit weights, unusual time weights, or poor pre-treatment fit can indicate fragile identification. This suggests a substantive reading of SDID in which pre-treatment alignment is part of the evidentiary basis for the causal claim, rather than a purely technical preprocessing step.

6. Extensions, adjacent frameworks, and substantive uses

Several later papers extend SDID beyond the original block-treatment panel design. Sequential Synthetic Difference-in-Differences addresses event studies with staggered treatment adoption by performing iterative imputation on cohort averages: estimates for earlier-treated cohorts are used to impute untreated counterfactuals for later-treated cohorts, preventing contamination from already-treated outcomes. Under a linear model with interactive fixed effects, the resulting estimator is asymptotically equivalent to an infeasible oracle OLS estimator, which yields asymptotic normality, valid Bayesian bootstrap inference, and, under homoskedastic serially uncorrelated aggregate errors, an efficiency result relative to that oracle benchmark (Arkhangelsky et al., 2024).

A separate extension adapts SDID to repeated cross-sectional data. In that design, treatment is assigned at the group level, outcomes are aggregated by group-period, and the original unit and time weights are supplemented with a third weight

$T_1=T-T_0$4

which corrects for unequal numbers of observations across group-period cells. Simulation results show lower bias and RMSE relative to both DiD and unadjusted SDID when interactive fixed effects are present and cell sizes vary (Morin, 2024).

Broader theoretical frameworks place SDID within a larger unifying literature. Generalized DiD introduces a stable bias assumption and interprets the estimator as a post-period ignorability-style imputation debiased by an analogous pre-period imputation, thereby connecting DiD to synthetic-control-style use of pre-treatment outcomes without adopting SDID’s specific unit-weight-plus-time-weight construction (Agniel et al., 2023). A doubly robust DiD–synthetic control integration proposes a single semiparametric ATT score that identifies the effect under either parallel trends or a group-level synthetic-control assumption, extends to repeated cross-sections and staggered adoption, and uses a multiplier bootstrap because the asymptotic variance differs across the two identifying regimes (Sun et al., 14 Mar 2025). A groupwise-matching framework, in turn, treats DiD, synthetic control, and SDID as distinguished by the matching conditions they impose on weighted group means after differencing (Rincón et al., 30 Oct 2025).

Other related methods should not be conflated with SDID. Entropy-balanced DiD reweights controls to exactly match treated units on moments of pre-intervention outcome trends and is justified by a weighted parallel-trends logic rather than by SDID’s joint unit and time weighting (Cefalu et al., 2020). “Synthesized Difference in Differences” is a different estimator entirely: it combines randomized clinical trial data with observational data to learn a linear adjustment to non-parallel slopes and targets the conditional average treatment effect for a clinical population, rather than the panel ATT under synthetic weighting (Strobl et al., 2021). A synthetic triple-difference procedure rewrites DDD as DID on a transformed outcome and then applies SDID-style weights, making it an extension by transformation rather than a redefinition of SDID itself (Zhuang, 2024).

Substantive applications illustrate SDID’s empirical role. In a country-quarter panel using GitHub Innovation Graph data, one study compares DiD, synthetic control, and SDID to estimate the effect of ChatGPT availability on software development activity. The design includes 151 economies from 2020 Q1 to 2023 Q1, 11 pre-treatment quarters, 2 post-treatment quarters, and treatment beginning in 2022 Q4. SDID is used as the hybrid estimator that reweights both countries and pre-treatment periods; the reported findings indicate positive effects on git pushes, repositories, and unique developers per 100,000 people, while also noting the short post-treatment window and the distinction between activity quantity and software quality (Quispe et al., 2024).

Across these extensions and applications, a stable theme remains: SDID is a weighted DiD estimator whose central innovation is to learn a synthetic comparison group in both the unit and time dimensions before applying the treatment contrast. Related frameworks generalize, reinterpret, or adapt this idea, but the defining object remains the weighted two-way fixed-effects comparison induced by jointly optimized unit and time weights.