Synthetic Difference In Differences

The paper studies causal inference in panel data where treatment occurs for some units at some times. It reframes the Synthetic Control (SC) method as a weighted least squares regression with time fixed effects and unit-specific weights, and contrasts it with standard Difference-in-Differences (DID), which uses both unit and time fixed effects but no weights. Building on this, the authors propose the Synthetic Difference-in-Differences (SDID) estimator, which introduces both unit and time weights within a two-way fixed effects regression. The research aims to improve bias properties and robustness relative to SC and DID by combining outcome modeling with balancing weights. The key insight is a double robustness property: SDID is consistent if either a two-way fixed effects model is correctly specified (even with broad classes of data-dependent weights) or, when that model is misspecified, if the underlying signal matrix is approximately low-rank and penalized SC-style weights are used. The paper also develops asymptotic theory as N and T grow, shows conditions for consistency and asymptotic normality, and provides inference procedures using standard robust methods while treating data-adaptive weights as fixed. Empirical and simulation evidence suggests improved predictive accuracy and better-calibrated inference compared to DID and SC.

The paper builds on SC methods introduced by Abadie and coauthors (2003; 2010; 2015), including extensions and alternatives (Doudchenko and Imbens, 2016; Xu, 2017; Athey et al., 2017; Carvalho et al., 2018; Li and Bell, 2017). It relates to Augmented Synthetic Control (Ben-Michael, Feller, and Rothstein, 2018), which combines outcome modeling with SC weights, yielding double robustness but lacking a weighted regression interpretation. The authors connect SDID to the broader program evaluation literature on combining outcome models and balancing/weighting for double robustness (Robins et al., 1994; Scharfstein et al., 1999; Newey et al., 2004; Belloni et al., 2014; Chernozhukov et al., 2018; Athey, Imbens, and Wager, 2018; Hirshberg and Wager, 2018). They also draw on panel models with latent factor/low-rank structure (Bai and Ng, 2002; Bai, 2009; Bonhomme and Manresa, 2015; Xu, 2017) and discuss inference traditions in DID and panel regressions (Arellano, 2003; Hansen, 2007; Liang and Zeger, 1986; Stock and Watson, 2008).

Setup considers a balanced panel with outcomes Y_it, binary treatment W_it, and an underlying signal matrix L with Y_it = L_it + τ W_it + ε_it. The SC estimator is characterized as a weighted least squares regression with time fixed effects and unit-specific weights: minimize Σ_i Σ_t (Y_it − μ − β_t − τ W_it)^2 ω_i^SC. The DID estimator uses unweighted two-way fixed effects: minimize Σ_i Σ_t (Y_it − μ − α_i − β_t − τ W_it)^2. The proposed SDID estimator combines both unit and time weights with two-way fixed effects, minimizing Σ_i Σ_t (Y_it − μ − α_i − β_t − τ W_it)^2 ω_i λ_t. Weights factorize into unit weights ω_i and time weights λ_t. Constraints ensure nonnegativity, summing to one within control vs treated groups for units and pre- vs post-periods for times, and equal weighting within treated unit/time groups. Weight construction options include: (a) penalized synthetic control-style unit weights chosen to balance treated unit pre-trends, (b) time-analog SC weights to balance pre-periods to resemble post-period(s), optionally with an intercept to handle trends; (c) kernel or nearest-neighbor schemes. L2 (ridge) penalties are recommended to spread positive weights across many controls/time periods, helping asymptotics. The SDID estimator admits bias-reduction representations: it corrects a weighted control average by two bias adjustments—one for unit imbalance in pre-periods and one for time imbalance—mirroring DID’s double correction but in a weighted form. Generalization to multiple treated units and periods averages treated units and post-periods, constructs ω to balance treated vs control units in pre-periods and λ to balance pre- vs post-periods for controls, and then runs the weighted two-way fixed effects regression (optionally including covariates X_it). For inference, treat ω and λ as fixed and use robust panel standard errors (e.g., jackknife/HC3), accommodating within-row correlation. Asymptotic theory: Under large N,T, the authors analyze two cases: (1) correctly specified two-way fixed effects (L_it = μ + α_i + β_t), showing SDID is consistent and asymptotically normal for broad classes of data-dependent weights provided weights do not use the target row/column and are not too concentrated; (2) approximate low-rank L (interactive fixed effects/latent factors) where DID is inconsistent, but SDID with penalized SC weights is consistent; feasible weights learned on Y are shown to balance L sufficiently via empirical risk minimization tools, with bounds involving weight norms, gaussian widths, and approximate rank. They also provide extensions to multiple treated units/time periods with AR(1) errors, giving conditions under which √(N1 T1)(τ̂ − τ) is asymptotically normal with variance determined by treated-cell noise.

• Conceptual contribution: SC can be written as a weighted regression with time fixed effects; DID is unweighted two-way FE; SDID augments DID with unit and time weights, yielding a doubly weighted DID.
• Double robustness: SDID is consistent if (a) the two-way fixed effects model is correctly specified (for a wide variety of weighting schemes that avoid using the target cell and are sufficiently diffuse), or (b) the outcome model is approximately low-rank and penalized SC-style weights are used, even if two-way fixed effects is misspecified.
• Asymptotic results: Under well-specified two-way fixed effects and appropriate weight conditions, SDID is consistent and asymptotically normal. Under approximate low-rank L, SDID with penalized SC weights is consistent; they also provide conditions for SC consistency but those are stronger than for SDID due to SDID’s time-weighted bias removal.
• Inference: Standard large-sample inference for weighted DID regressions is valid when treating ω and λ as fixed. Jackknife/HC3 standard errors provide reliable coverage even though weights are data-adaptive.
• Placebo empirical evaluation (smoking data, Abadie et al. 2010): Predict one-step-ahead outcomes (1980–1988) using earlier data. SDID achieves substantially lower RMSE than SC and DID for almost all states; median state-wise RMSE improvement: SDID vs SC ≈ 15%, and vs DID ≈ 50%.
• Simulations (low-rank data): Across multiple (N,T), σ², and rank settings, SDID exhibits substantially lower bias and often lower RMSE than DID and SC. Tables report numerous configurations; e.g., N=T=200, σ=0.5, rank=2: RMSE(DID)=1.38, SC=0.29, SDID=0.11 (exchangeable); in non-exchangeable designs, SDID retains bias advantages.
• Confidence intervals: In a design with multiple treated units (N=100, N1=20; T=120, T1=5; σ=2; τ=1; rank=2), SDID 95% CIs had ≈98% coverage (slightly conservative), while unweighted DID had ≈82% coverage. With correlated errors (ρ=0.7), SDID coverage ≈93% vs DID ≈88%.
• Practical guidance: L2-penalized SC weights improve large-sample properties by spreading weights; time weights with intercept help under trends; robust, row-wise jackknife variance estimation works well.

The SDID framework unifies SC and DID by expressing both within a weighted regression perspective and then combining their strengths. By adding unit and time weights to a two-way fixed effects regression, SDID relaxes the global parallel trends assumption to a local, weighted version, targeting balance where it matters for imputing counterfactual outcomes. The double robustness result is central: consistency is achieved if either the simple model (two-way fixed effects) is correct or the weighting balances the relevant low-rank structure, mitigating risks of model misspecification or imperfect balance. Empirical and simulation evidence shows that SDID meaningfully reduces bias and improves predictive performance relative to SC and DID, especially when the treated unit is not well represented by a convex combination of controls or when temporal patterns differ. For inference, despite weights being data-dependent, treating them as fixed in standard robust DID regressions yields valid standard errors under large-sample conditions, providing a practical path for uncertainty quantification. These contributions make SDID broadly relevant for panel treatment effect estimation with staggered or single-shot treatments, particularly in settings with complex latent structures.

The paper introduces Synthetic Difference in Differences (SDID), a doubly weighted two-way fixed effects estimator that generalizes SC and DID. It provides: (i) a regression-based interpretation of SC, (ii) a new estimator with double robustness—consistent under correct two-way fixed effects with general weights or under approximate low-rank structure with penalized SC weights, and (iii) asymptotic theory and practical inference procedures based on standard weighted DID regressions. Empirical and simulation studies demonstrate improved accuracy and better-calibrated inference relative to DID and SC. The framework naturally extends to multiple treated units and periods and accommodates covariates. Future research directions include refining data-adaptive weight construction and penalties, extending inference under richer dependence structures, exploring heterogeneous treatment effects and dynamic treatments, and integrating high-dimensional covariates or advanced factor models within the SDID weighting-regression framework.

• Theoretical guarantees rely on large N and T asymptotics; finite-sample performance depends on design features such as factor structure and noise.
• Consistency under misspecification requires approximate low-rank structure and penalized SC weights with sufficient weight dispersion; if low-rank approximation is poor or weights concentrate on few units/periods, properties may degrade.
• Inference procedures treat data-adaptive weights as fixed; while justified asymptotically, small-sample behavior may be conservative or sensitive.
• The approach assumes balanced panels (as presented) and specific treatment timing patterns; extensions to unbalanced panels or complex staggered adoption require care.
• When treated units lie outside the convex hull of controls or pre-periods offer weak information about the post-period, weight-based balancing may be challenging and require diagnostics/regularization.
• Autocorrelation and heteroskedasticity are addressed under certain structures (e.g., AR(1) rows); more complex dependence may require tailored variance estimators or block jackknife schemes.